feat(platform): add community-maintained Pi wrapper (#636)

* feat(platform): add pi wrapper

* feat(pi): scope ARS activation and add evidence

* fix(pi): gate automatic ARS skill invocation

* fix(pi): preserve args and tree activation

* docs(pi): scope prompt guarantees to idle runs

* fix(pi): canonicalize hidden skill paths

Pi preserves lexical resource paths while Node resolves extension imports through symlinks. Compare real filesystem identity after XML decoding.

* fix(pi): substitute command arguments once

Avoid reprocessing placeholder-like text introduced from user arguments.

* fix(pi): bound inactive skill block matching

Prevent alternate skill formatting from extending a match into an adjacent unrelated skill block.

examples/pi/README.md

@@ -0,0 +1,85 @@
+# Pi degraded-mode end-to-end evidence
+
+This directory is a privacy-minimized evidence bundle from one completed `/ars-full` run through the thin Pi wrapper. It demonstrates checkpoint enforcement, source-read provenance gating, Stage 5 finalization, Stage 6 process recording, and persisted terminal completion when no Pi orchestration capability was installed.
+
+This is evidence from one synthetic test scenario, not a general conformance certificate or a claim that degraded execution is equivalent to independent multi-agent review.
+
+## Run identity
+
+| Field | Value |
+| --- | --- |
+| Run ID | `ars-pi-e2e-gpt-5.6-sol-20260803T051847Z` |
+| Pi | `0.83.0` |
+| Provider/model | `openai-codex/gpt-5.6-sol` |
+| ARS pipeline | `3.19.0` |
+| Repository HEAD used for curation | `682b158ea02b6bdcc1f70b1ba577f237645bf7f5` |
+| Execution | sequential, same recovered Pi/model context |
+| Terminal state | Stage 6 `completed`, checkpoint confirmed, acknowledgement received, 36/36 round trips |
+
+The run crossed a disclosed transport recovery and automatic context compaction. Continuity was established from persisted artifacts, state, and hashes; it was not represented as a new academic run.
+
+After the run, the wrapper's system-prompt injection was scoped to ARS-active sessions and inactive prompts were hardened to hide this package's ARS skill entries. That guarantee and its probes cover prompts submitted while Pi is idle. Pi 0.83.0 does not rebuild the system prompt for prompts queued into an in-flight run, and its direct RPC `steer` and `follow_up` methods bypass the wrapper's `input` handler; those residuals are documented rather than represented as covered. The compatibility text used after `/ars-full` is unchanged; the final wrapper received fresh idle-path activation, automatic-invocation, and package-load regression tests, but these small post-run changes did not receive another full pipeline execution. See [`model-portability.md`](model-portability.md).
+
+## Load-bearing Pi boundaries
+
+1. **No wrapper-provided agent isolation.** [`runtime/ars-pi-doctor.txt`](runtime/ars-pi-doctor.txt) reported `Orchestration: none; sequential/degraded mode`. Specialist roles therefore ran sequentially in one recovered Pi/model context. No independent reviewer, writer, verifier, observer, formatter, or cross-model judgment is claimed. The external RPC driver delivered synthetic test-operator checkpoint responses only; it was not available to the inner Pi session as an orchestration tool or skill.
+2. **No Claude hook enforcement.** The same doctor record reported that Claude hooks are unavailable in Pi. ARS write-scope enforcement was prompt-level in this run; no deterministic Claude `PreToolUse` hook boundary is claimed.
+
+The installed `/skill:websearch` capability was visible to Pi. Source verification remained bounded as disclosed in the provenance and process records. [`runtime/package-load-validation.json`](runtime/package-load-validation.json) records fresh offline root/nested package loads that found all three wrapper commands, all 16 `/ars-*` prompt templates, and all four ARS skills, plus nine passing unit tests (including filesystem-backed symlink/XML-location, block-boundary, and single-pass argument-substitution probes) and five real Pi idle-prompt-scope probes.
+
+## Evidence chain
+
+### Full stage chain
+
+[`run-log.md`](run-log.md) provides the chronological checkpoint record. Representative byte-for-byte outputs make each executed stage inspectable:
+
+- Stage 1 Research: [`stage-01/revised-research-report.md`](stage-01/revised-research-report.md)
+- Stage 2 Write: [`stage-02/course-paper-draft.md`](stage-02/course-paper-draft.md)
+- Stage 2.5 Integrity: [`stage-02.5/focused-reverification-report.md`](stage-02.5/focused-reverification-report.md)
+- Stage 3 Review: [`stage-03/editorial-decision-and-revision-roadmap.md`](stage-03/editorial-decision-and-revision-roadmap.md)
+- Stage 4 Revise: [`stage-04/response-to-reviewers.md`](stage-04/response-to-reviewers.md)
+- Stage 3′ Re-review: [`stage-03p/verification-review-report.md`](stage-03p/verification-review-report.md)
+- Stage 4′ Re-revise: [`stage-04p/validation-report.md`](stage-04p/validation-report.md)
+- Stage 4.5 Final integrity: [`stage-04.5/final-integrity-verification-report.md`](stage-04.5/final-integrity-verification-report.md)
+- Stage 5 Finalize and Stage 6 Process summary: final package, process records, checkpoints, and terminal state below
+
+### Source-read hard gate
+
+1. [`stage-05/cite-time-gate-refusal-report.md`](stage-05/cite-time-gate-refusal-report.md) and [`stage-05/cite-time-gate-refusal.json`](stage-05/cite-time-gate-refusal.json) show that Stage 5 refused converted output while all 79 citation occurrences remained unacknowledged `LOW-WARN`.
+2. After the required human attestation was consumed, [`stage-05/cite-time-gate-pass.json`](stage-05/cite-time-gate-pass.json) recorded 21 attested sources, 79/79 markers resolved to `OK`, and a hard-gate `PASS` before clean outputs were emitted.
+3. Exact attestation wording, Pi session identifiers, event identifiers, and human-read ledger were intentionally omitted from the public bundle. The aggregate gate state is sufficient for port evidence and is not independent proof that reading occurred.
+
+### Stage 5
+
+- Retained public evidence: final [`paper.md`](stage-05/package/paper.md), [format validation](stage-05/package/format-validation.md), cite-time refusal/PASS records, and [FULL checkpoint](stage-05/full-checkpoint.md).
+- The run also generated a DOCX and advisory submission-verifier artifacts. Those redundant binary/snapshot files are intentionally omitted; their hashes and outcomes remain in the checkpoint and persisted state.
+- No paper PDF exists because the test operator declined paper LaTeX. HTML-to-PDF was not used.
+- The submission-package policy was advisory. B1–B5 and C1–C2 remained `NOT-CHECKED`; this is not a clean universal submission-verifier pass.
+
+### Stage 6 and terminal state
+
+- The concise public process evidence is the [English Markdown record](stage-06/process-record-en.md). Redundant PDF, LaTeX, translated copy, and build-validation files are omitted.
+- [Terminal checkpoint](stage-06/terminal-checkpoint.md) was correctly nonterminal when presented.
+- [`state/pipeline-state.json`](state/pipeline-state.json) is the controlling proof: `pipeline_state: completed`, Stage 6 `completed`, checkpoint confirmed, terminal acknowledgement received, and completion timestamp persisted. Exact acknowledgement wording and transport identifiers are omitted.
+
+## Model portability
+
+[`model-portability.md`](model-portability.md) compares observed `openai-codex/gpt-5.6-sol` behavior with the repository's illustrative Claude-oriented example. It separates runtime-established divergences from output differences that cannot be causally attributed to the model.
+
+## Privacy boundary
+
+- No maintainer name, email, account, home-directory path, credential, API key, authorization token, or Git remote is included.
+- “Scholar” decisions in generated artifacts are synthetic test-operator inputs for the canonical example. Funding, conflict, affiliation, authorship, and placeholder statements are not biographical declarations about the Pi port maintainer.
+- The public bundle records only aggregate human-read gate state; exact attestation wording and internal session/event identifiers are excluded.
+- Public-source UUIDs appearing inside government document URLs are source locators, not Pi or user identifiers.
+
+## Preserved limitations
+
+- Sequential same-context role simulation is weaker than independent agents or context windows.
+- Source verification and the scholar's source-read attestation have distinct evidentiary meanings.
+- The paper is a non-systematic evidence synthesis and retains the nonblocking PRISMA-trAIce/RAISE reporting warning.
+- The submission verifier was advisory and incomplete for checks requiring absent venue-profile or corpus inputs.
+- Four title-page placeholders remain.
+- The run experienced model-stream stalls, recovery, and compaction.
+
+See [`inventory.json`](inventory.json) for source-to-curated path and hash provenance. `SHA256SUMS` covers every curated file except itself. Raw RPC streams, full session logs, internal identifiers, downloaded source files, redundant binaries/translations, and intermediate caches are intentionally excluded.

examples/pi/SHA256SUMS

@@ -0,0 +1,23 @@
+fb8f6fc3edebe9c52bd41aa47a8f4ccb43c0530ba70c1ce1ec1f560f0b7d7f4e  ./inventory.json
+2a4316b250cc7fe4cd57ed4d42a0185bddf1e10d4dfd2fc4af0ddbce7cecf8a6  ./model-portability.md
+3da15ecfd8896bc35a0bf5af218fb64b9a69b08cad0fad160642b50c66169e95  ./README.md
+5d6e43ed109a26412048ca192f7622435d0186274d760796cb35e77fdcfa3bad  ./run-log.md
+47731cf8e217769f22844cc8a6ba8174c0d01279997e3d71fff46eceae30871e  ./runtime/ars-pi-doctor.txt
+3708902123deebf65757eee09182b63b6670c267b8f2cb6849174020ec6ab017  ./runtime/package-load-validation.json
+cee08d1b349854631a5634302852e58c16e0c51f480042fd467d1daa76e3abab  ./stage-01/revised-research-report.md
+614081675275d7b21b395c191b88545e779c130a626e91c941888b83b9c93a19  ./stage-02.5/focused-reverification-report.md
+953badfa35dab3ae6bcd183918b5d319b8575e22df944a3f0fa2ac54429f6507  ./stage-02/course-paper-draft.md
+972dafc30787c2407508f02f8c01eb4deab45e289087d930bb158a7e37f1ffa6  ./stage-03/editorial-decision-and-revision-roadmap.md
+926d7702847ff45d00a0157d71bede3119b291e80623988679f3518ad1087e46  ./stage-03p/verification-review-report.md
+d68db78036962dbcb6588a888343e5b611647203835ff263b7ef22e81be0bc9f  ./stage-04.5/final-integrity-verification-report.md
+c53ed200fe21b5ccc43a17ad3339995112ea53efd10adc5b03c4fe49a4feafd3  ./stage-04/response-to-reviewers.md
+02584274c834b082b132f48d57033efa054257f5d8f57848c377b1a265cfd803  ./stage-04p/validation-report.md
+90c75b0cc4748461bfae6b235c28d2e75fe700b44b87203d5ad1aca5c093735f  ./stage-05/cite-time-gate-pass.json
+4a92cb245675846e68ae105f8ce1478e8db6c13048e81f74781443d2b8ddedc2  ./stage-05/cite-time-gate-refusal-report.md
+6385ad669ab385b721de3e48c262f791f2e56e168893be761973c70406cf6f86  ./stage-05/cite-time-gate-refusal.json
+78d05d2c7bca23e3928931892ca1e86ac4bb637ed67bb93943410dffa2fd0e02  ./stage-05/full-checkpoint.md
+1400fa0809c62fea37eaac92c8302e97314c5904c36a0e67588ba25c8108cd14  ./stage-05/package/format-validation.md
+64ac9f2d5794f1f920136259869fe6a8f4c248f74a5d052f8d349818796ad219  ./stage-05/package/paper.md
+d378ec4100d64c2041083352af09a4d1b5f0acfb778ec047590d310a861f7ea3  ./stage-06/process-record-en.md
+11f6dd94e360f3119a10b9298ed585243fe3a93908a994772b621843523e3d87  ./stage-06/terminal-checkpoint.md
+4422a6db8075023e02612945e06e6072c1ebcd2ec6913ec9ccadfb7216a76933  ./state/pipeline-state.json

examples/pi/inventory.json

@@ -0,0 +1,216 @@
+{
+  "schema_version": 1,
+  "run_id": "ars-pi-e2e-gpt-5.6-sol-20260803T051847Z",
+  "source_root": "<redacted-temporary-run-root>",
+  "repository_head": "682b158ea02b6bdcc1f70b1ba577f237645bf7f5",
+  "runtime": {
+    "pi_version": "0.83.0",
+    "provider": "openai-codex",
+    "model": "gpt-5.6-sol",
+    "thinking": "high",
+    "pipeline_version": "3.19.0"
+  },
+  "execution_disclosure": "sequential same recovered Pi/model context; no independent or cross-model roles claimed",
+  "terminal_state": {
+    "pipeline_state": "completed",
+    "stage6_status": "completed",
+    "checkpoint_confirmed": true,
+    "terminal_acknowledgement_received": true,
+    "completed_at": "2026-08-03T18:19:42Z",
+    "round_trips": "36/36"
+  },
+  "privacy": {
+    "scenario": "synthetic test scenario; generated scholar declarations are not maintainer biography",
+    "removed": [
+      "maintainer-identifying fields (none found)",
+      "Pi session/event/request identifiers",
+      "exact human attestation wording",
+      "machine-local paths except generic redaction tokens",
+      "document binaries and redundant translated/build outputs"
+    ],
+    "credentials_found": false
+  },
+  "files": [
+    {
+      "curated_path": "README.md",
+      "source_path": "curator-authored privacy-minimized evidence guide",
+      "sha256": "3da15ecfd8896bc35a0bf5af218fb64b9a69b08cad0fad160642b50c66169e95",
+      "bytes": 8404,
+      "transformation": "curator-authored"
+    },
+    {
+      "curated_path": "model-portability.md",
+      "source_path": "curator-authored comparison of retained run evidence with academic-pipeline/examples/full_pipeline_example.md",
+      "sha256": "2a4316b250cc7fe4cd57ed4d42a0185bddf1e10d4dfd2fc4af0ddbce7cecf8a6",
+      "bytes": 7151,
+      "transformation": "curator-authored from linked observed evidence; baseline explicitly treated as illustrative"
+    },
+    {
+      "curated_path": "run-log.md",
+      "source_path": "curator-authored projection from persisted terminal state and retained run artifacts",
+      "sha256": "5d6e43ed109a26412048ca192f7622435d0186274d760796cb35e77fdcfa3bad",
+      "bytes": 7470,
+      "transformation": "privacy-minimized chronological projection; not a verbatim transport transcript"
+    },
+    {
+      "curated_path": "runtime/ars-pi-doctor.txt",
+      "source_path": "doctor-rpc.jsonl custom message",
+      "sha256": "47731cf8e217769f22844cc8a6ba8174c0d01279997e3d71fff46eceae30871e",
+      "bytes": 511,
+      "transformation": "capability report extracted; machine-specific repository path replaced"
+    },
+    {
+      "curated_path": "runtime/package-load-validation.json",
+      "source_path": "fresh post-run root/nested Pi RPC registry and system-prompt activation probes",
+      "sha256": "3708902123deebf65757eee09182b63b6670c267b8f2cb6849174020ec6ab017",
+      "bytes": 3073,
+      "transformation": "generated validation result; no model response used as evidence"
+    },
+    {
+      "curated_path": "stage-01/revised-research-report.md",
+      "source_path": "pipeline-workspace/stage-01-research/19-phase6-revised-research-report.md",
+      "sha256": "cee08d1b349854631a5634302852e58c16e0c51f480042fd467d1daa76e3abab",
+      "bytes": 38561,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-02.5/focused-reverification-report.md",
+      "source_path": "pipeline-workspace/stage-02.5-integrity-pre-review/04-focused-reverification-report.md",
+      "sha256": "614081675275d7b21b395c191b88545e779c130a626e91c941888b83b9c93a19",
+      "bytes": 6410,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-02/course-paper-draft.md",
+      "source_path": "pipeline-workspace/stage-02-write/11-complete-stage2-course-paper-draft.md",
+      "sha256": "953badfa35dab3ae6bcd183918b5d319b8575e22df944a3f0fa2ac54429f6507",
+      "bytes": 34191,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-03/editorial-decision-and-revision-roadmap.md",
+      "source_path": "pipeline-workspace/stage-03-review/editorial/editorial-decision-and-revision-roadmap.md",
+      "sha256": "972dafc30787c2407508f02f8c01eb4deab45e289087d930bb158a7e37f1ffa6",
+      "bytes": 11612,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-03p/verification-review-report.md",
+      "source_path": "pipeline-workspace/stage-03p-rereview/08-verification-review-report.md",
+      "sha256": "926d7702847ff45d00a0157d71bede3119b291e80623988679f3518ad1087e46",
+      "bytes": 7338,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-04.5/final-integrity-verification-report.md",
+      "source_path": "pipeline-workspace/stage-04.5-final-integrity/13-final-integrity-verification-report.md",
+      "sha256": "d68db78036962dbcb6588a888343e5b611647203835ff263b7ef22e81be0bc9f",
+      "bytes": 12623,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-04/response-to-reviewers.md",
+      "source_path": "pipeline-workspace/stage-04-revision/07-response-to-reviewers.md",
+      "sha256": "c53ed200fe21b5ccc43a17ad3339995112ea53efd10adc5b03c4fe49a4feafd3",
+      "bytes": 5579,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-04p/validation-report.md",
+      "source_path": "pipeline-workspace/stage-04p-rerevision/06-stage4p-validation-report.md",
+      "sha256": "02584274c834b082b132f48d57033efa054257f5d8f57848c377b1a265cfd803",
+      "bytes": 4200,
+      "transformation": "one machine-local temporary virtualenv path replaced with <temporary-virtualenv>; otherwise unchanged"
+    },
+    {
+      "curated_path": "stage-05/cite-time-gate-pass.json",
+      "source_path": "pipeline-workspace/stage-05-finalize/phase7-format/06-cite-time-hard-gate-pass.json",
+      "sha256": "90c75b0cc4748461bfae6b235c28d2e75fe700b44b87203d5ad1aca5c093735f",
+      "bytes": 1240,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-05/cite-time-gate-refusal-report.md",
+      "source_path": "pipeline-workspace/stage-05-finalize/01-cite-time-provenance-gate-report.md",
+      "sha256": "4a92cb245675846e68ae105f8ce1478e8db6c13048e81f74781443d2b8ddedc2",
+      "bytes": 3052,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-05/cite-time-gate-refusal.json",
+      "source_path": "pipeline-workspace/stage-05-finalize/phase7-format/03-cite-time-hard-gate.json",
+      "sha256": "6385ad669ab385b721de3e48c262f791f2e56e168893be761973c70406cf6f86",
+      "bytes": 1078,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-05/full-checkpoint.md",
+      "source_path": "pipeline-workspace/stage-05-finalize/05-stage5-output-package-and-full-checkpoint.md",
+      "sha256": "78d05d2c7bca23e3928931892ca1e86ac4bb637ed67bb93943410dffa2fd0e02",
+      "bytes": 3631,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-05/package/format-validation.md",
+      "source_path": "pipeline-workspace/stage-05-finalize/package/format_validation.md",
+      "sha256": "1400fa0809c62fea37eaac92c8302e97314c5904c36a0e67588ba25c8108cd14",
+      "bytes": 2877,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-05/package/paper.md",
+      "source_path": "pipeline-workspace/stage-05-finalize/package/paper.md",
+      "sha256": "64ac9f2d5794f1f920136259869fe6a8f4c248f74a5d052f8d349818796ad219",
+      "bytes": 36585,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-06/process-record-en.md",
+      "source_path": "pipeline-workspace/stage-06-process-summary/paper_creation_process_en.md",
+      "sha256": "d378ec4100d64c2041083352af09a4d1b5f0acfb778ec047590d310a861f7ea3",
+      "bytes": 17935,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "stage-06/terminal-checkpoint.md",
+      "source_path": "pipeline-workspace/stage-06-process-summary/03-stage6-artifact-index-and-terminal-checkpoint.md",
+      "sha256": "11f6dd94e360f3119a10b9298ed585243fe3a93908a994772b621843523e3d87",
+      "bytes": 3023,
+      "transformation": "byte-for-byte copy"
+    },
+    {
+      "curated_path": "state/pipeline-state.json",
+      "source_path": "pipeline-workspace/state/pipeline-state.json",
+      "sha256": "4422a6db8075023e02612945e06e6072c1ebcd2ec6913ec9ccadfb7216a76933",
+      "bytes": 22844,
+      "transformation": "byte-for-byte copy"
+    }
+  ],
+  "excluded": [
+    {
+      "pattern": "provider RPC streams and full Pi session JSONL",
+      "reason": "multi-gigabyte transport detail, internal identifiers, and machine-local noise; not needed for stage or terminal proof"
+    },
+    {
+      "pattern": "exact human attestation event and human-read ledger",
+      "reason": "privacy minimization; retained refusal/PASS records preserve aggregate 21-source and 79-marker gate state"
+    },
+    {
+      "pattern": "DOCX, PDF, LaTeX, translated process copy, and build-validation outputs",
+      "reason": "redundant binary/rendered evidence; retained Markdown and format/checkpoint records preserve outcomes and hashes"
+    },
+    {
+      "pattern": "Material Passport and submission-verifier snapshot/report",
+      "reason": "verbose duplicate state and synthetic scenario declarations; retained terminal state, final paper, checkpoint, and integrity reports are sufficient"
+    },
+    {
+      "pattern": "downloaded source bodies, raw retrieval payloads, and intermediate caches",
+      "reason": "bulky inputs are not required to establish wrapper execution, degraded disclosure, stage progression, or terminal state"
+    }
+  ],
+  "integrity_notes": [
+    "Files marked byte-for-byte have matching source and curated SHA-256 values from the retained local audit workspace.",
+    "The Stage 4p validation report has one explicitly recorded machine-path redaction.",
+    "inventory.json does not hash itself; SHA256SUMS covers inventory.json and every other curated file except SHA256SUMS."
+  ]
+}

examples/pi/model-portability.md

@@ -0,0 +1,55 @@
+# Model portability: `openai-codex/gpt-5.6-sol`
+
+## Comparison boundary
+
+This note records one observed Pi run with `@earendil-works/pi-coding-agent` 0.83.0 and `openai-codex/gpt-5.6-sol`. The comparison target is [`academic-pipeline/examples/full_pipeline_example.md`](../../academic-pipeline/examples/full_pipeline_example.md), which documents the intended Claude-oriented experience but is an illustrative conversation log, not a paired empirical run. Differences below therefore must not be interpreted as a controlled model benchmark.
+
+The run also differs in ARS version, available tools, user decisions, and evidence availability. A behavior is attributed specifically to the Pi runtime only when the execution record establishes that boundary; other differences are reported as observations without causal attribution.
+
+## Tested configuration
+
+| Field | Value |
+| --- | --- |
+| Pi | `@earendil-works/pi-coding-agent` 0.83.0 |
+| Provider/model | `openai-codex/gpt-5.6-sol` |
+| Thinking | high |
+| ARS | 3.19.0 |
+| Orchestration | none exposed inside Pi |
+| Web retrieval | `/skill:websearch` exposed |
+| Claude hooks | unavailable |
+| Execution | sequential in one model context, with persisted-session recovery |
+
+No Claude, Opus, or Sonnet run was executed as part of this evidence. “Claude baseline” below means the repository's documented illustrative behavior, not a fresh Claude result.
+
+## Observed divergences
+
+| Area | Documented Claude-oriented baseline | Observed Pi/model behavior | Evidence and interpretation |
+| --- | --- | --- | --- |
+| Model routing | ARS is calibrated for Opus architecture/review and Sonnet execution; command frontmatter can request Sonnet. | Every role used `openai-codex/gpt-5.6-sol`; Claude `model: sonnet` metadata was ignored. | Runtime difference. See [`runtime/ars-pi-doctor.txt`](runtime/ars-pi-doctor.txt) and [`state/pipeline-state.json`](state/pipeline-state.json). |
+| Agent execution | Example describes parallel independent reviewer agents and specialist handoffs. | Specialist roles and five reviewer perspectives ran sequentially in the same recovered context. No independent reviewer, writer, verifier, formatter, observer, or cross-model judgment is claimed. | Load-bearing runtime divergence. Actual disclosure appears in [`stage-03/editorial-decision-and-revision-roadmap.md`](stage-03/editorial-decision-and-revision-roadmap.md), [`stage-04.5/final-integrity-verification-report.md`](stage-04.5/final-integrity-verification-report.md), and [`stage-06/process-record-en.md`](stage-06/process-record-en.md). |
+| Write-scope enforcement | Claude distribution wires a `PreToolUse` write-scope guard. | Claude hooks did not execute. Scope discipline was prompt-level and artifact-checked after writes. | Load-bearing runtime divergence; no deterministic Claude hook boundary is claimed. |
+| Review independence | Example reports independent review reports before editorial synthesis. | Reports were structurally separated but shared one model context. The final self-reflection mechanically classified sycophancy risk as HIGH because both explicit Devil's Advocate concessions were accepted. | Same-context execution may correlate errors. This is not evidence that independent Claude reviewers would reach different verdicts. |
+| Re-review outcome | Illustrative example reaches Minor Revision after one revision and proceeds to finalization. | Stage 3′ returned Major Revision, causing focused Stage 4′ re-revision before final integrity. | Observed output divergence; cannot be attributed solely to model because manuscript content and current ARS contracts differ. See [`stage-03p/verification-review-report.md`](stage-03p/verification-review-report.md). |
+| Interaction length | Example abbreviates several Socratic rounds and presents a compact happy path. | Run consumed all 36 advisory round trips, crossed 20 mandatory/full checkpoints, and required persisted-session recovery. | Not a direct efficiency comparison because baseline rounds are intentionally omitted. It does show materially higher operational friction in this run. |
+| Transport/context | Example does not describe stream recovery or compaction. | Stage 3′ encountered stalled model streams; execution resumed from persisted artifacts, state, and hashes across two Pi sessions. | Observed provider/runtime behavior. Continuity was disclosed rather than represented as cognitive independence. |
+| Length interpretation | Example targets a 5,200-word paper. | Model initially treated “6,000 words” too literally for Traditional Chinese; scholar corrected target to approximately 6,000 Chinese characters. | Observed model behavior corrected through checkpoint interaction. |
+| Claim discipline | Example presents a smooth draft/review sequence. | Model initially overclaimed that all findings were verified and carried an incorrect O7 locator. Stage 2.5 correctly failed, a narrow correction was applied, and focused re-verification passed. | Observed error plus successful gate behavior. See [`stage-02.5/focused-reverification-report.md`](stage-02.5/focused-reverification-report.md). |
+| Human-read provenance | Example proceeds directly from accepted re-review to formatting. | Stage 5 refused conversion despite 21/21 technical source verification because no scholar-read attestation existed. It resumed only after explicit full-text attestations resolved 79/79 markers. | Current ARS gate behavior exercised in Pi. See [`stage-05/cite-time-gate-refusal-report.md`](stage-05/cite-time-gate-refusal-report.md) and [`stage-05/cite-time-gate-pass.json`](stage-05/cite-time-gate-pass.json). |
+| Final formats | Example emits Markdown, DOCX, LaTeX, and paper PDF. | Run emitted Markdown and DOCX. Scholar explicitly declined paper LaTeX/PDF; bilingual process-record PDFs were still produced. | User decision, not a model or Pi limitation. |
+
+## Behavior that remained effective
+
+Despite the divergences, this run completed the full current pipeline:
+
+- research, writing, integrity, review, revision, re-review, re-revision, final integrity, finalization, and process summary all persisted terminal states;
+- no mandatory/full checkpoint was automatically skipped;
+- Stage 2.5 produced a real FAIL followed by a bounded correction and PASS;
+- Stage 5 produced a real provenance refusal followed by a user-authorized PASS;
+- deterministic patch/apply evidence and hashes survived transport recovery;
+- degraded sequential execution was disclosed at runtime rather than only predicted by documentation.
+
+These observations show functional execution on the tested model. They do not establish parity with Claude or independence equivalent to a multi-agent panel.
+
+## Portability conclusion
+
+`openai-codex/gpt-5.6-sol` completed ARS 3.19.0 on Pi, but the observed run diverged materially from the documented Claude-oriented experience in model routing, agent independence, hook enforcement, transport stability, re-review path, and several model errors corrected by gates. The strongest supported claim is therefore “completed with disclosed degraded same-context execution,” not Claude runtime parity.

examples/pi/run-log.md

@@ -0,0 +1,65 @@
+# Curated academic-pipeline run log
+
+This is a chronological, human-readable projection of persisted pipeline state and generated artifacts. It is not a verbatim replacement for the excluded RPC streams. Stage status, hashes, timestamps, and checkpoint facts come from [`state/pipeline-state.json`](state/pipeline-state.json) and linked run artifacts. “Scholar” language inside generated artifacts refers to the synthetic test operator, not the Pi port maintainer's identity or biography.
+
+## Invocation and capability boundary
+
+The run used the original scenario prompt through `/ars-full`:
+
+```text
+I want to write a paper on the impact of declining birthrate on enrollment
+strategies of private universities in Taiwan. Help me from start to finish.
+This is my first time writing an academic paper and I'm not sure about the process.
+```
+
+Pi was started with discovered extensions, skills, prompt templates, and context files disabled. Only the wrapper, four original ARS skills, original ARS commands, and `/skill:websearch` were explicitly loaded. [`runtime/ars-pi-doctor.txt`](runtime/ars-pi-doctor.txt) recorded:
+
+```text
+Orchestration: none; sequential/degraded mode
+Web retrieval: skill: /skill:websearch
+Claude hooks: unavailable in Pi; write-scope enforcement remains prompt-level
+```
+
+The external RPC/workflow driver delivered synthetic test-operator checkpoint responses and recovered transport. It was not exposed inside Pi as a tool, skill, reviewer, or orchestration capability.
+
+## Chronology
+
+| Order | Stage | Persisted result | Checkpoint/operator action | Representative evidence |
+| ---: | --- | --- | --- | --- |
+| 1 | Research | Completed. RQ brief, methodology, bounded search, source verification, synthesis, reviewed and revised research report. | Test operator prohibited invented interviews, participants, original-data analysis, statistical results, and strategy-effect rankings; required paired registration-rate and actual-freshman outcomes. Stage checkpoint confirmed. | [`stage-01/revised-research-report.md`](stage-01/revised-research-report.md) |
+| 2 | Write | Completed in plan-to-full mode. Fresh course-paper draft, bilingual abstracts, and citation check emitted. | Test operator required a fresh paper rather than relabeling Stage 1 and corrected the target from “6,000 words” to approximately 6,000 Chinese characters. Outline, argument, and draft checkpoints confirmed. | [`stage-02/course-paper-draft.md`](stage-02/course-paper-draft.md) |
+| 3 | Integrity 2.5 | Round 1 **FAIL**, bounded correction, Round 2 **PASS**. | Test operator accepted the O7 locator mismatch and authorized only the verified URL and stale process-status corrections. Compliance WARN acknowledged without converting it to PASS. | [`stage-02.5/focused-reverification-report.md`](stage-02.5/focused-reverification-report.md) |
+| 4 | Review | Five role reports executed sequentially in one context. Editorial decision: **Major Revision**. | Reviewer configuration and decision confirmed. Test operator completed revision coaching and reframed contribution around measurement core, scope stratification, and prototype demotion. | [`stage-03/editorial-decision-and-revision-roadmap.md`](stage-03/editorial-decision-and-revision-roadmap.md) |
+| 5 | Revise | Three deterministic patch rounds; apply chain passed; 10/10 response coverage. | Test operator prohibited new search, required patch-only revision, and confirmed the inspection checkpoint. | [`stage-04/response-to-reviewers.md`](stage-04/response-to-reviewers.md) |
+| 6 | Re-review 3′ | Hash-bound contract and checker passed; decision remained **Major Revision**. Transport recovery resumed the same academic run. | Mandatory decision accepted. Residual coaching skipped; no new search authorized. | [`stage-03p/verification-review-report.md`](stage-03p/verification-review-report.md) |
+| 7 | Re-revise 4′ | One authorized block changed; 115/116 blocks preserved byte-identically; validation passed. | Test operator inspected focused diff and accepted the token-conservation advisory while preserving the declared limitation. FULL checkpoint confirmed. | [`stage-04p/validation-report.md`](stage-04p/validation-report.md) |
+| 8 | Final integrity 4.5 | **PASS**. 21/21 references, 58/58 citation contexts, 41/41 claim clusters, data reproduction, originality sampling, and failure-mode checks completed. | Test operator inspected report/checklist, accepted 12 advisories and nonblocking compliance WARN, then authorized Stage 5. | [`stage-04.5/final-integrity-verification-report.md`](stage-04.5/final-integrity-verification-report.md) |
+| 9 | Finalize | Initial cite-time gate **REFUSE** because 79 markers lacked human-read attestation. After the required human attestation, 79/79 resolved to `OK`; Markdown and DOCX emitted. | Exact attestation wording and internal identifiers are omitted. Test operator selected APA 7, declined paper LaTeX/PDF, retained placeholders and warnings, and confirmed FULL checkpoint. | [`stage-05/cite-time-gate-refusal-report.md`](stage-05/cite-time-gate-refusal-report.md), [`stage-05/cite-time-gate-pass.json`](stage-05/cite-time-gate-pass.json), [`stage-05/package/paper.md`](stage-05/package/paper.md) |
+| 10 | Process summary | Bilingual records and PDFs were validated during the run. Pipeline and Stage 6 persisted `completed`. | Terminal checkpoint was presented as nonterminal until acknowledgement arrived. Successful retry and acknowledgement persisted; 36/36 advisory round trips used. | [`stage-06/process-record-en.md`](stage-06/process-record-en.md), [`stage-06/terminal-checkpoint.md`](stage-06/terminal-checkpoint.md), [`state/pipeline-state.json`](state/pipeline-state.json) |
+
+## Actual degraded-mode disclosure
+
+Degradation was not inferred only from `/ars-pi-doctor`. Generated runtime artifacts repeatedly disclosed it. For example, the Stage 3 editorial record states that the five roles ran sequentially in one Pi session, without independent context windows or genuine cognitive paper blindness. Stage 3′, Stage 4′, Stage 4.5, and Stage 6 repeat the same boundary.
+
+This establishes that fallback behavior actually fired. It does not establish equivalence to independent Claude agents.
+
+## Checkpoint and terminal audit
+
+Persisted state records:
+
+- mandatory/full checkpoints automatically skipped: 0;
+- Stage 2.5 integrity history: FAIL then PASS;
+- Stage 3 decision: Major Revision;
+- Stage 3′ decision: Major Revision;
+- Stage 4.5 final integrity: PASS and scholar-confirmed;
+- Stage 5 cite-time provenance: REFUSE then PASS after explicit attestation;
+- Stage 6: completed;
+- terminal checkpoint: confirmed;
+- terminal acknowledgement: received;
+- pipeline state: completed.
+
+The first terminal-response transport attempt was not consumed during compaction. Internal transport/request identifiers are omitted from this public bundle; the persisted terminal state controls the completion claim.
+
+## Completeness and exclusions
+
+Representative output from every executed stage is committed, along with final Markdown, process record, terminal state, inventory, and hashes. Provider streams, session JSONL, internal identifiers, exact personal attestation text, downloaded source bodies, redundant binaries/translations, and intermediate caches are excluded. Their omission is disclosed in [`inventory.json`](inventory.json); no excluded file is required to establish the persisted stage sequence or final terminal state.

examples/pi/runtime/ars-pi-doctor.txt

@@ -0,0 +1,12 @@
+ARS Pi doctor
+Repository: <repository-root>
+Orchestration: none; sequential/degraded mode
+Web retrieval: skill: /skill:websearch
+Python: Python 3.14.4
+PyYAML: PyYAML 6.0.3
+Pandoc: pandoc 3.10
+Tectonic: Tectonic 0.17.0
+Sandbox: not detected
+Claude hooks: unavailable in Pi; write-scope enforcement remains prompt-level
+
+Curation note: captured before the pipeline run with /ars-pi-doctor. Only the machine-specific repository path was replaced with <repository-root>; capability and version lines are unchanged.

examples/pi/runtime/package-load-validation.json

@@ -0,0 +1,97 @@
+{
+  "pi_version": "0.83.0",
+  "post_run_wrapper_validation": true,
+  "root_invocation": "pi --mode rpc --no-extensions --no-skills --no-prompt-templates --no-context-files -e . --offline",
+  "nested_invocation": "pi --mode rpc --no-extensions --no-skills --no-prompt-templates --no-context-files -e ./pi --offline",
+  "root_package": {
+    "success": true,
+    "prompt_templates": [
+      "ars-3w",
+      "ars-abstract",
+      "ars-cache-invalidate",
+      "ars-citation-check",
+      "ars-disclosure",
+      "ars-format-convert",
+      "ars-full",
+      "ars-lit-review",
+      "ars-mark-read",
+      "ars-outline",
+      "ars-plan",
+      "ars-rebuttal-audit",
+      "ars-reviewer",
+      "ars-revision",
+      "ars-revision-coach",
+      "ars-unmark-read"
+    ],
+    "skills": [
+      "skill:academic-paper",
+      "skill:academic-paper-reviewer",
+      "skill:academic-pipeline",
+      "skill:deep-research"
+    ],
+    "wrapper_commands": [
+      "ars-pi-doctor",
+      "ars-pi-start",
+      "ars-pi-stop"
+    ],
+    "stderr_empty": true
+  },
+  "nested_package": {
+    "success": true,
+    "prompt_templates": [
+      "ars-3w",
+      "ars-abstract",
+      "ars-cache-invalidate",
+      "ars-citation-check",
+      "ars-disclosure",
+      "ars-format-convert",
+      "ars-full",
+      "ars-lit-review",
+      "ars-mark-read",
+      "ars-outline",
+      "ars-plan",
+      "ars-rebuttal-audit",
+      "ars-reviewer",
+      "ars-revision",
+      "ars-revision-coach",
+      "ars-unmark-read"
+    ],
+    "skills": [
+      "skill:academic-paper",
+      "skill:academic-paper-reviewer",
+      "skill:academic-pipeline",
+      "skill:deep-research"
+    ],
+    "wrapper_commands": [
+      "ars-pi-doctor",
+      "ars-pi-start",
+      "ars-pi-stop"
+    ],
+    "stderr_empty": true
+  },
+  "system_prompt_scope": {
+    "claim_scope": "prompts submitted while Pi is idle",
+    "ordinary_idle_prompt_ars_text_active": false,
+    "ordinary_idle_prompt_package_ars_skills_visible": false,
+    "unrelated_colliding_skill_preserved": true,
+    "symlink_spelled_xml_escaped_skill_hiding_harness": "PASS",
+    "missing_skill_location_preserved": true,
+    "adjacent_skill_block_preserved_on_alternate_formatting": true,
+    "idle_ars_command_ars_text_active": true,
+    "idle_direct_skill_ars_text_active": true,
+    "idle_manual_start_enables_automatic_selection": true,
+    "idle_stop_hides_ars_text_and_skills_on_next_run": true,
+    "resume_persistence_harness": "PASS",
+    "tree_navigation_state_recovery": true,
+    "argument_safe_script_path_rewrite": true,
+    "single_pass_argument_placeholder_substitution": true,
+    "node_regression_tests": "9/9 PASS",
+    "real_pi_rpc_idle_prompt_scope_probes": "5/5 PASS",
+    "documented_residuals": {
+      "mid_stream_prompt": "queued ARS requests retain the in-flight run's existing system prompt",
+      "mid_stream_stop": "state changes immediately, but the in-flight system prompt remains unchanged",
+      "rpc_steer_follow_up": "these methods bypass the input event and do not activate the wrapper"
+    }
+  },
+  "result": "PASS"
+}

examples/pi/stage-01/revised-research-report.md

@@ -0,0 +1,325 @@
+# 少子化情境下臺灣私立大學招生策略證據之可判定性：公開資料趨勢與最低證據稽核框架
+
+**學生：**［學生姓名］  
+**系所：**［系所名稱］  
+**課程：**［課程名稱］  
+**授課教師：**［教師姓名］  
+**日期：**2026年8月3日  
+
+## 中文摘要
+
+臺灣高等教育在生源縮減下持續面臨招生與校務治理壓力，但註冊率上升不必然代表新生增加，校方公開的招生作法也不等於成效證據。本研究探討：公開可驗證資料在何種程度上能支持私立一般大學與私立技專校院對招生策略成效的管理判斷？研究採混合次級證據設計，整合有限結構化文獻、政策與文件分析，以及教育部公開資料的描述性分析。官方資料依私立一般大學與私立技專校院分層，檢視109至113學年度新生註冊率、實際新生人數及分母構成。研究另區分「策略存在」、「結果共同變化」與「足以判斷效果」三種主張層級。
+
+結果顯示，109至113學年度私立一般大學實際新生數由68,505人降至58,768人，註冊率由92.95%降至90.02%；私立技專校院實際新生數由96,155人降至79,451人，但註冊率由76.27%升至77.29%。後者顯示名額與分母收縮可與實際人數下降同時發生。在本次取得並完成原文檢查的有限語料中，僅Lin與Yang對臺灣技專校院系所調整提出直接關聯證據；該研究不能延伸為因果判斷，也不能推及私立一般大學。作為公開材料可取得性測試的六校目的性稽核樣本，只能證明特定校方主張在檢索時可見，不能推論盛行率、實施或效果。
+
+本研究據此提出「最低公開證據稽核與證據取得篩檢」。它是校務決策前的外部證據最低檢查層，不能取代學校內部的策略時間、觸及、成本與學生結果資料。現有公開證據不足以對任何策略提出正式建議；「證據生成試辦」只表示可考慮用小規模、可逆且預先規劃的設計補足證據，不表示策略可能有效。此篩檢僅具有程序一致性與可追溯性，不代表已通過效度、信度、預測力或一般化驗證。
+
+**關鍵詞：**少子化、私立大學、技專校院、新生註冊率、招生策略、證據分級、校務治理
+
+## Abstract
+
+Taiwan’s shrinking student cohorts place sustained pressure on private higher education, yet a rising freshman registration rate does not necessarily indicate growth in actual enrollment, and a publicly visible recruitment initiative does not itself demonstrate effectiveness. This study asks to what extent publicly verifiable evidence can support managerial judgments about recruitment-strategy effectiveness in Taiwan’s private general universities and private technological and vocational institutions. A mixed secondary-evidence design combined a bounded structured review of academic, policy, and institutional documents with descriptive analysis of Ministry of Education data for academic years 109–113. Freshman registration rates, actual freshman counts, and denominator components were examined separately by institutional stratum.
+
+Actual freshmen in private general universities declined from 68,505 to 58,768, while the registration rate changed from 92.95% to 90.02%. In private technological and vocational institutions, actual freshmen declined from 96,155 to 79,451 even as the rate rose from 76.27% to 77.29%. Thus, capacity and denominator contraction can coexist with declining enrollment. Within the bounded retrieved corpus, one full-text study directly linked defined strategies to outcomes; its findings were associative and confined to Taiwan TVET. Six purposively selected webpages documented issuer-authored claims only. The resulting minimum public-evidence audit distinguishes visibility, co-movement, and effect judgment. No strategy merits recommendation from this public corpus. The screen cannot replace institutional data, and “pilot” means evidence generation, not expected benefit. It is procedural, unvalidated, and non-generalizable.
+
+**Keywords:** demographic decline, private higher education, TVET, freshman registration rate, student recruitment, evidence-informed management
+
+# 緒論
+
+## 研究背景與問題
+
+少子化對高等教育的影響，不只是學生總量減少，也涉及招生名額、校型差異、系所調整、境外生政策與學校退場等制度變化。立法院的政策研析已將出生人口下降與學校生源不足連結為長期公共問題（盧延根，2022）<!--ref:O7--><!--anchor:section:我國少子化問題之概況-->。教育部公開資料則顯示，大專校院新生註冊率的統計範圍與公式曾在103、107及109學年度調整，因此跨期比較不能忽略定義變動（教育部，2026a）<!--ref:O1--><!--anchor:section:說明-->。
+
+私立學校對學費與招生的依賴，使生源縮減更容易轉化為財務與治理壓力。比較研究指出，人口停滯或下降可能優先衝擊缺乏明確差異化優勢的私立次部門，但私立高等教育的「衰退」可能指市場占比、絕對人數或特定次部門下降，不能由少數案例推論整體必然衰退（Levy, 2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->。臺灣早期政策分析也指出，私立大學承受的招生與財務壓力通常高於公立大學，並提出治理、特色發展、產學合作、國際化與組織調整等回應方向（Kuo, 2016）<!--ref:A2--><!--anchor:section:4.Conclusions-->。然而，「可採取的政策選項」與「已證明有效的招生策略」是不同命題。
+
+註冊率會隨核定名額、保留資格或境外新生改變，因而可能在實際新生下降時改善。校方網頁也只能證明作法被公開描述，通常欠缺實施日期、強度、比較、成本與反事實。若不區分主張層級，管理者可能把可見度當盛行率、把同期變動當因果，或由單一比率誤判生源變化。
+
+## 研究目的與研究問題
+
+本研究目的不是找出「最有效」的招生策略，而是判定公開證據允許何種推論，並建立可追溯的最低公開證據稽核與證據取得篩檢。此篩檢是校務內部判斷前的最低外部證據層，不是完整管理決策模型。主要研究問題為：
+
+> 在臺灣少子化情境下，公開可驗證資料在何種程度上能支持私立一般大學與私立技專校院對招生策略成效的管理判斷？
+
+三個子問題如下：
+
+1. 在定義可比的109至113學年度，兩類私立校院的新生註冊率與實際新生人數呈現何種趨勢？原設計亦規劃區域與規模分層，但本研究僅完成校型分層；區域與穩定規模趨勢仍屬未完成部分。
+2. 政策文件、學校材料及學術研究分別能支持策略存在、結果共同變化或效果判斷中的哪一層主張？
+3. 如何依證據強度、雙重招生指標，以及品質、適配、公平與成本風險，形成公開證據已達門檻、可設計證據生成試辦、延後判斷或停止採用／擴大策略的暫定分類？
+
+## 研究重要性
+
+本研究把外部證據檢查置於策略選擇之前。最低稽核不能取代校內招生漏斗、預算、執行強度、留存與學生結果；合格內部資料可能支持更強判斷。其用途是減少分母驅動的表面改善、宣傳材料驅動的盛行率推論，以及把關聯提升為因果等錯誤。
+
+# 文獻回顧與分析架構
+
+## 人口收縮、私立部門與制度分層
+
+高參與體系仍可能因公私立、學校階層、學科與學費形成不均等機會（Marginson, 2016）<!--ref:A7--><!--anchor:section:Abstract-->，故填滿名額不能取代品質與公平。私立部門可能較易受人口收縮影響，但聲譽、區位、需求、公共競爭與政策也會改變結果（Levy, 2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->；少子化不是個別校表現的單一原因。
+
+臺灣政策論述把合併、退場、特色發展、產學合作與國際招生列為調整方向（Kuo, 2016）<!--ref:A2--><!--anchor:section:4.Conclusions-->，但未提供可比較效果。另一政策評估發現投入增加不保證更佳表現；因其研究頂尖大學計畫而非私校招生，本文只轉用「投入不能代替績效辨識」的警示（Chang et al., 2009）<!--ref:A9--><!--anchor:section:Abstract-->。
+
+## 招生選擇機制與移植限制
+
+馬來西亞私立大學研究提出學習環境、影響者、服務、成本、設施、社交與區位等選擇因素（Padlee et al., 2010）<!--ref:A4--><!--anchor:section:Abstract-->；波札那研究另辨識課外活動、就業、品質、同儕、宣傳、財務支持與成本（Baliyan & Mokoena, 2024）<!--ref:A5--><!--anchor:section:6.Conclusion_and_Recommendations-->。兩者皆為異地學生知覺研究，不能證明臺灣策略提高註冊率。
+
+臺灣一所私立大學國際學程的11人焦點團體指出，互動教學、同儕支持、課程資訊、社會融入與國際環境影響新生經驗（Chalapati et al., 2018）<!--ref:A3--><!--anchor:section:Conclusion-->。小樣本不能代表私校，卻提醒國際招生須配置語言、教學與融入支持，不能只追求人數。
+
+## 系所調整的直接關聯證據
+
+Lin與Yang（2026）<!--ref:A1--><!--anchor:section:4.Methodology-->整合全臺技專校院行政資料，分析2020至2024年間74所機構的1,087件系所調整案例，包括更名、合併、新增、停招或關閉及名額擴充。研究以描述性趨勢與固定效果模型檢視新生註冊結果。結果顯示，系所更名是唯一持續與招生改善相關的調整類型，且私立機構的關聯較為明顯；合併通常沒有正向招生變化，關閉多發生在長期低註冊系所，新增與擴額在醫療健康及部分應用科技領域呈現較佳表現（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->。
+
+該研究不能支持「更名造成招生改善」。調整決策可能回應既有趨勢，固定效果無法控制所有隨時間變動的混淆，觀察期較短，合併或新設系所也可能有延遲效果；研究期間還與疫情後環境重疊。作者因此明確要求將結果解讀為關聯，而非因果（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:6.3.Methodological_and_Contextual_Constraints_on_Causal_Interpretation-->。此外，研究對象是技專校院，不能推及私立一般大學。
+
+## 教育品質、公平與利害關係人
+
+招生策略牽涉的不只是市場溝通。高參與體系中的學費與機構階層可能加深社會分層（Marginson, 2016）<!--ref:A7--><!--anchor:section:Abstract-->；大學與地方、產業及其他利害關係人的合作，也涉及任務選擇、治理與問責，過度擴張可能造成任務負荷（Jongbloed et al., 2008）<!--ref:A8--><!--anchor:section:Introduction-->。因此，註冊改善不能抵銷誤導招生、課名與課程不符、學生適配不良、債務負擔、退學風險、隱私違規或不可逆財務風險。
+
+## 主張、證據與風險架構
+
+本研究將證據判斷分為三層。第一層是策略存在，即可靠來源顯示某項措施被宣布或描述；第二層是結果共同變化或關聯，即策略分類與結果在明確單位與期間同時變動；第三層才是效果判斷，要求時間順序、實施強度、可信比較、基線或前趨勢、同期措施、延遲效果與替代解釋。其後再加入教育品質、學生適配、公平、法律、隱私與成本風險。這是依本研究暫定規則形成的最低公開證據稽核，用來顯示證據在哪一關停止及下一步需取得何種資料；它不預測策略成功，也不直接決定資源配置。
+
+# 研究方法
+
+## 研究設計
+
+本研究採務實取向的混合次級證據設計，包含有限結構化文獻、政策與文件分析，以及可驗證官方資料的描述性分析。它不是原創混合方法實證研究，也沒有訪談、問卷、實驗、人類參與者或新資料蒐集。國際研究僅用於提出機制、候選變項與風險條件，不替代臺灣效果證據。各校未公開的招生漏斗、預算、策略強度與學生結果不在資料範圍；缺少這些資料不表示校方無法以合格內部證據形成較強判斷。
+
+## 資料來源與檢索
+
+資料來源分為四類：教育部統計與校務資訊公開資料、政府政策文件、可確認書目身分且取得原文的學術研究，以及依事前規則抽取的學校網頁。檢索以繁體中文與英文進行，共執行43組網路查詢及16次Crossref/OpenAlex API查詢。原始發現紀錄382筆，經DOI與標題連結後形成316個候選群組。檢索過程保存查詢字串、UTC時間、URL、回傳數量、零結果、驗證碼、限流、403、404及Cloudflare等失敗。
+
+納入實質分析的學術原文共9篇；另有9筆僅完成書目或摘要驗證的資料，因無法檢查原文而不進入本文主張或參考文獻。Semantic Scholar多數請求回傳HTTP 429，因此未將「未匹配」錯誤編碼為不存在。兩個錯誤候選DOI，`10.1177/1745499920916532`與`10.1080/13538322.2020.1736512`，分別校正為`10.1177/1745499920910581`與`10.1080/08841241.2020.1728465`；原錯誤字串仍保留於稽核紀錄，但校正後兩篇因未取得原文，同樣未用於實質論證。
+
+## 官方資料與變項
+
+主要量化來源是教育部「大學、技專校院各系所註冊率」114學年度累積CSV快照，下載於2026年8月3日（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。分析保留109至113學年度、設立別為私立、校型為一般大學或技專校院，以及學士班（含四技）、二年制大學（二技）、二專、五專與學士後第二專長學士班（4+X）。研究所與附設進修學校類別不納入。
+
+系所資料欄位為核定名額A、保留入學資格人數B、總量內實際註冊人數C及境外新生實際註冊人數E。重算公式為：
+
+\[
+\text{新生註冊率}=\frac{C+E}{A-B+E}\times100\%
+\]
+
+教育部說明指出，109學年度起實際註冊人數納入特定擴充名額；113學年度欄位另納入《國家重點領域產學合作及人才培育創新條例》相關人數（教育部，n.d.-a）<!--ref:O8--><!--anchor:section:統計說明-->。因此，本文將109至113學年度視為較一致的分析窗，仍對113學年度的新增範圍保留版本註記。資料快照與衍生檔均保存SHA-256；後續官方修訂可能使數值改變，故結果應理解為該下載版本的可重製輸出，而非不可變的歷史檔。
+
+## 學校材料抽樣
+
+學校材料使用113學年度學生總數建立私立一般大學與私立技專校院兩個母層，再各分為小、中、大規模，每層抽取一校，共6校。擷取內容是各校一個系所或學程頁面在2026年8月3日可見的招生特色或課程主張。6個URL均成功回傳，但抽樣目的只是測試公開材料的可取得性。它不能估計策略在全體私校的盛行率，更不能連結至109至113學年度效果。
+
+## 分析步驟與品質控制
+
+分析先檢查定義、分母、年度、單位與缺值，再分別彙整兩類私校的A、B、C與E。所有納入欄位必須可解析為整數；遇到空白、省略號或非數字時，程式會停止，而不是自動補值。結果同時呈現註冊率與C+E實際新生數，避免單一比率掩蓋規模變化。
+
+文件分析矩陣記錄策略主張、可觀察證據、最高主張層級、替代解釋、風險與驗證狀態。暫定輸出為證據已達門檻、可設計證據生成試辦、延後判斷，或停止採用／擴大特定策略；最後一類不表示自動關校、停系或退場。單一研究者重新套用規則只能檢查可追溯性與一致性，不能稱為信度、效度或一般化。
+
+引文稽核區分標記存在、來源已取得及精確段落已定位。本稿有54組來源／定位標記與21個已取得並檢查脈絡的來源，但部分定位是內部標籤，PDF頁碼預檢未完成，不能宣稱所有引文皆完成頁級驗證（詳附錄C）。
+
+# 研究結果
+
+## 私立一般大學與私立技專校院趨勢
+
+表1呈現109至113學年度兩個私立校型的實際新生數與合計後重算註冊率。數值取自教育部UDB快照並由固定程式重算（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。該率是先按校型合計A、B、C、E再計算的合成校型比率，不是各校公開註冊率的平均值。
+
+**表1**  
+*109至113學年度私立校院大學部新生趨勢*
+
+| 學年度 | 私立一般大學校數 | 實際新生C+E | 註冊率 | 私立技專校院校數 | 實際新生C+E | 註冊率 |
+|---:|---:|---:|---:|---:|---:|---:|
+| 109 | 36 | 68,505 | 92.95% | 67 | 96,155 | 76.27% |
+| 110 | 36 | 66,799 | 91.02% | 64 | 90,159 | 76.25% |
+| 111 | 35 | 56,311 | 77.64% | 62 | 81,922 | 71.78% |
+| 112 | 34 | 55,898 | 83.09% | 59 | 78,785 | 73.68% |
+| 113 | 34 | 58,768 | 90.02% | 57 | 79,451 | 77.29% |
+
+*註。* 校數是當年度納入範圍中出現的機構數，並非平衡追蹤樣本。表中註冊率是依系所CSV之A、B、C、E先按校型與年度合計後重算的合成比率，不是機構公開率的算術平均；來源為教育部（2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。
+
+109至113學年度，私立一般大學實際新生減少9,737人（14.2%），合成率下降2.93個百分點；私立技專校院減少16,704人（17.4%），合成率卻增加1.02個百分點。後者核定名額A由123,366減至92,294（25.2%），顯示分母收縮可與實際人數下降並存。
+
+兩類校院的率在111學年度較低、113學年度回升，但本文不歸因於策略。名額、機構組成、境外生、學科需求、疫情後變化與同期政策皆可能解釋。各年校數不同，表1是重複橫斷面，不是同校前後比較。
+
+本文未建立跨年度區域、規模與退場交叉表。學生數資料只支持113學年度抽樣脈絡（教育部，2026c）<!--ref:O4--><!--anchor:section:資料說明-->，故第一子問題的區域與規模部分僅獲部分回答。
+
+## 招生策略的證據層級
+
+六校目的性稽核樣本的頁面呈現跨域課程、創新創業、影音專業、地方產業、實習、競賽、證照與醫療合作等校方主張（佛光大學，n.d.）<!--ref:U01--><!--anchor:section:招生資訊-->（中國醫藥大學，n.d.）<!--ref:U02--><!--anchor:section:最新消息-->（世新大學，n.d.）<!--ref:U03--><!--anchor:section:大學部-->（南開科技大學，n.d.）<!--ref:U04--><!--anchor:section:招生資訊-->（中國科技大學，n.d.）<!--ref:U05--><!--anchor:section:Latest_News-->（中華醫事科技大學，n.d.）<!--ref:U06--><!--anchor:section:招生特色-->。這些頁面只達「策略主張可見」層級。它們未提供一致的實施起點、預算、觸及人數、比較條件或連結後果。
+
+行政院2023年頁面則證明2024至2028年境外生擴大招生與留臺政策的存在及其政策目的（行政院，2023）<!--ref:O6--><!--anchor:section:Policy_content-->。該文件沒有提供私校策略效果評估，故不能用來宣稱國際招生已改善註冊。
+
+在本次取得並完成檢查的有限學術原文語料中，Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->是唯一直接比較臺灣技專校院調整類型與招生結果的研究。其結果讓私立技專校院的更名及部分需求導向新增或擴額達到關聯層級，但仍未達因果層級。合併沒有呈現穩定正向招生變化；關閉是長期低註冊後的結果型態，不應被當作招生增長策略。
+
+## 最低公開證據篩檢結果
+
+**表2**  
+*依本研究暫定規則形成的證據取得狀態示例*
+
+| 策略類別 | 公開證據最高層級 | 暫定證據取得狀態 | 啟動下一步前的必要條件 |
+|---|---|---|---|
+| 私立技專校院系所更名 | 關聯 | 延後；條件齊備才設計證據生成試辦 | 課名相符；成本上限、基線、比較、雙重結果、留存及停止條件 |
+| 私立技專校院合併 | 關聯但多數無正向變化 | 招生效果延後 | 治理目的另評；補成本、延遲效果、比較與學生權益資料 |
+| 需求導向新增或擴額 | 稀少情境化關聯 | 延後；條件齊備才設計證據生成試辦 | 驗證需求、師資設備與品質；設定成本、比較、分母及停止條件 |
+| 停招或關閉 | 長期低註冊之前兆 | 停止當作招生增長策略 | 不等於自動關校、停系或退場；治理評估另走學生權益程序 |
+| 國際招生 | 政策存在與間接機制 | 效果延後 | 補轉換、支持、留存、成本、法規及境外生分母資料 |
+| 高中連結、廣告、個人接觸 | 國外機制 | 延後 | 建立校級基線、觸及、招生漏斗、成本與比較 |
+| 獎助學金與財務支持 | 國外機制 | 延後 | 補淨價格、負擔、公平、留存及經常成本 |
+| 產學、地方合作、實習與就業定位 | 可見主張與間接機制 | 延後 | 確認實施，再取得實習品質、負擔及就業資料 |
+| 私立一般大學各策略 | 無直接關聯研究 | 延後 | 補策略時間、強度、比較、成本與學生結果 |
+
+*註。* 本表是最低公開證據稽核的示例，不是策略效果排名或資源配置建議。私立技專校院調整類型的關聯依據為Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->；其餘國外研究只用於機制與監測變項，不作臺灣效果證據。
+
+沒有策略依暫定規則達到正式建議門檻。「證據生成試辦」不表示可能有效；只有在地機制、成本上限、預定結果、比較、學生風險審查與停止條件齊備時，才可小規模、可逆地補充證據，否則仍延後。分類只描述證據取得姿態，不決定資源配置。
+
+# 討論
+
+## 公開資料能回答什麼
+
+研究結果顯示，公開資料最能支持三項最低稽核工作。第一，教育部資料可描述兩類私校的招生規模與合成註冊率，並辨識公式與版本限制。第二，率與人數並列可揭露分母驅動的表面改善。第三，證據分級可判斷一項主張是否只停在可見性、是否已有關聯，或是否具備效果判斷條件。
+
+公開資料無法普遍回答策略因果效果。部門總量沒有策略暴露資訊，學校頁面沒有反事實，政策文件沒有機構層級成效設計，國外學生選擇研究也不能直接移植。這個結論是「在本次有限、可取得且完成原文檢查的語料中無法判定」，不是「策略沒有作用」，也不是「管理者不可能判斷」。學校若具有可信的分期執行、觸及、成本、比較與學生結果等內部資料，可能形成比本文公開證據篩檢更強的判斷。
+
+## 私立一般大學與私立技專校院不能合併解讀
+
+兩類私校的基準註冊率、學制、系所結構與機構退出速度不同。表1顯示，私立技專校院在實際新生數下降時仍可出現註冊率回升；私立一般大學則在113學年度回升後仍低於109學年度。將兩者合併會掩蓋分母收縮幅度與制度差異。更重要的是，直接策略關聯研究只涵蓋TVET。即使一般大學也採行更名、合併或特色發展，也不能用TVET係數替代其效果。
+
+## 更名關聯的管理意義
+
+系所更名可能改變課程辨識，也可能伴隨課程重整或品牌重定位；Lin與Yang（2026）<!--ref:A1--><!--anchor:section:6.3.Methodological_and_Contextual_Constraints_on_Causal_Interpretation-->無法分離這些機制。若設計證據生成試辦，須確認課名、課程、師資與職涯承諾一致，並預定成本上限、基線、比較、招生漏斗、留存、轉退學及停止條件。期待落差或不適配風險升高時應停止擴大策略，但不表示自動停系或關校。
+
+合併則需先澄清目標。若目標是節省行政成本、整合師資或維持課程，招生沒有改善不必然表示合併失敗；但若管理者把招生增長列為主要理由，現有TVET證據不支持直接投入。這也解釋了政策文獻將合併列為治理選項，與A1未發現招生增益並非真正矛盾。
+
+## 國際招生與學生支持
+
+境外生政策可能增加可招收人口，但境外生同時進入註冊率分子與分母，數值解讀必須分開呈現本國與境外新生。政策目標也不能代替學校能力。教學互動、同儕支持、課程資訊與語言環境應成為國際招生的最低監測面向（Chalapati et al., 2018）<!--ref:A3--><!--anchor:section:Conclusion-->。若支持容量不足，增加招生可能把人口壓力轉化為學生適應、留存或教育品質問題。
+
+## 最低公開證據稽核與證據取得篩檢
+
+本研究提出五道程序閘門：
+
+1. **來源可用性：**原始資料或完整原文可取得，身分與版本可核對；摘要或搜尋片段不能進入實質主張。定位欄位存在不等於精確段落已驗證。
+2. **主張層級：**先判斷證據只能支持存在、共同變化或效果，不能因來源權威而自動升級。
+3. **測量完整性：**同時檢查註冊率、實際新生、分母、年度、校型、退出與組成。
+4. **效果辨識：**要求策略時間、強度、基線或前趨勢、比較、同期措施、延遲效果與替代解釋。
+5. **風險閘門：**檢查品質、適配、留存、公平、費用、法律、隱私、師資設備、可逆性及財務下行風險。
+
+輸出是證據已達門檻、可設計證據生成試辦、延後，或停止採用／擴大特定策略，不是決策命令。停止不等於自動關校、停系或退場；重大品質、公平、法律、隱私、學生福利或不可逆財務警訊也不能由註冊改善抵銷。此最低外部證據層須結合校內資料與治理程序，且未經評分者研究、閾值校準或預測檢驗，不能稱為具有信度、效度、準確度、管理有效性或一般化能力。
+
+## 研究限制
+
+第一，本研究是有限結構化檢索，不是系統性回顧。驗證碼、限流、403、Cloudflare與付費資料庫限制，可能低估中文研究、學位論文與負面案例；完整原文門檻提高可核對性，也縮小證據面。
+
+第二，官方分析使用2026年8月3日取得的114學年度累積CSV快照。後續修訂可能改變數值，且113學年度欄位範圍增加，跨年解讀須保留版本註記。
+
+第三，表1是重複橫斷面。校數下降可能來自退場、停招、合併或分類變動，不能當作同一批學校的平均效果。本文未完成區域與穩定規模分層，因此第一子問題只獲部分回答。
+
+第四，策略材料存在可見性、宣傳、存活與時間錯置偏誤。6校頁面是2026年檢索時的網頁，不能代表109至113學年度實施狀態。退出學校與失敗策略可能更不容易被搜尋到。
+
+第五，在本次有限原文語料中，Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->是唯一直接的臺灣策略關聯研究。移除該研究後，本文仍能支持趨勢描述、雙指標規則與「多數效果不可判定」，但所有特定策略關聯與試辦狀態都須撤回。
+
+第六，成本、機會成本、教育品質、留存、公平、負債與學生福利沒有和策略暴露連結。故本文不能進行成本效益分析或資源配置排名。
+
+第七，標記存在、來源取得與精確段落驗證不同。部分章節標籤是內部定位名稱，PDF頁碼預檢未完成，專用時間線稽核亦缺輸入檔。因此只能主張有限的來源與脈絡檢查，不能宣稱完整頁級或獨立驗證。
+
+## 未來研究
+
+未來可登錄策略日期、對象、觸及、投入、並行措施與停止規則，再連結招生漏斗、C+E、留存、轉退學及負擔。分期實施且假設成立時，才考慮中斷時間序列、差異中的差異或配對比較。
+
+私立一般大學須另行研究；區域與規模分析需要穩定校碼、跨年規則及退場處理。未來的訪談、學生研究或使用者檢驗均屬新的人類參與者設計，須另作倫理判定，不能描述為本研究已完成。
+
+# 結論
+
+公開資料能支持趨勢描述、指標稽核與最低外部證據篩檢，卻不足以普遍判斷招生策略效果。109至113學年度兩類私校實際新生數均下降，私立技專校院合成註冊率則小幅上升，說明單看比率可能誤判。本次有限原文語料只支持一項狹窄關聯：2020至2024年臺灣技專校院的系所調整類型與招生表現不同，其中私立機構更名的正向關聯較明顯；這不是因果結果，也不能推及一般大學。
+
+現有公開證據不足以形成策略排名或資源配置建議。只有在地機制、成本上限、預定結果、比較、學生風險與停止條件齊備時，部分TVET調整才可進入小規模、可逆的證據生成試辦；這不表示預期有效。其他策略在補足臺灣實施、成本與學生結果前應延後。重大風險應使學校停止採用或擴大策略，但不推導自動停系或關校。
+
+本文貢獻是最低公開證據稽核與證據取得篩檢，不是完整決策框架或已驗證模型。它要求先檢查版本、分母、實際人數、策略時間與強度、比較、替代解釋及學生風險。「公開證據尚不能判定」不妨礙學校以可信內部資料形成較強判斷。
+
+# AI協作揭露
+
+**AI Disclosure:** This report was produced with AI-assisted research tools. The research pipeline included AI-powered literature search, source-identity and source-to-claim checking, evidence synthesis, code-assisted descriptive analysis, and report drafting. Substantive claims were checked against the retrieved official data or full-text sources cited in this bounded review, subject to documented retrieval, snapshot, PDF-preflight, and section-anchor limitations. This checking did not independently replicate each underlying study or establish that every source finding is correct. Human oversight was applied throughout the process.
+
+本報告使用AI協作工具支援研究問題整理、公開資料與文獻檢索、來源身分及主張對照、證據綜合、程式化描述統計與初稿撰寫。研究者逐階段審閱並確認研究範圍、來源限制、跨文獻張力與推論邊界。實質主張依本次取得的官方資料或完整原文進行有限對照，但受到檢索缺漏、資料快照、PDF頁碼預檢不可用及部分內部章節定位標籤限制；此程序未獨立重做各篇研究，也不證明每項來源發現皆正確。AI未執行訪談、實驗或原始資料蒐集。
+
+# 經費與利益衝突聲明
+
+本課程論文未接受外部經費。作者與本文所列學校及招生策略沒有任職、治理、顧問或商業利益衝突。
+
+# 參考文獻
+
+Baliyan, S. P., & Mokoena, S. (2024). Students’ choice management: Recipe for improving enrollment in private higher education institutions. *Asian Journal of University Education, 20*(1), 206–220. https://doi.org/10.24191/ajue.v20i1.26030 <!--ref:A5--><!--anchor:section:6.Conclusion_and_Recommendations-->
+
+Chalapati, S., Leung, R., & Chalapati, N. (2018). Exploring factors affecting first-year students’ learning experiences: A case study of a private university in Taiwan. *Student Success, 9*(4), 25–39. https://doi.org/10.5204/ssj.v9i4.651 <!--ref:A3--><!--anchor:section:Conclusion-->
+
+Chang, D.-F., Wu, C.-T., Ching, G. S., & Tang, C.-W. (2009). An evaluation of the dynamics of the plan to develop first-class universities and top-level research centers in Taiwan. *Asia Pacific Education Review, 10*(1), 47–57. https://doi.org/10.1007/s12564-009-9010-7 <!--ref:A9--><!--anchor:section:Abstract-->
+
+Jongbloed, B., Enders, J., & Salerno, C. (2008). Higher education and its communities: Interconnections, interdependencies and a research agenda. *Higher Education, 56*(3), 303–324. https://doi.org/10.1007/s10734-008-9128-2 <!--ref:A8--><!--anchor:section:Introduction-->
+
+Kuo, Y.-Y. (2016). Taiwan universities: Where to go? *Humanities, 5*(1), Article 12. https://doi.org/10.3390/h5010012 <!--ref:A2--><!--anchor:section:4.Conclusions-->
+
+Levy, D. C. (2013). The decline of private higher education. *Higher Education Policy, 26*(1), 25–42. https://doi.org/10.1057/hep.2012.26 <!--ref:A6--><!--anchor:section:Demographic_decline-->
+
+Lin, W.-B., & Yang, C.-M. (2026). Demographic change and higher education governance: Evidence from departmental restructuring and enrollment dynamics in Taiwan’s TVET institutions. *Education Sciences, 16*(3), Article 371. https://doi.org/10.3390/educsci16030371 <!--ref:A1--><!--anchor:section:7.1.Conclusions-->
+
+Marginson, S. (2016). The worldwide trend to high participation higher education: Dynamics of social stratification in inclusive systems. *Higher Education, 72*(4), 413–434. https://doi.org/10.1007/s10734-016-0016-x <!--ref:A7--><!--anchor:section:Abstract-->
+
+Padlee, S. F., Kamaruddin, A. R., & Baharun, R. (2010). International students’ choice behavior for higher education at Malaysian private universities. *International Journal of Marketing Studies, 2*(2), 202–211. https://doi.org/10.5539/ijms.v2n2p202 <!--ref:A4--><!--anchor:section:Abstract-->
+
+中華醫事科技大學醫務暨健康事業管理系。（n.d.）。*招生特色*。https://hca.hwai.edu.tw/ <!--ref:U06--><!--anchor:section:招生特色-->
+
+中國科技大學視覺傳達設計系。（n.d.）。*中國科技大學視覺傳達設計系暨碩士班*。https://web.cute.edu.tw/dvcd/ <!--ref:U05--><!--anchor:section:Latest_News-->
+
+中國醫藥大學科技管理碩士學位學程。（n.d.）。*科技管理碩士學位學程*。https://mptm-cmu1819.cmu.edu.tw/ <!--ref:U02--><!--anchor:section:最新消息-->
+
+世新大學廣播電視電影學系。（n.d.）。*大學部*。https://rtf.wp.shu.edu.tw/?page_id=547 <!--ref:U03--><!--anchor:section:大學部-->
+
+佛光大學產品與媒體設計學系。（n.d.）。*產品與媒體設計學系、研究所*。https://pmd.fgu.edu.tw/ <!--ref:U01--><!--anchor:section:招生資訊-->
+
+行政院。（2023年9月20日）。*Boosting enrollment and retention of international students*。https://english.ey.gov.tw/News3/9E5540D592A5FECD/5ba2391b-d763-4441-9a85-a77cd10ad3e7 <!--ref:O6--><!--anchor:section:Policy_content-->
+
+南開科技大學行銷與流通管理系。（n.d.）。*行銷與流通管理系*。https://nkmlm.nkut.edu.tw/ <!--ref:U04--><!--anchor:section:招生資訊-->
+
+教育部。（n.d.-a）。*學12-2：新生（含境外生）註冊率－以校（含學制班別）統計*。大專校院校務資訊公開平臺。https://udb.moe.edu.tw/udata/StatSystemRegistrationRate <!--ref:O8--><!--anchor:section:統計說明-->
+
+教育部。（2026a）。*93–114學年度大專校院新生註冊率－按學制別分*［資料集］。https://ws.moe.edu.tw/001/Upload/15/relfile/7851/103816/12f43c6a-9e69-4212-9e31-39883e27ce4f.xlsx <!--ref:O1--><!--anchor:section:說明-->
+
+教育部。（2026b）。*大學、技專校院各系所註冊率*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/26228 <!--ref:O3--><!--anchor:section:AY109-113_data_rows-->
+
+教育部。（2026c）。*大專院校各校科系別學生數*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/9622 <!--ref:O4--><!--anchor:section:資料說明-->
+
+盧延根。（2022年2月）。*我國少子化問題與對策之研析*（R01617）。立法院。https://www.ly.gov.tw/Pages/Detail.aspx?nodeid=6590&pid=165915 <!--ref:O7--><!--anchor:section:我國少子化問題之概況-->
+
+# 附錄A：可重製資料規則摘要
+
+- 分析期間：109至113學年度。
+- 設立別：私立。
+- 校型：一般大學、技專校院，分開彙整。
+- 納入學制：學士班（含四技）、二年制大學（二技）、二專、五專、學士後第二專長學士班（4+X）。
+- 主要結果：先按校型與年度合計官方欄位後重算的合成註冊率，不是各校公開率的平均。
+- 配對結果：實際新生人數C+E。
+- 分母：A−B+E。
+- 缺值規則：非數字即停止，不補值。
+- 組成規則：保留各年度機構數；不宣稱平衡追蹤。
+- 程式：`stage-01-research/phase2-search/derive_private_enrollment_trends.py`。
+- 輸出：`stage-01-research/phase2-search/verification/private_enrollment_trends_109_113.csv`。
+
+# 附錄B：未解決事項
+
+1. 區域與穩定規模的跨年度趨勢未完成。
+2. 私立一般大學缺直接策略關聯研究。
+3. 策略時間、強度、成本、同期措施與反事實普遍缺漏。
+4. 品質、留存、適配、公平、負債與學生福利未與策略連結。
+5. 退出學校、失敗策略、消失網頁及中文付費文獻可能不足。
+6. 最低稽核與證據取得篩檢尚未接受獨立編碼、使用者研究、閾值校準或預測驗證。
+7. 部分引文仍缺頁級精確段落定位；專用時間線稽核與期刊撤稿／更正資料庫檢查尚未完成。
+
+# 附錄C：引文可追溯性狀態
+
+| 狀態 | 本稿意義 | Phase 6結果 |
+|---|---|---|
+| 標記存在 | 引文後有機器可讀來源與定位欄位 | 54/54組完成；只證明通道完整 |
+| 來源已取得 | 研究期間取得官方檔案、完整原文或網頁 | 21/21個參考來源已取得並完成脈絡檢查 |
+| 精確段落已驗證 | 可確認頁碼、出版社章節或段落與主張直接對應 | 部分完成；內部合成章節標籤與PDF頁碼預檢限制使其不能宣稱全面完成 |
+
+`AY109-113_data_rows`等標籤是本研究的內部資料定位名稱，不是來源發布者的正式章節。HTML標記完整性不得被描述成頁級驗證或獨立重作研究結果。
+
+# 字數說明
+
+學者於2026年8月3日澄清：本繁體中文課程論文的目標是約6,000個中文字元，而非6,000個英語式單字。本修訂稿以腳本估算，緒論至結論含表格文字約6,601個CJK字元；排除Markdown表格列後約6,164個。此處只透明報告計數口徑，不主張中文字元與英語單字等值；正式計數仍依課程排版規則。英文摘要為227個空白分隔詞。參考文獻、附錄、揭露及聲明不計入正文目標。

examples/pi/stage-02.5/focused-reverification-report.md

@@ -0,0 +1,137 @@
+# Stage 2.5 Focused Re-verification Report
+
+## Scope
+
+**Correction authorization:** `fix all`  
+**Round:** 2 of maximum 3  
+**Re-verified artifact:** `02-corrected-pre-review-course-paper.md`  
+**Corrected SHA-256:** `867ab46797b7270fdbe3edb8814fa24cfe1712a2aa8ceb09f27b7cbcee320d76`  
+**Original Stage 2 SHA-256:** `953badfa35dab3ae6bcd183918b5d319b8575e22df944a3f0fa2ac54429f6507` (unchanged)
+
+# Verdict: PASS
+
+Both Round-1 issues are resolved. No SERIOUS or MEDIUM issue remains. The focused changes introduced no new citation, data, temporal, experiment-provenance, or failure-mode block. Non-blocking advisories remain explicit.
+
+## Correction closure
+
+| Issue | Round-1 severity | Round-2 result | Verification |
+|---|---|---|---|
+| IL-MEDIUM-1 — O7 wrong locator | MEDIUM | RESOLVED | Corrected URL returned HTTP 200 and contained exact title, author, and R01617 code. |
+| IL-MINOR-1 — stale check status | MINOR | RESOLVED | Both stale phrases are absent; replacement statements match preserved Stage 2.5 logs and retain boundedness. |
+
+## Preservation and scope control
+
+- The scholar-approved Stage 2 artifact remains unchanged at its original path and hash.
+- The corrected paper is a distinct pre-review artifact.
+- `diff -u` contains exactly three one-line replacement hunks:
+  1. Methods process-status statement.
+  2. Limitations process-status statement.
+  3. O7 URL.
+- No abstract result, table value, body citation marker, source title/author/year, research question, inference boundary, or conclusion was changed.
+
+Patch: `round2-focused-diff.patch`  
+Patch SHA-256: `c4b1999aafac010c8fd8278e7b1f02fd08ee4f1bff1a84e26e7d5f3c67dfcf50`
+
+## O7 focused verification
+
+- Corrected locator: `https://www.ly.gov.tw/Pages/Detail.aspx?nodeid=6590&pid=217372`
+- Fresh response: HTTP 200; 19,134 bytes.
+- Exact content checks:
+  - title `我國少子化問題與對策之研析`: present;
+  - author `盧延根`: present;
+  - report code `R01617`: present.
+- Corrected paper counts: old locator 0; corrected locator 1.
+- Body author-year citation and section anchor remain unchanged.
+
+Result: IL-MEDIUM-1 closed.
+
+## Process-status focused verification
+
+The corrected Methods/Limitations accurately report:
+
+1. Nine DOI-bearing articles were checked against the saved Crossref Labs Retraction Watch snapshot, OpenAlex retraction field, Crossref relations, and title/DOI web searches.
+2. No exact indexed retraction/correction match was found as of 2026-08-03.
+3. The result is explicitly limited by date and database coverage and is not framed as permanent or universal clearance.
+4. The dedicated temporal tool checked the manuscript body.
+5. `pypdf` remains unavailable.
+6. `pdfinfo`/`pdftotext` page mapping does not establish universal sentence-level exact-passage verification.
+
+The Round-2 temporal audit, using corrected O7 provenance, returned **0 findings**. Round-1 full-file reference-list false positives remain preserved rather than deleted.
+
+Result: IL-MINOR-1 closed.
+
+## Structural and citation checks
+
+| Check | Result |
+|---|---|
+| Three-layer citation lint | PASS |
+| Body citation markers | 58 |
+| Unique body sources | 21 |
+| Reference entries | 21 |
+| Paired anchors | 58 |
+| Orphan body/reference sources | 0 / 0 |
+| Metadata-only substantive sources | 0 |
+
+A fresh bounded count from the text after `## 一、緒論` through the text before `## 資料可得性聲明`, including intervening headings and tables, yields 5,935 CJK characters. The target remains approximate.
+
+## Quantitative check
+
+The official enrollment derivation was run again during focused re-verification:
+
+- Exit code: 0.
+- stderr: empty.
+- Derived CSV SHA-256: `ae1802065413b8cd0ff11686382eb6202e671cc18836243156536ebe75251fff`.
+- No table or numerical-result line changed in the correction patch.
+
+Result: PASS.
+
+## AI Research Failure Mode Checklist
+
+The mandatory seven-mode checklist was completed:
+
+| Mode | Result |
+|---|---|
+| Implementation bug passing self-review | CLEAR |
+| Hallucinated citation | CLEAR |
+| Hallucinated experimental result | CLEAR |
+| Shortcut reliance | CLEAR |
+| Bug reframed as insight | CLEAR |
+| Methodology fabrication | CLEAR |
+| Early frame-lock | CLEAR |
+
+Totals: 7 CLEAR; 0 SUSPECTED; 0 INSUFFICIENT EVIDENCE. No failure-mode override is required.
+
+Artifact: `round2-focused-reverification/ai-research-failure-mode-checklist.md`.
+
+## Compliance extension
+
+The mode-aware compliance check classified the paper as **other evidence synthesis**, not a systematic review. Therefore:
+
+- PRISMA-trAIce was used only as a non-blocking adaptation/information check.
+- RAISE returned `warn` for incomplete AI reporting/reproducibility detail: model/provider/version, prompts/parameters, no-fine-tuning declaration, output formats, formal human-review qualifications/adjudication detail, and full stochastic reproducibility are not in the paper.
+- These gaps do not invalidate the paper’s evidence claims and do not block Stage 2.5 under non-systematic mode, but they require explicit checkpoint acknowledgement.
+- PRISMA-trAIce is itself a foundational pre-Delphi proposal and has not been empirically validated across diverse contexts.
+
+Schema 12 validation: PASS.  
+Artifact: `round2-focused-reverification/compliance-report.json`.
+
+## Retained advisories
+
+The following remain open by scholar instruction and are not silent passes:
+
+| Advisory | Retained status |
+|---|---|
+| `pypdf` availability | Unavailable |
+| Universal sentence-level exact-passage verification | Incomplete |
+| Author-publication self-plagiarism search | Unavailable while author identity is a placeholder |
+| Table 1/2 `figure_table_trace[]` sidecars | Absent; both tables retain PASS WITH NOTES |
+| O1 fresh endpoint | Timed out; prior official snapshot/hash and MOE index remain |
+| Search coverage / E5 novelty wording | Corpus/date bounded; platform names not all stated in the same novelty sentence |
+| Search zero hits/API/TLS/access limits | Preserved in raw audit trail |
+| Cross-model verification | Not enabled |
+
+## Gate determination
+
+The legacy integrity gate and mandatory seven-mode failure checklist now PASS. Compliance contributes a non-blocking WARN. Stage 2.5 is therefore **technically PASS with explicit advisories**, but the mandatory integrity-boundary checkpoint still requires scholar confirmation before Stage 3 may begin.
+
+No Stage 3 transition has been performed.

examples/pi/stage-02/course-paper-draft.md

@@ -0,0 +1,242 @@
+# 少子化情境下臺灣私立大學招生策略證據之可判定性：公開資料趨勢與最低公開證據稽核
+
+**學生：**［學生姓名］  
+**系所：**［系所名稱］  
+**課程：**［課程名稱］  
+**授課教師：**［教師姓名］  
+**日期：**2026年8月3日
+
+## 中文摘要
+
+少子化使臺灣私立校院持續承受招生與治理壓力，但註冊率上升不必然表示新生增加，公開可見的招生措施也不等於效果證據。本研究探討公開可驗證資料在何種程度上能支持私立一般大學與私立技專校院的招生策略判斷。研究採整合式次級證據設計，結合有限結構化文獻、政策與文件分析，以及教育部109至113學年度資料的描述性重算。結果顯示，私立一般大學實際新生由68,505人降至58,768人，合成分層註冊率由92.95%降至90.02%；私立技專校院實際新生由96,155人降至79,451人，該率卻由76.27%升至77.29%，顯示分母收縮可與人數下降並存。於截至2026年8月3日取得並檢查全文的有限語料中，直接策略研究僅支持臺灣技專校院2020至2024年的關聯判斷，不能推論因果或移用至私立一般大學；六校目的性頁面只證明校方主張可見。本研究據此提出最低公開證據稽核與證據取得篩檢。其用途是區分可見性、關聯與效果判斷，並標示下一步資料需求，而非策略排名或已驗證的管理模型。
+
+**關鍵詞：**人口變遷、高等教育治理、技職教育、招生指標、次級資料、證據分級、學生風險
+
+## Abstract
+
+Demographic contraction places sustained enrollment and governance pressure on Taiwan’s private higher education sector, but a higher freshman registration rate does not necessarily indicate growth in actual enrollment. Publicly visible recruitment initiatives likewise do not establish effectiveness. This study examines the extent to which publicly verifiable evidence can support strategy judgments in private general universities and private technological and vocational institutions. An integrated secondary-evidence design combined a bounded structured review of academic, policy, and institutional documents with descriptive recomputation of Ministry of Education data for academic years 109–113. Actual freshmen in private general universities declined from 68,505 to 58,768, while the synthetic stratum-level registration ratio moved from 92.95% to 90.02%. In private technological and vocational institutions, actual freshmen declined from 96,155 to 79,451 even as the ratio rose from 76.27% to 77.29%, demonstrating that denominator contraction can coexist with lower enrollment. Within the bounded full-text corpus retrieved and examined as of 3 August 2026, direct strategy evidence supported only association-level judgments for Taiwan TVET during 2020–2024; it did not establish causality or warrant transfer to private general universities. Six purposively selected institutional pages documented issuer-authored visibility only. The resulting minimum public-evidence audit distinguishes visibility, association, and effect judgment while identifying further data needs. It is an evidence-acquisition screen, not a strategy ranking or validated management model.
+
+**Keywords:** demographic contraction, higher education governance, technical and vocational education, enrollment measurement, secondary evidence, evidence classification, student risk
+
+## 一、緒論
+
+### （一）少子化、私校壓力與指標陷阱
+
+臺灣出生人口長期下降，使高等教育生源不足成為持續性的政策問題（盧延根，2022）<!--ref:O7--><!--anchor:section:我國少子化問題之概況-->。私立校院較依賴學費與招生收入，因而較容易承受人口與市場變化，但「私立高等教育衰退」可能指市場占比、絕對人數或特定次部門下降，不能推論每所私校必然失敗（Levy, 2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->。臺灣政策研究亦曾提出特色發展、產學合作、國際化與組織調整等因應方向（Kuo, 2016）<!--ref:A2--><!--anchor:section:4.Conclusions-->。這些方向是治理選項，不是已證實的招生效果。
+
+招生壓力也涉及測量問題。教育部新生註冊率的統計範圍與公式曾於103、107與109學年度調整，跨期比較須先確認定義（教育部，2026a）<!--ref:O1--><!--anchor:section:說明-->。註冊率由核定名額、保留入學資格、實際註冊及境外新生共同構成，因此比率上升可能源於分母收縮，而非新生增加（教育部，n.d.-a）<!--ref:O8--><!--anchor:section:統計說明-->。若只讀取單一比率，管理者可能將名額調整誤認為招生策略成效。
+
+### （二）研究缺口、問題與貢獻
+
+公開資料呈現的「學校宣稱採取某策略」、「策略與結果同時變動」及「策略造成結果」是不同層級的主張。政策頁面或招生網頁可以證明內容曾被公開，但通常沒有策略起始時間、投入強度、比較條件與學生後續結果。依本研究截至2026年8月3日完成的繁體中文與英文有限結構化檢索，在取得並檢查完整原文的語料中，Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->是唯一直接比較臺灣技專校院系所調整類型與招生結果的研究。此說法只適用於本次檢索，不能推論其他公開、付費或內部研究不存在。
+
+本研究據此提出問題：「在臺灣少子化情境下，公開可驗證資料在何種程度上能支持私立一般大學與私立技專校院對招生策略成效的管理判斷？」子問題分別處理109至113學年度兩類私校的新生趨勢、不同來源能支持的最高主張層級，以及公開證據如何轉化為最低稽核與證據取得狀態。本文主張，公開資料足以支持分層趨勢描述、分母敏感的指標稽核與最低證據分類，卻不足以普遍判定策略效果。其用途應是校內決策前的最低外部證據層，而非策略排名或已驗證的管理模型。
+
+## 二、文獻回顧與分析視角
+
+### （一）人口收縮、私立部門與制度分層
+
+人口收縮不會以相同方式作用於所有大學。私立部門的結果仍受聲譽、區位、需求、公立競爭與政策影響，故人口變化只能構成風險背景，不能作為個別學校表現的充分解釋（Levy, 2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->。高參與高等教育體系也可能依公私立別、機構階層、學科與學費維持社會分層（Marginson, 2016）<!--ref:A7--><!--anchor:section:Abstract-->。因此，私立一般大學與私立技專校院的學制、系所結構及市場位置不同，合併分析會掩蓋制度差異。分層的目的不是預設哪一類校院較脆弱，而是避免用總體平均抵銷彼此不同的名額基準、退出速度與策略證據範圍。只有先固定分析單位，後續趨勢與策略文獻才具有一致的解讀基礎。
+
+臺灣治理文獻列出的合併、特色發展及國際化，主要說明制度可採取的調整方向（Kuo, 2016）<!--ref:A2--><!--anchor:section:4.Conclusions-->。另一項臺灣高等教育政策評估發現，增加投入並不保證更佳表現，但研究對象是頂尖大學計畫，不能轉用為私校招生效果估計（Chang et al., 2009）<!--ref:A9--><!--anchor:section:Abstract-->。兩者共同支持一項方法警示：政策存在與資源投入都不能代替結果辨識。
+
+### （二）學生選擇、國際化與入學後支持
+
+學生選擇研究可協助界定未來應觀察的變項。馬來西亞私立大學研究辨識學習環境、服務、成本、設施、社交與區位等選擇因素（Padlee et al., 2010）<!--ref:A4--><!--anchor:section:Abstract-->。波札那私立高等教育研究則納入就業、品質、宣傳、財務支持及同儕影響（Baliyan & Mokoena, 2024）<!--ref:A5--><!--anchor:section:6.Conclusion_and_Recommendations-->。這些研究來自不同制度與學生群體，只能提供候選機制，不能證明臺灣私校採取廣告、獎助或就業定位後會提高註冊。
+
+境外招生同時牽涉入學後支持。臺灣一所私立大學國際學程的焦點團體顯示，互動教學、同儕支持、課程資訊、語言環境與社會融入影響新生經驗（Chalapati et al., 2018）<!--ref:A3--><!--anchor:section:Conclusion-->。行政院頁面則確認2024至2028年擴大境外生招生與留臺政策的政策目的（行政院，2023）<!--ref:O6--><!--anchor:section:Policy_content-->，但沒有機構層級效果評估。招生人數與學生適應因而必須分開觀察，政策目標也不能替代學校支持能力。
+
+### （三）系所調整、治理與投入績效
+
+Lin與Yang（2026）<!--ref:A1--><!--anchor:section:4.Methodology-->分析2020至2024年間臺灣74所技專校院的1,087件系所調整案例。研究顯示，更名持續呈現較有利的招生關聯，且私立機構的關聯較明顯。合併通常未呈現正向招生變化，停招或關閉多發生於長期低註冊系所，部分需求導向的新設或擴額則呈現較佳關聯（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->。然而，調整可能回應既有趨勢，固定效果也無法排除所有隨時間改變的混淆。作者因此將發現限定為關聯，而非因果（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:6.3.Methodological_and_Contextual_Constraints_on_Causal_Interpretation-->。
+
+合併未帶來招生增益，不等於合併在治理上必然失敗。大學與地方、產業及其他利害關係人的合作涉及任務選擇、問責與資源負荷，成效不能只用招生量衡量（Jongbloed et al., 2008）<!--ref:A8--><!--anchor:section:Introduction-->。若合併目的在整合師資或維持課程，應另設治理指標。若以招生增長為主要理由，現有TVET關聯證據則不足以支持直接擴大。
+
+### （四）三層主張與最低稽核
+
+本文將證據分為策略可見性、結果共同變化或關聯，以及效果判斷三層。第一層只回答措施是否被宣布或描述。第二層須有明確單位與期間的策略分類及結果。第三層另要求時間順序、實施強度、可信比較、基線或前趨勢、同期措施與替代解釋。判斷後還須檢查教育品質、適配、公平、法律、隱私、學生福利與財務風險（Chalapati et al., 2018）<!--ref:A3--><!--anchor:section:Conclusion-->（Jongbloed et al., 2008）<!--ref:A8--><!--anchor:section:Introduction-->（Marginson, 2016）<!--ref:A7--><!--anchor:section:Abstract-->。此分類只界定公開證據允許的主張，不預測策略成功。三層之間也不是來源數量的累加關係：多個校方頁面仍只表示可見，多項國外機制研究仍不能替代臺灣比較資料。證據升級取決於研究設計是否回答時間、暴露與替代解釋，而非引用篇數或發布者權威。
+
+## 三、研究方法
+
+### （一）研究設計與檢索
+
+本研究採整合式次級證據設計，結合有限結構化文獻、政策與文件分析，以及官方公開資料的描述性重算。研究沒有訪談、問卷、實驗、人類參與者或原始資料蒐集，也不是完整系統性回顧。直接證據限於臺灣。國際研究只用於機制、候選變項與風險條件。
+
+檢索以繁體中文及英文進行，共執行43組網路查詢與16次Crossref/OpenAlex API查詢。382筆原始紀錄經DOI及標題連結後形成316個候選群組，最後有9篇學術全文進入實質分析。另9筆僅完成書目或摘要驗證的紀錄未用於主張或參考文獻。每次查詢均保留查詢字串、平台、UTC時間、URL及可見的回傳數，過程也保留零結果、驗證碼、限流、403、404與Cloudflare失敗。檢索片段只作發現線索，無法取代原始頁面或全文。候選DOI `10.1177/1745499920916532`及`10.1080/13538322.2020.1736512`分別校正為`10.1177/1745499920910581`及`10.1080/08841241.2020.1728465`。因未取得校正後全文，兩筆仍未進入實質分析。
+
+### （二）官方資料、分層與合成比率
+
+量化資料取自2026年8月3日下載的教育部「大學、技專校院各系所註冊率」114學年度累積CSV快照（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。分析保留109至113學年度、設立別為私立、校型為一般大學或技專校院，以及大學部相關學制，兩類校院分開彙整。欄位包括核定名額A、保留入學資格B、總量內實際註冊C及境外新生E。合成分層註冊率為`(C+E)/(A−B+E)`（教育部，n.d.-a）<!--ref:O8--><!--anchor:section:統計說明-->。
+
+本研究先按年度與校型加總A、B、C、E，再計算合成比率。這不是各校公開註冊率的算術平均。109學年度起與113學年度新增的欄位範圍仍保留版本註記（教育部，n.d.-a）<!--ref:O8--><!--anchor:section:統計說明-->。各年度納入校數不同，因此結果屬重複橫斷面，而非同一批學校的平衡追蹤。
+
+### （三）六校材料與分析程序
+
+學校材料以113學年度學生總數作為規模脈絡（教育部，2026c）<!--ref:O4--><!--anchor:section:資料說明-->，在私立一般大學與私立技專校院中各選取小、中、大規模一校，共六校。佛光大學（n.d.）<!--ref:U01--><!--anchor:section:招生資訊-->、中國醫藥大學（n.d.）<!--ref:U02--><!--anchor:section:最新消息-->及世新大學（n.d.）<!--ref:U03--><!--anchor:section:大學部-->頁面呈現跨域、創新或影音專業等主張。南開科技大學（n.d.）<!--ref:U04--><!--anchor:section:招生資訊-->、中國科技大學（n.d.）<!--ref:U05--><!--anchor:section:Latest_News-->及中華醫事科技大學（n.d.）<!--ref:U06--><!--anchor:section:招生特色-->頁面則呈現實務、設計或醫療合作主張。樣本只測試公開材料是否可取得，不估計策略盛行率、實施或效果。
+
+分析先檢查年度、單位、分母與缺值。任何空白、省略號或非數字欄位均使程式停止，不自動補值。文件矩陣記錄來源身分、主張層級、替代解釋、風險與驗證狀態。引文可追溯性另分為標記存在、來源已取得及精確段落已驗證。三者不能互相替代。部分PDF未完成頁碼預檢，專用時間線與撤稿／更正檢查亦未完整執行，故本文只主張有限來源及脈絡檢查。
+
+## 四、研究結果
+
+### （一）兩類私校的新生趨勢
+
+表1呈現109至113學年度兩類私校的實際新生C+E及合成分層註冊率。數值由固定程式重算教育部CSV欄位（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。
+
+**表1**  
+*109至113學年度私立校院大學部新生趨勢*
+
+| 學年度 | 私立一般大學校數 | 實際新生C+E | 合成註冊率 | 私立技專校院校數 | 實際新生C+E | 合成註冊率 |
+|---:|---:|---:|---:|---:|---:|---:|
+| 109 | 36 | 68,505 | 92.95% | 67 | 96,155 | 76.27% |
+| 110 | 36 | 66,799 | 91.02% | 64 | 90,159 | 76.25% |
+| 111 | 35 | 56,311 | 77.64% | 62 | 81,922 | 71.78% |
+| 112 | 34 | 55,898 | 83.09% | 59 | 78,785 | 73.68% |
+| 113 | 34 | 58,768 | 90.02% | 57 | 79,451 | 77.29% |
+
+*註。* 校數為各年度納入範圍內出現的機構數。比率是加總A、B、C、E後重算的合成分層比率，不是各校公開率平均。資料來自教育部（2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。
+
+私立一般大學的實際新生由68,505人降至58,768人，減少9,737人（14.2%），合成比率由92.95%降至90.02%。私立技專校院的實際新生由96,155人降至79,451人，減少16,704人（17.4%），合成比率卻由76.27%升至77.29%。同期核定名額A由123,366降至92,294（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。比率與人數在TVET呈現相反方向，是本研究最直接的指標稽核結果。它顯示分母收縮可與實際人數下降並存，但不能說明何種策略造成變化。
+
+兩類校院在111學年度的比率較低，113學年度回升。名額、機構退出或分類、境外生、學科需求、疫情後環境與同期政策皆可能構成替代解釋。由於本文未建立跨年度區域與穩定規模交叉表，教育部學生數資料只支持113學年度抽樣脈絡（教育部，2026c）<!--ref:O4--><!--anchor:section:資料說明-->，第一子問題仍屬部分回答。
+
+### （二）策略證據的最高層級
+
+策略來源的證據層級不一致。六校頁面只顯示2026年8月3日檢索時可見的校方主張。行政院頁面只證明境外生政策及其目的存在（行政院，2023）<!--ref:O6--><!--anchor:section:Policy_content-->。Chalapati等人（2018）<!--ref:A3--><!--anchor:section:Conclusion-->、Padlee等人（2010）<!--ref:A4--><!--anchor:section:Abstract-->及Baliyan與Mokoena（2024）<!--ref:A5--><!--anchor:section:6.Conclusion_and_Recommendations-->提供學生選擇與支持機制。Levy（2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->、Marginson（2016）<!--ref:A7--><!--anchor:section:Abstract-->、Jongbloed等人（2008）<!--ref:A8--><!--anchor:section:Introduction-->及Chang等人（2009）<!--ref:A9--><!--anchor:section:Abstract-->則提供私立部門、分層、治理及投入績效警示。這些來源均不能建立臺灣私校策略的因果效果。
+
+本次完整原文語料中，只有Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->把明確策略類型與臺灣招生結果直接連結。其私立TVET更名及部分需求導向新設／擴額最多達到關聯層級。合併未呈現穩定正向招生關聯，停招或關閉是低註冊後的結果型態。私立一般大學缺乏相當的直接關聯研究，不能借用TVET結果填補。
+
+### （三）最低公開證據與取得狀態
+
+表2依三層主張及風險條件呈現例示性的證據取得狀態。它不評分策略，也不決定資源配置。
+
+**表2**  
+*策略公開證據與後續證據取得狀態*
+
+| 策略類別 | 公開證據最高層級 | 暫定狀態 | 主要缺口或條件 |
+|---|---|---|---|
+| TVET系所更名 | 關聯 | 延後；條件齊備才設計試辦 | 課名相符、成本、基線、比較、雙重結果、留存 |
+| TVET系所合併 | 關聯但多未呈正向 | 招生效果延後 | 治理目的、延遲效果、比較、學生權益 |
+| 需求導向新增或擴額 | 稀少情境關聯 | 延後；條件齊備才設計試辦 | 需求、師資設備、品質、分母、停止條件 |
+| 停招或關閉 | 低註冊後的型態 | 停止作為增長策略 | 另行處理教學與學生權益，非自動停系或關校 |
+| 國際招生 | 政策存在與間接機制 | 延後 | 轉換、支持、留存、成本、法規及分母 |
+| 接觸、廣告、獎助 | 國外機制 | 延後 | 校級漏斗、淨價格、公平、成本與比較 |
+| 產學、地方合作、實習 | 可見主張與間接機制 | 延後 | 實施強度、實習品質、負擔及就業結果 |
+| 私立一般大學各策略 | 無直接關聯研究 | 延後 | 時間、強度、比較、成本與學生結果 |
+
+*註。* TVET調整類型的關聯依據限於Lin與Yang（2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->。其餘來源只支援政策存在、機制或風險條件。試辦僅指證據生成，不表示預期有效。
+
+現有公開語料沒有任何策略達到正式建議門檻。若要進入小規模證據生成試辦，至少須有在地機制合理性、成本上限、預定結果、比較邏輯、學生風險審查、可逆性及停止條件。條件不足時維持延後。停止僅表示停止採用或擴大特定策略，不推導自動停系、關校或退場。
+
+## 五、討論
+
+### （一）雙重指標與校型分層的管理意義
+
+兩類私校實際新生均下降，但TVET合成比率小幅上升，說明「比率改善」與「生源增加」不能畫上等號。合成比率仍有管理價值，因為它呈現名額使用程度。實際新生則反映招生規模。兩者連同A−B+E分母及機構組成一起呈現，才能避免把名額收縮誤判為策略成效（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->（教育部，n.d.-a）<!--ref:O8--><!--anchor:section:統計說明-->。 
+
+校型分層同樣不能省略。私立一般大學與私立技專校院的基準比率、學制及退出速度不同，且唯一直接策略關聯研究只涵蓋TVET（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:4.Methodology-->。合併兩類資料會同時遮蔽分母變化及證據可轉用範圍。人口收縮是共同壓力，卻不是將不同制度位置視為同一分析單位的理由（Levy, 2013）<!--ref:A6--><!--anchor:section:Demographic_decline-->。
+
+### （二）公開證據與校內證據的分工
+
+公開資料的優勢在於定義可查、跨校涵蓋及可重算，弱點則是沒有策略暴露資訊。部門總量不能顯示某校何時採取何種措施，學校網頁沒有反事實，政策文件也沒有機構層級成效設計。因此，本研究提出的最低公開證據稽核，合理功能是指出主張停在哪一層，以及下一步缺少哪些資料。
+
+這項功能不代表篩檢已能改善決策。它尚未接受獨立編碼、使用者研究、閾值校準或預測檢驗，也沒有證明管理成效。Jongbloed等人（2008）<!--ref:A8--><!--anchor:section:Introduction-->所討論的利害關係人與問責問題，以及Chang等人（2009）<!--ref:A9--><!--anchor:section:Abstract-->提出的投入不等於績效警示，只能支持擴大判斷面向，不能驗證本研究程序。其貢獻限於讓來源、主張層級、缺口與風險可被追查。
+
+學校若具有可信的策略日期、觸及、成本、招生漏斗、同期措施、留存與學生結果，可以形成比公開語料更強的判斷。內部資料也不會因「內部」身分自動具有因果力，仍須檢查基線、比較與替代解釋。公開稽核因此是最低層，不是完整決策環境。
+
+### （三）證據生成試辦與停止語義
+
+人口壓力可能要求快速行動，但速度不必等同大規模採用。以A1的更名關聯為例，更名可能伴隨課程重整、品牌重定位或既有需求趨勢，研究無法分離各項機制（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:6.3.Methodological_and_Contextual_Constraints_on_Causal_Interpretation-->。若學校欲取得在地證據，應先確認名稱與課程、師資及職涯承諾一致，再設定成本上限、基線、比較、招生漏斗、留存與停止條件。此試辦的目的在減少不確定性，不是預告正面結果。
+
+合併案例也顯示目標須先拆分。若目標是整合師資或降低行政重複，招生未增加不必然代表治理失敗。若目標明定為招生增長，現有TVET研究沒有提供穩定正向支持（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->。停止擴大某項招生策略，也不等於自動停止系所或學校。後者需要另行處理教學延續、學生權益、財務及區域可近性。
+
+### （四）品質、公平與國際招生風險
+
+招生增加不能抵銷教育品質與學生風險。高參與體系可能在擴張中維持階層與費用不平等（Marginson, 2016）<!--ref:A7--><!--anchor:section:Abstract-->。國際學程經驗亦顯示，教學互動、語言及社會支持會影響新生適應（Chalapati et al., 2018）<!--ref:A3--><!--anchor:section:Conclusion-->。境外招生除須拆分本國生與境外生對分子、分母的影響，也要追蹤留存、支持負荷、法規與就業協助。否則，招生壓力可能轉化為不適配、退學或品質問題。
+
+風險閘門不要求零風險，而是禁止用註冊改善掩蓋重大警訊。課名與課程不符、誤導招生、過度負債、隱私違規、支持容量不足或不可逆財務曝險，都應阻止策略擴大。利害關係人問責與任務負荷也要求學校說明誰承擔成本、誰獲得利益，以及招生目標是否偏離教育使命（Jongbloed et al., 2008）<!--ref:A8--><!--anchor:section:Introduction-->。
+
+### （五）限制與未來研究
+
+本研究是有限結構化檢索，不是系統性回顧。驗證碼、限流、付費與網站阻擋可能使中文研究、學位論文、退出學校及負面案例不足。完整原文門檻提高可核對性，卻縮小可使用的證據面。專用撤稿／更正篩檢、完整PDF頁碼預檢及時間線稽核尚未完成，來源身分已核對不等於所有精確段落均完成頁級驗證。
+
+官方分析依2026年8月3日的資料快照，後續修訂可能改變數值。113學年度欄位範圍亦有新增，表1又是機構數改變的重複橫斷面。本文未完成區域與穩定規模趨勢，策略時間、強度、成本、反事實及品質、留存、公平和福利結果也未連結，不能進行成本效益或資源配置排名。
+
+未來研究可建立校級策略登錄，記錄日期、對象、觸及、投入、同期措施與停止規則，再連結申請至註冊漏斗、C+E、留存、轉退學及負擔。分期實施且假設成立時，才可考慮中斷時間序列、差異中的差異或配對比較。私立一般大學、區域與穩定規模須另行研究。未來訪談或學生研究屬新的人類參與者設計，必須另作倫理審查。
+
+## 六、結論
+
+### （一）研究問題的回答
+
+公開可驗證資料對招生策略管理判斷的支持程度並不一致。教育部資料能描述109至113學年度兩類私校的實際新生、合成分層比率及分母變化，並揭示TVET比率與人數方向不同（教育部，2026b）<!--ref:O3--><!--anchor:section:AY109-113_data_rows-->。這些資料可以回答「結果如何變動」及「指標如何形成」，不能回答「哪項策略造成變動」。政策與學校頁面主要支持策略或主張可見性。本次完整原文語料只提供一項狹窄的臺灣TVET系所調整關聯，不能形成因果判斷或推及私立一般大學（Lin & Yang, 2026）<!--ref:A1--><!--anchor:section:7.1.Conclusions-->。因此，研究問題的答案不是單一的可或不可，而是依主張層級而異：描述與測量具有較強支持，特定TVET調整僅有關聯支持，普遍策略效果、成本效益與跨校排名則缺乏足夠公開證據。
+
+### （二）貢獻與使用邊界
+
+本文的貢獻是最低公開證據稽核與證據取得篩檢。它要求管理者先核對來源、版本、分母、實際人數、策略時間與強度、比較、替代解釋及學生風險，再判斷應補充何種證據。這項程序把「沒有足夠公開證據」轉化為可追查的資料需求，例如需要補足招生漏斗、投入成本、同期措施與留存結果。此程序具有可追溯性，但未證明信度、效度、預測力、管理效果或一般化能力。可信校內資料可能支持較強結論，也可能因缺乏比較或資料品質不足而維持不確定。
+
+「公開證據尚不能判定」是合理研究結果，不等於策略無效，也不要求組織停止思考。條件齊備時，學校可以用小規模、可逆的證據生成試辦降低不確定性。重大品質、公平、法律、隱私、學生福利或財務風險則不能由註冊改善抵銷。公開稽核的作用，是讓行動與主張不超過證據。
+
+## 資料可得性聲明
+
+本研究使用的教育部公開資料、下載日期、來源URL、衍生程式與CSV輸出均保存於研究工作區。主要量化結果依2026年8月3日下載的官方快照重算，後續官方修訂可能產生不同數值。受限於授權或存取的來源不重新散布；檢索失敗與排除紀錄仍保留於稽核檔。
+
+## 倫理聲明
+
+本研究未涉及人類參與者、訪談、問卷、實驗、介入或可識別個人資料，因此目前不適用人體研究審查。未來若進行管理者或學生研究，須依研究設計另行辦理倫理審查與知情同意。
+
+## 作者貢獻
+
+依CRediT分類，單一作者負責Conceptualization、Methodology、Investigation、Data Curation、Formal Analysis、Writing – Original Draft及Writing – Review & Editing。AI工具不列為作者。
+
+## 經費與利益衝突聲明
+
+本課程論文未接受外部經費。作者與本文所列學校及招生策略沒有任職、治理、顧問或商業利益衝突。
+
+## AI協作揭露
+
+本稿使用AI協作工具支援問題整理、有限檢索、來源身分與主張對照、程式化描述分析及草稿撰寫。實質主張依本次取得的官方資料或完整原文進行有限對照，但仍受檢索缺漏、資料快照、PDF頁碼預檢及部分內部章節定位限制。此程序未獨立重做各篇研究，也不證明每項來源發現皆正確。研究者確認研究範圍、推論邊界與最終文字，AI未執行訪談、實驗或原始資料蒐集。
+
+## 參考文獻
+
+行政院。（2023年9月20日）。*Boosting enrollment and retention of international students*。https://english.ey.gov.tw/News3/9E5540D592A5FECD/5ba2391b-d763-4441-9a85-a77cd10ad3e7 <!--ref:O6--><!--anchor:section:Policy_content-->
+
+中華醫事科技大學醫務暨健康事業管理系。（n.d.）。*招生特色*。https://hca.hwai.edu.tw/ <!--ref:U06--><!--anchor:section:招生特色-->
+
+中國科技大學視覺傳達設計系。（n.d.）。*中國科技大學視覺傳達設計系暨碩士班*。https://web.cute.edu.tw/dvcd/ <!--ref:U05--><!--anchor:section:Latest_News-->
+
+中國醫藥大學科技管理碩士學位學程。（n.d.）。*科技管理碩士學位學程*。https://mptm-cmu1819.cmu.edu.tw/ <!--ref:U02--><!--anchor:section:最新消息-->
+
+世新大學廣播電視電影學系。（n.d.）。*大學部*。https://rtf.wp.shu.edu.tw/?page_id=547 <!--ref:U03--><!--anchor:section:大學部-->
+
+佛光大學產品與媒體設計學系。（n.d.）。*產品與媒體設計學系、研究所*。https://pmd.fgu.edu.tw/ <!--ref:U01--><!--anchor:section:招生資訊-->
+
+南開科技大學行銷與流通管理系。（n.d.）。*行銷與流通管理系*。https://nkmlm.nkut.edu.tw/ <!--ref:U04--><!--anchor:section:招生資訊-->
+
+教育部。（n.d.-a）。*學12-2：新生（含境外生）註冊率－以校（含學制班別）統計*。大專校院校務資訊公開平臺。https://udb.moe.edu.tw/udata/StatSystemRegistrationRate <!--ref:O8--><!--anchor:section:統計說明-->
+
+教育部。（2026a）。*93–114學年度大專校院新生註冊率－按學制別分*［資料集］。https://ws.moe.edu.tw/001/Upload/15/relfile/7851/103816/12f43c6a-9e69-4212-9e31-39883e27ce4f.xlsx <!--ref:O1--><!--anchor:section:說明-->
+
+教育部。（2026b）。*大學、技專校院各系所註冊率*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/26228 <!--ref:O3--><!--anchor:section:AY109-113_data_rows-->
+
+教育部。（2026c）。*大專院校各校科系別學生數*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/9622 <!--ref:O4--><!--anchor:section:資料說明-->
+
+盧延根。（2022年2月）。*我國少子化問題與對策之研析*（R01617）。立法院。https://www.ly.gov.tw/Pages/Detail.aspx?nodeid=6590&pid=165915 <!--ref:O7--><!--anchor:section:我國少子化問題之概況-->
+
+Baliyan, S. P., & Mokoena, S. (2024). Students’ choice management: Recipe for improving enrollment in private higher education institutions. *Asian Journal of University Education, 20*(1), 206–220. https://doi.org/10.24191/ajue.v20i1.26030 <!--ref:A5--><!--anchor:section:6.Conclusion_and_Recommendations-->
+
+Chalapati, S., Leung, R., & Chalapati, N. (2018). Exploring factors affecting first-year students’ learning experiences: A case study of a private university in Taiwan. *Student Success, 9*(4), 25–39. https://doi.org/10.5204/ssj.v9i4.651 <!--ref:A3--><!--anchor:section:Conclusion-->
+
+Chang, D.-F., Wu, C.-T., Ching, G. S., & Tang, C.-W. (2009). An evaluation of the dynamics of the plan to develop first-class universities and top-level research centers in Taiwan. *Asia Pacific Education Review, 10*(1), 47–57. https://doi.org/10.1007/s12564-009-9010-7 <!--ref:A9--><!--anchor:section:Abstract-->
+
+Jongbloed, B., Enders, J., & Salerno, C. (2008). Higher education and its communities: Interconnections, interdependencies and a research agenda. *Higher Education, 56*(3), 303–324. https://doi.org/10.1007/s10734-008-9128-2 <!--ref:A8--><!--anchor:section:Introduction-->
+
+Kuo, Y.-Y. (2016). Taiwan universities: Where to go? *Humanities, 5*(1), Article 12. https://doi.org/10.3390/h5010012 <!--ref:A2--><!--anchor:section:4.Conclusions-->
+
+Levy, D. C. (2013). The decline of private higher education. *Higher Education Policy, 26*(1), 25–42. https://doi.org/10.1057/hep.2012.26 <!--ref:A6--><!--anchor:section:Demographic_decline-->
+
+Lin, W.-B., & Yang, C.-M. (2026). Demographic change and higher education governance: Evidence from departmental restructuring and enrollment dynamics in Taiwan’s TVET institutions. *Education Sciences, 16*(3), Article 371. https://doi.org/10.3390/educsci16030371 <!--ref:A1--><!--anchor:section:7.1.Conclusions-->
+
+Marginson, S. (2016). The worldwide trend to high participation higher education: Dynamics of social stratification in inclusive systems. *Higher Education, 72*(4), 413–434. https://doi.org/10.1007/s10734-016-0016-x <!--ref:A7--><!--anchor:section:Abstract-->
+
+Padlee, S. F., Kamaruddin, A. R., & Baharun, R. (2010). International students’ choice behavior for higher education at Malaysian private universities. *International Journal of Marketing Studies, 2*(2), 202–211. https://doi.org/10.5539/ijms.v2n2p202 <!--ref:A4--><!--anchor:section:Abstract-->
+
+<!--protected-hedges: 依本研究截至2026年8月3日完成的繁體中文與英文有限結構化檢索 | 在取得並檢查完整原文的語料中 | 此說法只適用於本次檢索-->

examples/pi/stage-03/editorial-decision-and-revision-roadmap.md

@@ -0,0 +1,138 @@
+# Editorial Decision Package
+
+## Sprint Contract Audit
+
+dimension_verdicts: [D1=block, D2=warn, D3=block, D4=warn, D5=warn, D6=block]
+fired_conditions: [F2, F3, F5]
+da_critical_adjudications: []
+editorial_decision=major_revision
+
+## Part 1: Editorial Decision Letter
+
+敬啟者：
+
+感謝提交〈少子化情境下臺灣私立大學招生策略證據之可判定性：公開資料趨勢與最低公開證據稽核〉。本稿依《高等教育》區域專業期刊適配標準接受五席完整審查。
+
+## Review Panel Provenance (#540)
+
+All five reviewer personas ran on a single model family (openai-codex/gpt-5.6-sol). Persona diversity is not model diversity — blind spots may be correlated across reviewers (Ren et al. 2026, arXiv:2607.13104 §5.2).
+
+五席在同一Pi工作階段依序執行，沒有獨立context windows，也沒有真正認知paper blindness。每席Phase 1與Phase 2分檔並通過機械檢查，報告彼此未交叉引用；這只能維持程序分離，不能消除單一模型與共享對話脈絡造成的相關錯誤。
+
+### Decision: Major Revision
+
+### Reviewer Summary
+
+| Reviewer | Role | Recommendation or signal | Confidence |
+|---|---|---|---:|
+| Journal-Fit | 臺灣高等教育期刊定位與結構 | Major Revision | 4/5 |
+| R1 | 行政資料、重複橫斷面、次級證據方法 | Major Revision | 4/5 |
+| R2 | 臺灣私立高教與TVET治理 | Minor Revision | 4/5 |
+| R3 | 教育管理、決策科學與學生風險 | Minor Revision | 4/5 |
+| DA | 負面證據與論證壓力測試 | D3 repairable block; no CRITICAL | 5/5 |
+
+### Points of Agreement
+
+- **[CONSENSUS-3, positive]** EIC、R1與R2均肯定稿件維持私立一般大學／TVET分層，並把官方趨勢、校方可見性及Lin與Yang的關聯限制在各自證據層級；R3未反對但未直接審查此領域／方法主張。
+- **Corroborated finding** R1與DA均指出，有限取得語料的邊界需要更清楚地與「臺灣公開證據總體不足」分開；R2另從臺灣領域定位不足提供相鄰支持。
+- **Corroborated finding** R1與DA均指出，表2由證據層級轉為延後／停止狀態的規則尚未操作化；EIC從貢獻可辨識性提出相鄰問題。
+- 四位非DA審查者都確認稿件具有真實優點，但各自發現不同的可修復缺口。沒有任何審查者主張稿件存在不可修復的資料錯誤或因果誇大。
+
+### Points of Disagreement
+
+沒有形成同一原子主張上「存在／處置／嚴重度」相互衝突的SPLIT。推薦差異來自角色評估不同維度：R2與R3認為其領域與實務缺口可用局部補充處理；EIC、R1與DA則分別在D6、D1與D3識別可修復block。依Sprint Contract，這些owner-seat block觸發F2，不能用其他維度的Minor建議平均或軟化。
+
+### Decision Rationale
+
+稿件最穩固的成果是109至113學年度兩類私校雙重指標的可重現重算，以及對分母、校型與關聯／因果邊界的持續節制。這些成果讓論文不至於因資料或推論錯誤而被否決。然而，三個核心維度尚未達區域專業期刊門檻。第一，R1指出正文無法重建316個候選群組到9篇全文的納入排除與篩選規則；這直接限制「公開證據可支持到何種程度」的可評估性。第二，DA指出本次可取得語料的不足仍可能在論證上承擔過廣的管理結論，且表2幾乎一律延後的輸出欠缺可反駁規則。第三，EIC指出最低公開證據稽核尚未與既有做法區別或驗證，因此期刊貢獻仍像一套審慎清單。這些問題均可用既有稽核紀錄、重新定位與程序操作化修復，不需要捏造新效果或進行人體研究；故決定為Major Revision而非Reject。機械判定由F2優先驅動，F3與F5亦觸發。
+
+### Top Blocking Issues
+
+| Rank | Blocking issue | Source reviewer(s) | Evidence anchor | Resolving roadmap item |
+|---:|---|---|---|---|
+| 1 | 文獻／文件篩選邊界與流程不足以重建 | R1 | absence: §三（一）研究設計與檢索 — expected 明確納入排除標準、檢索期間與逐階段篩選規則; checked §三（一）、中文摘要、§五（五） | R1 |
+| 2 | 本次取得語料不足與臺灣公開證據總體不足尚未完全拆開 | DA | text: §一（二） "此說法只適用於本次檢索，不能推論其他公開、付費或內部研究不存在。" | R2 |
+| 3 | 最低公開證據稽核的新增價值與待驗證狀態未充分區別 | EIC | text: §五（二） "它尚未接受獨立編碼、使用者研究、閾值校準或預測檢驗，也沒有證明管理成效。" | R3 |
+
+### Required Revisions
+
+| # | Revision Item | Sub-Claim(s) | Severity | Evidence Anchor | Confidence | Source | Priority | Estimated Effort |
+|---|---|---|---|---|---|---|---|---|
+| R1 | 補足有限檢索的資格、時間範圍、篩選流程與各階段數量，必要時核對現有候選清單 | SC-3 | major | absence: §三（一）研究設計與檢索 — expected 明確納入排除標準、檢索期間與逐階段篩選規則; checked §三（一）、中文摘要、§五（五） | 4 — 有限證據檢索與可重現篩選 | R1 | P1 | 2–4 days |
+| R2 | 將官方資料可判定、已取得全文可判定及未搜索／不可得範圍分成三個明確結論層級 | SC-10 | major | text: §一（二） "此說法只適用於本次檢索，不能推論其他公開、付費或內部研究不存在。" | 5 — 負面證據與推論邊界 | DA | P1 | 1–2 days |
+| R3 | 重寫程序貢獻：界定概念來源、新增元素、與一般因果／資料品質警示的差異及待驗證prototype地位 | SC-1 | major | text: §五（二） "它尚未接受獨立編碼、使用者研究、閾值校準或預測檢驗，也沒有證明管理成效。" | 4 — 高教期刊貢獻判準 | EIC | P1 | 2–3 days |
+| R4 | 操作化證據層級到處置狀態的規則，說明可產生非「延後」輸出的條件與反例 | SC-4 | major | table: 表2 — 八項策略之證據層級、暫定狀態與缺口 | 5 — 論證結構與可反駁性 | DA; R1 corroborating | P1 | 2–3 days |
+
+### Required Item Details
+
+**R1: 重建有限檢索流程**
+- **Problem**: 搜尋數量透明，但納入排除、時間範圍及逐階段決策規則不足。
+- **Source**: R1 W1。
+- **Requirement**: 從既有稽核檔建立精簡流程表或附錄；不得把有限檢索改稱系統性回顧。
+- **Acceptance criteria**: 讀者可由稿件與附錄重建316個候選群組如何形成9篇實質全文及9筆保留紀錄，所有數量與既有日誌一致。
+
+**R2: 分離三種證據範圍**
+- **Problem**: 本次可取得全文的不足仍可能被讀成臺灣公開證據整體不足。
+- **Source**: DA M1，與R1 W1及R2 W1相鄰。
+- **Requirement**: 在摘要、緒論、結果與結論使用一致的三層範圍語句，明示未知範圍。
+- **Acceptance criteria**: 每個「唯一／不足／缺乏」主張都能被歸入官方資料、已取得完整原文或未取得／未搜索範圍，且不把零取得寫成零存在。
+
+**R3: 界定prototype的可出版增量**
+- **Problem**: 程序可追溯，但與既有一般警示的差異、概念來源及驗證狀態不夠清楚。
+- **Source**: EIC W1。
+- **Requirement**: 新增精簡定位段落，將已成立貢獻限於程序可追溯性，並提出可靠性、使用者與判定效度的後續評估方式。
+- **Acceptance criteria**: 文章可用一段文字回答「相較一般因果與資料品質檢查新增什麼」，且未宣稱信效度、預測力或管理效果。
+
+**R4: 操作化表2輸出規則**
+- **Problem**: 延後／停止輸出可能只是預設保守規則，缺少可反駁性。
+- **Source**: DA M2；R1 W2 corroborating。
+- **Requirement**: 為可見性、關聯、效果與風險閘門提供明確decision rules，加入至少一個在何種公開／內部證據下可形成不同輸出的假設例。
+- **Acceptance criteria**: 第三方可依相同輸入得到相同prototype狀態，並可指出至少一組會使輸出不同的條件；停止仍不得被寫成自動停系或關校。
+
+### Suggested Revisions
+
+| # | Revision Item | Sub-Claim(s) | Severity | Evidence Anchor | Confidence | Source | Priority | Estimated Effort |
+|---|---|---|---|---|---|---|---|---|
+| S1 | 補強臺灣私校退場、區域招生差異與校務研究的有界領域定位 | SC-5 | minor | absence: §二文獻回顧 — expected 近年臺灣私校招生治理、退場或區域差異研究之定位; checked §二（一）至§二（四）及參考文獻 | 4 — 臺灣高教政策脈絡 | R2 | P2 | 1–3 days |
+| S2 | 將表2策略分成招募、學程容量、組織治理及合作支持類別 | SC-6 | minor | table: 表2 — 策略類別欄八項異質治理與招募行動 | 4 — 高教治理概念區分 | R2 | P2 | 0.5–1 day |
+| S3 | 增加prototype使用者、責任角色、決策時點與紀錄workflow | SC-7 | minor | absence: §四（三）與§六（二） — expected 使用者、決策時點、責任角色與可稽核輸出流程; checked 表2、§五（二）至§五（四）、§六（二） | 4 — evidence-to-decision流程 | R3 | P2 | 1 day |
+| S4 | 加入受影響者證據與最低諮詢條件，明示尚非共同設計 | SC-8 | minor | text: §五（四） "誰承擔成本、誰獲得利益" | 4 — 利害關係人治理 | R3 | P2 | 0.5 day |
+| S5 | 為未來內部資料升級加入資料最小化、存取、揭露與刪除規則 | SC-9 | minor | text: §五（五） "再連結申請至註冊漏斗、C+E、留存、轉退學及負擔" | 4 — 教育資料治理 | R3 | P2 | 0.5 day |
+| S6 | 正文簡化工具與頁面預檢細節，完整稽核紀錄移至附錄 | SC-2 | minor | text: §三（三） "由於pypdf不可用，雖以pdfinfo與pdftotext讀取九份PDF" | 4 — 期刊結構與讀者適配 | EIC | P3 | 0.5 day |
+
+## Part 2: Revision Roadmap
+
+### Priority 1 — Structural Revisions
+- [ ] R1：補足並核對有限檢索流程。
+- [ ] R2：分離三種可判定／未知的證據範圍。
+- [ ] R3：界定最低公開證據稽核的新增價值與prototype狀態。
+- [ ] R4：操作化表2的decision rules與反例條件。
+
+### Priority 2 — Content Supplementation
+- [ ] S1：有界補強臺灣領域定位。
+- [ ] S2：重組異質策略類別。
+- [ ] S3：加入使用與責任workflow。
+- [ ] S4：加入受影響者證據欄位。
+- [ ] S5：加入內部資料治理先決條件。
+
+### Priority 3 — Text and Formatting
+- [ ] S6：把完整性工具細節移至附錄並精簡正文。
+- [ ] 最終排版時補學生、系所、課程及教師欄位，完成APA行距、頁碼與懸掛縮排。
+
+### Revision Deadline
+
+建議6週；若維持課程論文而非期刊投稿，可由授課期限縮短，但不可因此省略R1–R4的可驗證接受標準。
+
+### Response Letter
+
+修訂時須逐項回覆R1–R4及S1–S6，標示變更位置；若不採用建議項目，需說明證據與範圍理由。
+
+## Part 3: Reviewer Report Summary
+
+- **Journal-Fit:** 主題合適、邊界一致，但程序貢獻仍未形成可辨識的期刊增量。
+- **R1 Methodology:** 官方量化重算穩健；有限檢索流程與表2映射需操作化。
+- **R2 Domain:** 領域主張準確；臺灣當代定位與策略分類可補強。
+- **R3 Perspective:** 風險與停止語義有價值；使用者、workflow、參與與資料治理仍不足。
+- **DA:** 無CRITICAL；兩項MAJOR挑戰是負面證據範圍與表2輸出的可反駁性。
+
+敬請依上述路徑進行實質修訂。修訂稿將進入verification re-review，而非直接接受。

examples/pi/stage-03p/verification-review-report.md

@@ -0,0 +1,78 @@
+# Verification Review Report
+
+## Judge Record (#539)
+
+- **Verification judge**: `openai-codex/gpt-5.6-sol`, sequential same-context transport-recovery execution.
+- **Round-1 panel provenance**: all five Round-1 personas ran on `openai-codex/gpt-5.6-sol` in one shared Pi context, without independent context windows or genuine cognitive paper blindness; separate artifacts provided procedural separation only.
+- **Independent cross-model pass**: `not_configured`; each Priority-1 row records `not_configured`.
+- **Pre-committed criteria**: `263148e3cd0edbee8a92bf3ab26b9cfe9a67c93003adc5ea3b59cd5ea354bb30`.
+- **Prompt/rubric surfaces**: `academic-paper-reviewer/references/re_review_mode_protocol.md`, three-gate contract v1.0; `shared/contracts/re_review/*.schema.json`; `scripts/check_re_review_synthesis.py`.
+- **Reviewer configuration**: `round1_cards_reused`; `field_analyst_agent` was not rerun.
+- **Routing**: `card_mapped`; P1/P2 items used the first mapped non-DA seat, with DA-only REV-002 routed to EIC and REV-004 routed to R1 after its DA label.
+- **Apply-report chain**: `pass` — `e51e77f36f2c → 8ad94b3a3eab → cd87ea2971a8 → ac17240daede`.
+- **Evidence seen by the judge**: Round-1 Roadmap, Editorial Decision Letter, frozen findings bundle and configuration cards; original and revised manuscripts plus three paired patch/apply-report rounds at Phase 2A; Response to Reviewers only at Phase 2B.
+- **Judging budget**: three persisted sequential contract phases plus one deterministic checker run; no independent model calls; exact token accounting unavailable from the transport.
+
+This verification round ran on the same model family that drove the revisions; over-optimization to this judge's latent biases is possible (Ren et al. 2026, arXiv:2607.13104 §8.1.2).
+
+## Execution Independence Disclosure
+
+This report does **not** represent independent reviewers. The clean session is a disclosed transport recovery of the same academic run. Phase boundaries were enforced through frozen inputs, persisted artifacts, hash bindings, and the checker, but all judgment was produced sequentially in the same model context.
+
+## Decision
+
+**Major Revision**
+
+The mandatory synthesis checker passed with exit status 0. Rule **B4** governs: one Priority-1 item is only `PARTIALLY_ADDRESSED` and retains a `must_fix` residual. `reject_recommended` is false.
+
+## Revision Response Checklist
+
+### Priority 1 — Required Revisions
+
+| # | Original Review Comment | Author's Claim | Response Status | Revision Location | Verified? | Cross-model (#539) | Quality Assessment |
+|---|---|---|---|---|---|---|---|
+| R1 | Reconstruct eligibility, screening flow, and stage counts | Added dates, rules, and 382→316→96→18→9+9 counts | PARTIALLY_ADDRESSED | §三（一）; §五（五）; Appendix A | ⚠️ Partial | not_configured | The revision is substantial, but the manuscript never states the selection rule for **96 near-screened clusters → 16 initial core works** before two citation-traced additions. The 316→9+9 path therefore remains incompletely reconstructable. **Residual: must_fix.** |
+| R2 | Separate official-data, retrieved-full-text, and unknown scopes | Propagated three-scope language and zero-retrieval ≠ zero-existence | FULLY_ADDRESSED | Abstracts; §一（二）; §四（二）; §五; §六 | ✅ Yes | not_configured | The author claim matches the manuscript; central uniqueness and insufficiency statements are scope-qualified throughout. |
+| R3 | Distinguish the contribution and preserve unvalidated prototype status | Made the Taiwan measurement case primary and prototype secondary | FULLY_ADDRESSED | Abstracts; §一（二）; §五（二）; §六（二） | ✅ Yes | not_configured | The paper states a concrete measurement/scope-chain increment and explicitly disclaims reliability, validity, usability, prediction, management effect, and generalizability. |
+| R4 | Operationalize evidence-to-disposition rules and alternative outputs | Added four outputs, necessary inputs, strategy-group applications, and withdrawal conditions | FULLY_ADDRESSED | §二（四）; Table 2 and following text; §五; §六 | ✅ Yes | not_configured | Rules permit defer, pilot, bounded effect assessment, and risk-gate stop; different inputs can change outputs, and stop does not imply automatic department or institutional closure. |
+
+### Priority 2 — Suggested Revisions
+
+| # | Original Review Comment | Response Status | Notes |
+|---|---|---|---|
+| S1 | Strengthen bounded Taiwan exit, regional-difference, and institutional-research positioning | PARTIALLY_ADDRESSED | Existing-search traceability and access-gap mapping improved, but no substantive supplementary literature search or positioning was added. The author transparently records this as a deliberate limitation. **Residual: consider.** |
+| S2 | Separate heterogeneous strategy categories | FULLY_ADDRESSED | Four governance-relevant groups and matched outcome/evidence inputs are present. |
+| S3 | Add users, roles, decision moments, and auditable workflow | FULLY_ADDRESSED | Proposal, checking, risk review, decision record, and re-review timing are assigned to roles. |
+| S4 | Add affected-party evidence and minimum consultation conditions | FULLY_ADDRESSED | Required evidence and the non-co-design boundary are explicit. |
+| S5 | Add internal-data governance prerequisites | FULLY_ADDRESSED | Minimization, purpose limits, role access, aggregate disclosure, logs, retention, and deletion are present. |
+
+### Priority 3 — Nice to Fix
+
+| # | Original Review Comment | Response Status |
+|---|---|---|
+| N1 | Move detailed tool/page-preflight discussion to an appendix while preserving limitations | FULLY_ADDRESSED |
+
+## New Issues (Discovered During Revision)
+
+None. The frozen Phase-2A new-issue set is empty; therefore no regression, previously-missed, or indeterminate issue affects the decision.
+
+## Decision Rationale
+
+- P1 verdicts: 3 fully addressed, 1 partially addressed.
+- P2 addressed rate: 5/5 (100%), because both fully and partially addressed count under the contract.
+- No P1 item is not addressed, made worse, or cannot be verified.
+- No regression-attributed new issue exists.
+- Nevertheless, REV-001's residual is `must_fix`; rule B4 mechanically sets **Major Revision**.
+
+The author response does not alter the committed Phase-2A verdict: its pointers lead to the same manuscript passages and do not supply the missing 96→16 selection rule. No adjustment record was therefore admissible or needed.
+
+## Residual Issues
+
+1. **REV-001 / R1 — must fix:** Add one auditable sentence or compact appendix row stating the frozen rule by which the 96 near-screened candidate clusters yielded the 16 initial core works. Identify the persisted list/log that witnesses the selection and preserve all existing totals. Do not rerun any search.
+2. **REV-005 / S1 — acknowledged limitation / consider:** The Taiwan exit, regional-difference, and institutional-research literature remains thin. Under the run's frozen no-new-search constraint, retain this as an explicit limitation rather than inventing or metadata-filling a literature claim.
+
+## Mandatory Checker Result
+
+```text
+re-review synthesis ok: round 'stage3p-round1-20260803', revision 1, decision_state 'Major Revision', apply_chain_witness 'pass'
+```

examples/pi/stage-04.5/final-integrity-verification-report.md

@@ -0,0 +1,211 @@
+# Academic Integrity Verification Report — Final Check
+
+## Verification mode and input
+
+- Mode: **Mode 2 — Stage 4.5 final verification**.
+- Manuscript: `stage-04p-rerevision/03-re-revised-course-paper-for-inspection.md`.
+- Bound SHA-256: `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`.
+- Literature corpus: frozen; no source was added, removed, or promoted.
+- Live activity: exact-identity, DOI, direct-page, retraction/correction, and text-overlap verification only; these lookups did not reopen literature discovery.
+- Cross-model verification: not enabled.
+- Cache: no cached citation row was used; no `ADV-CACHE` row.
+- Execution: sequential in the same recovered `openai-codex/gpt-5.6-sol` Pi context; fresh artifact/source checks were performed, but cognitive independence is not claimed.
+
+## Verdict
+
+# PASS
+
+Zero SERIOUS issues, zero MEDIUM issues, zero MAJOR_DISTORTION, zero UNVERIFIABLE claims, zero citation-context errors, and zero data inconsistencies were found. No manuscript correction round is required.
+
+Non-gating notes and advisories remain visible: Semantic Scholar HTTP 429 degradation, one live-host retrieval timeout backed by the preserved official file, heuristic originality limits, unavailable identity-based self-plagiarism screening, two standalone-table trace notes, nonblocking PRISMA-trAIce/RAISE adaptation warnings, and token-conservation advisories.
+
+## Verification summary
+
+| Category | Total | Passed | Issues |
+|---|---:|---:|---:|
+| Reference existence and identity | 21 | 21 VERIFIED | 0 |
+| Bibliographic accuracy | 21 | 21 | 0 |
+| Ghost citations | 21 body/reference slugs | 21 | 0 orphan; 0 dangling |
+| Citation-context occurrences | 58 | 58 VERIFIED | 0 |
+| Statistical/data verification | 10 table rows + endpoint calculations | All | 0 |
+| Internal consistency | Complete manuscript | PASS | 0 |
+| Originality D1 | 24/46 eligible long body paragraphs (52.17%) | 24 | 0 CLOSE_MATCH; 0 VERBATIM |
+| Revised-paragraph originality | 24/24 | 24 | 0 |
+| Self-plagiarism D2 | — | Not executable | Author name remains a placeholder |
+| Claim verification E1–E3 | 41/41 (Mode 2: ALL) | 41 VERIFIED | 0 distortion; 0 unverifiable |
+| Scope E4 | 41 claim clusters | 41 conformant | 0 `ADV-E4` |
+| Novelty/absence E5 | 1 bounded corpus-absence formulation | Supported within search | 0 `ADV-E5` |
+| Strength drift E6 | 4 rounds / 27 patch operations | All | 0 `ADV-E6` |
+| AI failure modes | 7 | 7 CLEAR | 0 suspected/insufficient |
+| Figure/table fidelity C3 | 2 standalone tables | 2 PASS WITH NOTES | 0 blocking failure |
+| Experiment declaration C4 | 1 passport declaration | PASS | 0 |
+| Temporal audit | Complete body | PASS | 0 findings |
+
+## Phase A — Reference verification
+
+### A0 structured API layer
+
+All nine DOI-bearing articles were freshly queried.
+
+- Semantic Scholar: 9/9 returned HTTP 429 (`API_UNAVAILABLE`). This was recorded as degradation, never as a positive or negative verdict.
+- Crossref: 9/9 exact DOI records returned matching titles and containers.
+- OpenAlex: 9/9 exact DOI records returned matching titles; `is_retracted: false` for all nine.
+- DOI/direct publisher resolution: 9/9 returned HTTP 200.
+- A5 author metadata and A4 page metadata absent from Crossref were confirmed in the retrieved full texts.
+- A6's 2012 online deposit date and 2013 issue year were distinguished; the manuscript's 2013 issue citation matches the retrieved article.
+
+### A1/A2 existence and bibliographic accuracy
+
+Every reference received an exact title/author or cited-domain query and a direct canonical resolution. SearXNG returned few indexed hits during this run, so positive determinations rest on DOI/Crossref/OpenAlex records, fresh official/issuer pages, and preserved original files—not on search-result counts.
+
+- Official/institutional sources: 12/12 VERIFIED.
+- Academic articles: 9/9 VERIFIED.
+- NOT_FOUND: 0.
+- MISMATCH: 0.
+- DOI misdirection: 0.
+
+For U02, U03, U05, O7, and O8, the standard-library client encountered certificate-chain errors; `curl -k` reached the cited HTTPS pages with HTTP 200 and matching issuer/title content. O1's live host timed out, but the preserved official workbook/PDF text carries the exact cited title and 93–114 table content. This is an access note, not an identity uncertainty.
+
+Field-level audit trail: `01-reference-adjudication.tsv` and `reference-verification/raw/`.
+
+### A3 ghost citations and retraction/correction
+
+- Body citation markers: 58.
+- Anchor markers: 58.
+- Unique body source slugs: 21.
+- Reference entries: 21.
+- Orphans/dangling entries: 0/0.
+
+For all nine DOI articles, Crossref relation metadata was empty, OpenAlex returned `is_retracted: false`, and exact-title retraction/correction/expression-of-concern queries returned no notice. This is a dated, coverage-bounded result, not a permanent guarantee.
+
+## Phase B — 100% citation-context verification
+
+All 58 citation occurrences were checked against the cited source's assigned evidentiary role.
+
+- A1 remains Taiwan TVET, 2020–2024, association-only.
+- A2–A9 remain governance, mechanism, context, or safeguard sources; none is promoted to a Taiwan private-university strategy effect.
+- O3/O8 support official definitions and descriptive recomputation, not effects.
+- O4 supports AY113 sampling context only.
+- O6 supports policy existence/purpose only.
+- U01–U06 support issuer-authored visibility only.
+
+The compound safety language near A3/A7/A8 is explicitly treated as normative synthesis: the sources support student experience, stratification, and stakeholder-governance components, while legal/privacy/financial safeguards are not misattributed as empirical findings from those articles.
+
+Complete occurrence audit: `03-citation-context-audit.tsv`.
+
+## Phase C — data, tables, and experiments
+
+### C1/C2 statistical and internal consistency
+
+The frozen derivation was rerun from the official input:
+
+- Input CSV SHA-256: `8d11b6647f693b27e70f5289980a9710972c17db46fcdb2313d35b7ce7c30c6b`.
+- Derived CSV SHA-256 before/after rerun: `ae1802065413b8cd0ff11686382eb6202e671cc18836243156536ebe75251fff`.
+- Exit: 0; stderr: empty.
+- Ten Table 1 stratum-year rows: exact match.
+- Private-general decline: 9,737; 14.2% after rounding.
+- Private-Tvet decline: 16,704; 17.4% after rounding.
+- TVET approved quota A: 123,366 → 92,294.
+- Chinese abstract, English abstract, Results, Discussion, and Conclusion endpoints agree.
+
+Search accounting also passed mechanically: 382 raw records → 316 clusters → 96 retained/220 removed → 16 initial DOI-core works + 2 citation-chased works = 18 identity-verified works → 9 full text + 2 abstract-only + 7 metadata-only. `core-dois.txt` contains 16 unique entries. The Stage 4' record transparently states that no new per-record decision ledger was created; the exact list and frozen rule surfaces are the available witnesses.
+
+### C3 figure/table caption fidelity
+
+No figures exist. Both standalone tables pass their support and interpretation checks. Because no Figure Package or `figure_table_trace[]` exists, each carries a trace-unavailable note under the standalone-table rule; neither note is a blocking issue.
+
+### C4 experiment provenance
+
+The passport's `no_experiments_declared` declaration is present and consistent with an empty `experiment_provenance[]`. No own-experiment claim or `planned_experiment_ids[]` pointer exists.
+
+> This check verifies disclosure and claim-to-provenance fidelity. It does not judge whether the experiment was correctly designed, run, statistically adequate, or reproducible by ARS.
+
+## Phase D — originality
+
+- Eligible long body paragraphs: 46.
+- Sampled: 24 (52.17%).
+- Substantially revised eligible paragraphs: 24; checked: 24 (100%).
+- Every major numbered section and Appendix A is represented.
+- Quoted and unquoted exact-fragment searches returned no related match for the 24 sampled blocks.
+- Grades: 24 ORIGINAL; 0 CLOSE_MATCH; 0 VERBATIM.
+
+Identity-based self-plagiarism checking could not run because the title-page author name is still a placeholder. D3 produced one structural-parallelism observation, below the two-indicator alert threshold; it is not an authorship determination.
+
+> This originality check is heuristic WebSearch screening, not Turnitin or iThenticate. It covers publicly searchable material only, has cross-language and paywall limitations, and can miss overlap despite the 52.17% sample. Professional plagiarism screening remains recommended before formal submission.
+
+## Phase E — 100% claim verification
+
+All 41 quantitative/factual claim clusters were checked.
+
+| Verdict | Count |
+|---|---:|
+| VERIFIED | 41 |
+| MINOR_DISTORTION | 0 |
+| MAJOR_DISTORTION | 0 |
+| UNVERIFIABLE | 0 |
+| UNVERIFIABLE_ACCESS | 0 |
+
+The funding, conflict, author-contribution, and AI-use statements are verified as project/scholar declarations consistent with the recorded pipeline, not as externally audited personal circumstances.
+
+### E4 scope
+
+No population, timeframe, geography, or domain broadening was found. International studies remain mechanism-only; A1 remains TVET-only; private-general, region, stable-size, and Taiwan-domain gaps remain open limitations. `ADV-E4`: none.
+
+### E5 novelty/absence
+
+The manuscript makes no global novelty claim. Its “only” formulation is bounded to the dated, named, retrieved-full-text corpus and explicitly preserves unknown unsearched/unretrieved areas. Classification: `SUPPORTED_WITHIN_SEARCH`. `ADV-E5`: none.
+
+### E6 claim-strength drift
+
+All four patch rounds and 27 operations since Stage 2.5 PASS were audited. No causal/evidential rung moved without roadmap authorization, and no load-bearing scope, null, validation, or non-causality qualifier was silently dropped. `ADV-E6`: none.
+
+## AI Research Failure Mode Checklist
+
+All seven modes are CLEAR: implementation bug, citation hallucination, result hallucination, shortcut reliance, bug-as-insight, methodology fabrication, and frame-lock. No user override is needed. Full evidence: `08-ai-research-failure-mode-checklist.md`.
+
+## Stage 3' traceability and previous integrity issues
+
+The Stage 3' traceability sidecar was consumed:
+
+- previously-missed new issues: 0;
+- indeterminate new issues: 0;
+- regressions: 0.
+
+Stage 2.5 corrections remain resolved: O7 uses `pid=217372`, current integrity-process wording matches completed checks, and no stale incorrect locator or claim remains.
+
+## Advisory register
+
+### Token conservation
+
+Twelve patch-level token advisories are preserved across Stage 4 and Stage 4'. E6 found no unauthorized strength drift in any associated operation. `ADV-REV-S4P-R1-1` was explicitly accepted by the scholar. Eleven earlier Stage 4 rows remain open/nonblocking because their prior stage confirmation did not record row-specific dispositions; they are listed in `12-token-conservation-advisory-register.tsv`.
+
+### Compliance
+
+Schema 12 validation passed. Because the manuscript is `other_evidence_synthesis`, PRISMA-trAIce is adaptation-only and cannot block. RAISE returns WARN: the paper discloses AI tasks and author responsibility but does not provide model/provider/version, full prompts/parameters, per-tool validation, AI screening-performance metrics, or a dedicated Discussion-level AI limitations paragraph. These are transparent reporting gaps, not citation/data integrity failures.
+
+PRISMA-trAIce is a foundational, pre-Delphi proposal and has not been empirically validated across diverse contexts; this report does not present it as a consensus standard.
+
+## Execution recovery notes
+
+- The first reference-verification command reached 12/21 records before the harness timeout; the script was resumed at record 13 and the final adjudication contains all 21.
+- Follow-up exact-identity searches reached the harness timeout twice; completed query files were retained and the remaining records were resumed until all planned queries were present.
+- The first temporal lint invocation used a nonconforming audit-run ID and was rejected by schema lint; the audit was rerun with a valid ID and returned zero findings.
+- No incomplete partial output was used for a PASS determination. Final mechanical, schema, citation, hash, and state checks all passed.
+
+## Issue list
+
+### SERIOUS
+
+None.
+
+### MEDIUM
+
+None.
+
+### MINOR
+
+None requiring a correction round.
+
+## Release status
+
+The manuscript is released from Stage 4.5 with a **PASS and zero blocking issues**. Stage 5 has not started. The final verified draft remains byte-identical to the inspected input at SHA-256 `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`.

examples/pi/stage-04/response-to-reviewers.md

@@ -0,0 +1,104 @@
+# Response to Reviewers — Revision Round 1
+
+感謝五位審查者與編輯的具體建議。本次修訂採「測量核心、範圍分層、prototype降級」姿態，依R1→R2→R4→R3完成必修項目，再逐項處理S1–S6。修訂沒有新增文獻搜尋、來源或效果資料。
+
+## Summary
+
+- **Resolved:** 9
+- **Deliberate limitation:** 1
+- **Unresolvable / reviewer disagreement:** 0
+- **Canonical whitespace word-count delta:** −40（已排除HTML audit markers；中文仍應以字元解讀）
+- **New references:** 0
+- **Patch chain:** 3 deterministic patch rounds; no structural acknowledgment or full re-emission
+
+## Required revisions
+
+### REV-001 / R1 — Screening reconstruction
+
+**Reviewer comment:** 補足有限檢索的資格、時間範圍、篩選流程與各階段數量。
+
+**Status:** RESOLVED
+
+**Response:** 已依既有日誌補入凍結檢索期間、平台、納入排除規則、去重、初篩，以及382筆原始紀錄至316候選群組、96近審、18身分驗證、9全文與9保留紀錄的流程。附錄A提供逐階段規則與數量，且不將本研究改稱系統性回顧。
+
+**Changes:** §三（一）、§五（五）、附錄A；blocks B0030, B0031, B0070, B0113–B0116。
+
+### REV-002 / R2 — Three evidence scopes
+
+**Reviewer comment:** 分開官方資料、已取得全文及未搜索／不可得範圍。
+
+**Status:** RESOLVED
+
+**Response:** 中英文摘要與正文現在一致區分三種範圍，明示零取得不等於零存在，並放棄「臺灣公開證據整體不足」的主張。第二個聚焦patch補足討論中的最後一處「唯一」範圍限定並更新protected-hedges roster。
+
+**Changes:** 摘要、§一（二）、§二（四）、§三（一）、§四（二）、§五、§六及protected-hedges；blocks B0004, B0007, B0014, B0015, B0027, B0031, B0048, B0058, B0060, B0070, B0075, B0077, B0111。
+
+### REV-003 / R3 — Contribution and prototype demotion
+
+**Reviewer comment:** 界定程序貢獻與一般方法警示的差異及待驗證prototype地位。
+
+**Status:** RESOLVED
+
+**Response:** 主要貢獻改為可重製的臺灣雙指標／分母測量案例及三層範圍報告；prototype降為次要、可反駁但未驗證的程序提案。全文不主張信度、效度、可用性、預測力、管理成效或一般化。
+
+**Changes:** 摘要、§一（二）、§四（三）、§五（二）、§六；blocks B0004, B0007, B0014, B0015, B0050, B0053, B0054, B0060, B0061, B0075, B0077, B0078, B0111。
+
+### REV-004 / R4 — Input-sensitive decision rules
+
+**Reviewer comment:** 操作化證據到處置的規則並說明非延後輸出的條件。
+
+**Status:** RESOLVED
+
+**Response:** 表2現包含「效果評估門檻已達」、「可設計證據生成試辦」、「停止採用或擴大」及「延後判定」四項輸出，並按策略群組列出改變輸出的關鍵輸入。另加入編碼不一致、規則不敏感、正反例同態、選擇性填寫及不公平後果等撤回條件。
+
+**Changes:** §二（四）、§四（三）表2、§五（二）、§六（二）；blocks B0027, B0050, B0052, B0112, B0053, B0054, B0061, B0077, B0078。
+
+## Suggested revisions
+
+### REV-005 / S1 — Taiwan domain positioning
+
+**Status:** DELIBERATE_LIMITATION
+
+**Response:** 依核准策略，本輪沒有新增文獻搜尋，也不以中繼資料補成主張。正文與附錄僅用既有搜尋、排除、失敗及保留紀錄呈現有界檢索版圖。
+
+**Justification:** 新增文獻會移動凍結語料邊界；現有不可得／中繼資料只能證明檢索狀態，不能支持領域實質結論。因此透明保留缺口，而不假裝完成臺灣領域補強。
+
+**Changes:** §三（一）、§五（五）、附錄A；blocks B0030, B0031, B0070, B0113–B0116。
+
+### REV-006 / S2 — Strategy categories
+
+**Status:** RESOLVED
+
+表2已重組為學程與容量、組織治理、招募支持及合作支持，並另列私立一般大學的已取得全文範圍。Changes: B0052, B0112。
+
+### REV-007 / S3 — User workflow
+
+**Status:** RESOLVED
+
+已加入提案、校務研究查核、教務／學務與法遵風險審查、決策紀錄及重審日期，並標示為未測程序。Changes: B0060, B0062。
+
+### REV-008 / S4 — Affected-party evidence
+
+**Status:** RESOLVED
+
+已要求未來記錄學生、教師、行政人員及合作方的負擔與支持需求，並明示本研究不是共同設計且未蒐集參與者資料。Change: B0068。
+
+### REV-009 / S5 — Internal-data governance
+
+**Status:** RESOLVED
+
+已加入資料最小化、用途限制、角色型存取、聚合揭露、存取／決策紀錄與保存刪除條件。Changes: B0062, B0068。
+
+### REV-010 / S6 — Technical detail placement
+
+**Status:** RESOLVED
+
+正文保留必要驗證限制；查詢、失敗、篩選數量、撤稿／更正與頁面映射細節集中於附錄A。Changes: B0037, B0070, B0113–B0116。
+
+## Patch provenance
+
+1. `phase6-round1/revision_patch_round1.json` → `04-revised-course-paper-round1.md`; preserved ratio 0.8108.
+2. `phase6-round1/revision_patch_round2.json` → `05-revised-course-paper.md`; preserved ratio 0.9828.
+3. `phase6-round1/revision_patch_round3.json` → `09-revised-course-paper-for-inspection.md`; preserved ratio 0.9828. This final focused round retained exactly two numbered/tabular displays by converting the strategy-group and appendix accounting into prose lists.
+
+Apply-chain hashes are continuous from `e51e77f36f2c` to `8ad94b3a3eab` to `cd87ea2971a8` to `ac17240daede`. The immutable Stage 3 input was not edited.

examples/pi/stage-04p/validation-report.md

@@ -0,0 +1,66 @@
+# Stage 4' Focused Re-Revision Validation Report
+
+## Verdict
+
+**PASS FOR SCHOLAR INSPECTION — FINAL INTEGRITY NOT STARTED**
+
+The Stage 4' round modified only `B0031` for `REV-3P-001`. `REV-3P-002` remains an explicit limitation, and no search was rerun.
+
+## Focused patch and apply
+
+- Base: `00-anchored-pre-rerevision-base.md`
+- Base SHA-256: `ac17240daedee845185f875d395e6e03aa7d44981ce99b1a52c61d6c5461f3b7`
+- Patch: `phase6-round1/revision_patch_round1.json`
+- Patch SHA-256: `ce9519ba434fed3fcc47158d4d11d84c25881eac962d16ce6b67747e9758e152`
+- Applied output: `03-re-revised-course-paper-for-inspection.md`
+- Output SHA-256: `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`
+- Apply report: `03-re-revised-course-paper-for-inspection.md.apply-report.json`
+- Apply-report SHA-256: `800415dce2534ff33bf959a916d7211b468b1529abb754c7d651cf867398ed50`
+- Ordered chain: `ac17240daede → 513004b245ef` — PASS.
+- Patch-digest binding: PASS.
+- Operations: 1 `replace_block` on `B0031`.
+- Preserved blocks: 115/116 byte-identical (`preserved_ratio: 0.9914`).
+- Structural flags: none; no structural acknowledgment and no full re-emission.
+
+The first apply command under system Python stopped before validation because `jsonschema` was unavailable and wrote no output. The retry under the existing `<temporary-virtualenv>` completed successfully. This tooling recovery is recorded in `05-focused-patch-apply-record.json`.
+
+## REV-3P-001 verification
+
+The inserted clarification is limited to the frozen 96→16 selection stage:
+
+1. The 96 closer-review clusters were assessed under the frozen direct-relevance and comparative-mechanism criteria.
+2. A DOI suitable for identity/retrieval verification was required for the initial DOI core.
+3. The exact 16-item list is named at `stage-01-research/phase2-search/core-dois.txt`.
+4. The manuscript explicitly states that the 16-item core was a prioritized identity/retrieval set, not evidence that the other 80 clusters were nonexistent or all full-text excluded.
+5. The original downstream accounting remains unchanged: 16 initial core + 2 citation-chased works = 18 identity-verified works = 9 full text + 2 abstract-only + 7 metadata-only.
+
+Frozen-log witness check: PASS. No per-record decision ledger was newly created; the clarification reports the already-frozen rule surfaces and exact core list rather than pretending a new retrospective screen occurred.
+
+## REV-3P-002 preservation
+
+- `B0070`: byte-identical to the Stage 4' base.
+- `B0116`: byte-identical to the Stage 4' base.
+- Taiwan exit, regional-difference, and institutional-research coverage remain explicitly limited.
+- No metadata-only or inaccessible record was upgraded to substantive evidence.
+
+## Citation, data, and scope integrity
+
+- Three-layer citation lint: PASS.
+- Citation/ref/anchor stream: exactly unchanged; Cite-Time Provenance Finalizer was therefore a no-op.
+- New references: 0.
+- New searches: 0.
+- Existing accounting values: preserved.
+- Stage 3' frozen findings bundle SHA-256 remains `30387db1f64b0f3f22e016e5d22fbdc66c774d9c146022a9e5fb70822e3db5eb`.
+- Empirical findings, table values, scope hedges, and causal-strength language were not modified.
+
+## Token-conservation advisory
+
+`ADV-REV-1` was emitted because the authorized clarification repeats the existing 96 and 16 values, states the implied remaining 80 clusters, and names a path containing `stage-01`. Citation and protected-hedge multisets are unchanged. This is non-blocking but must remain visible for inspection and the Stage 4.5 E6 claim-strength audit.
+
+## Execution disclosure
+
+The patch-writing and orchestration steps ran sequentially in the same `openai-codex/gpt-5.6-sol` Pi context after the disclosed transport recovery. There was no independent writer, evaluator, or context window. Deterministic apply guarantees untouched-block preservation; it does not establish cognitive independence or certify the substantive edit.
+
+## Downstream status
+
+Stage 4.5 final integrity has not started. On scholar confirmation, the patch, base, residual roadmap, apply report, response record, and Stage 3' traceability sidecar must travel together as the Revision-Evidence Bundle.

examples/pi/stage-05/cite-time-gate-pass.json

@@ -0,0 +1,27 @@
+{
+  "source_path": "stage-04.5-final-integrity/14-final-verified-course-paper.md",
+  "source_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f",
+  "working_path": "stage-05-finalize/phase7-format/04-provenance-finalized-working-draft-pass.md",
+  "working_sha256": "fe92e8ec8ae93ecf6bfb4f6c0f7851c34d998b2216728f3a0e301ffb10b1dbe3",
+  "reference_adjudication_path": "stage-04.5-final-integrity/01-reference-adjudication.tsv",
+  "human_read_log_path": "state/material-passport_human_read_log.yaml",
+  "human_read_log_sha256": "6e2b0443ac75513745d0c5e51136809dfbfd15c7f7c652f002b00c58a582ffda",
+  "evidence_adaptation": "Stage 4.5 VERIFIED adjudication used for source_acquired/source_verified_against_original because passport literature_corpus[] is absent",
+  "human_read_log_present": true,
+  "human_read_inferred_from_report_inspection": false,
+  "human_read_attestation_scope": "full_text",
+  "citation_occurrences": 79,
+  "unique_sources": 21,
+  "anchor_kinds": {
+    "section": 79
+  },
+  "no_locator": 0,
+  "resolution": {
+    "OK": 79,
+    "LOW_WARN_UNACKNOWLEDGED": 0
+  },
+  "hard_gate": "PASS",
+  "literal_refusal_hits": {},
+  "high_block_occurrences": 0,
+  "converted_outputs_emitted_at_gate_time": false
+}

examples/pi/stage-05/cite-time-gate-refusal-report.md

@@ -0,0 +1,65 @@
+# Stage 5 Cite-Time Provenance Gate Report
+
+## Result
+
+**REFUSE — converted final artifacts were not emitted.**
+
+This is a formatter hard-gate refusal, not a reversal of the Stage 4.5 factual-integrity PASS.
+
+## Bound input
+
+- Source: `stage-04.5-final-integrity/14-final-verified-course-paper.md`
+- SHA-256: `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`
+- Entry hash check: PASS
+- Source mutation: none
+
+## Gate evidence
+
+- Citation marker occurrences: 79 (58 body/table-note occurrences plus 21 reference-list markers)
+- Unique cited sources: 21
+- Paired non-`none` anchors: 79/79
+- Anchor kind: section, 79/79
+- Stage 4.5 reference identity/source adjudication: 21/21 VERIFIED
+- Literal NO-ORIGINAL, NOT-CROSS-CHECKED, NO-LOCATOR, HIGH-WARN, or HIGH-BLOCK hits: 0
+- Passport `literature_corpus[]`: absent because this run began before that aggregate was emitted
+- Peer human-read log: absent
+- Human reading inferred from inspection of the integrity report: **no**
+
+The finalizer used the Stage 4.5 VERIFIED adjudication as the disclosed source-acquisition and AI/source-cross-check witness. It did not infer that the scholar personally read the original sources. Under the cite-time matrix, all 21 unique sources therefore resolve to `LOW-WARN` and all 79 marker occurrences remain unacknowledged. The formatter hard gate refuses conversion until the source-read state is resolved.
+
+## Working artifact
+
+A working copy was created only to demonstrate the deterministic gate result:
+
+- `phase7-format/01-provenance-finalized-working-draft.md`
+- SHA-256: `5cdbf1584ef261dd95dcb5f5365847161a04e643178b04fa5387448d09d2575d`
+- Difference from verified source: citation-marker status tokens only (`LOW-WARN`); manuscript prose is not revised.
+- Three-layer citation lint: PASS.
+
+The verified Stage 4.5 source remains byte-identical.
+
+## Required scholar action
+
+The read-attestation checklist is:
+
+- `phase7-format/02-human-read-attestation-checklist.tsv`
+- SHA-256: `4bef368341d7e903389234a370b73e6b18e63a37e551353ff4ef2333459ca574`
+
+It covers: `A1`–`A9`, `O1`, `O3`, `O4`, `O6`, `O7`, `O8`, and `U01`–`U06`.
+
+For each source, the scholar may truthfully attest either:
+
+1. `full_text`, or
+2. `sections` with locators covering every listed section anchor.
+
+Inspection of the integrity report, citation audit, or bibliography alone is not a source-read attestation. There is no “accept warning and bypass” option for this terminal formatter gate.
+
+If the originals have not been personally reviewed, the safe action is to pause Stage 5 and review them. Removing or replacing citations would be a content revision and would require a new integrity pass before finalization.
+
+## Other carried decisions
+
+- APA 7 remains selected.
+- All 12 token advisories are accepted.
+- The compliance WARN remains an acknowledged reporting limitation and has not been inserted into the manuscript.
+- No AI performance metric or independent-verification claim has been invented.
+- The LaTeX question is deferred until this hard gate passes.

examples/pi/stage-05/cite-time-gate-refusal.json

@@ -0,0 +1,25 @@
+{
+  "source_path": "stage-04.5-final-integrity/14-final-verified-course-paper.md",
+  "source_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f",
+  "working_path": "stage-05-finalize/phase7-format/01-provenance-finalized-working-draft.md",
+  "working_sha256": "5cdbf1584ef261dd95dcb5f5365847161a04e643178b04fa5387448d09d2575d",
+  "reference_adjudication_path": "stage-04.5-final-integrity/01-reference-adjudication.tsv",
+  "evidence_adaptation": "Stage 4.5 VERIFIED adjudication used for source_acquired/source_verified_against_original because passport literature_corpus[] is absent",
+  "human_read_log_present": false,
+  "human_read_inferred_from_report_inspection": false,
+  "citation_occurrences": 79,
+  "unique_sources": 21,
+  "anchor_kinds": {
+    "section": 79
+  },
+  "no_locator": 0,
+  "resolution": {
+    "LOW_WARN_UNACKNOWLEDGED": 79,
+    "OK": 0
+  },
+  "hard_gate": "REFUSE",
+  "hard_gate_reason": "unacknowledged LOW-WARN source-read state",
+  "literal_refusal_hits": {},
+  "high_block_occurrences": 0,
+  "converted_outputs_emitted": false
+}

examples/pi/stage-05/full-checkpoint.md

@@ -0,0 +1,60 @@
+# Stage 5 Output Package and FULL Completion Checkpoint
+
+## Completion decision
+
+- Citation style: APA 7 adapted for Traditional Chinese educational writing.
+- LaTeX: declined by scholar.
+- Paper PDF: not generated because LaTeX was declined; no HTML-to-PDF conversion was attempted.
+- Content inspection: scholar inspected the clean Markdown, DOCX, validation report, provenance summary, and cite-time PASS record.
+- Placeholders: `［學生姓名］`, `［系所名稱］`, `［課程名稱］`, and `［教師姓名］` intentionally preserved.
+- Warnings/limitations: preserved; no silent manuscript expansion.
+- Scholar explicitly authorized Stage 5 completion and transition to Stage 6 in the same decision turn.
+
+## Final package
+
+| File | SHA-256 | Status |
+|---|---|---|
+| `package/paper.md` | `64ac9f2d5794f1f920136259869fe6a8f4c248f74a5d052f8d349818796ad219` | PASS; marker-only transformation |
+| `package/paper.docx` | `23740f600c870235efcf8a58b6ea4c3a5b79720f06e91f1d9af24e3639cf728a` | PASS; ZIP and content checks |
+| `package/format_validation.md` | `1400fa0809c62fea37eaac92c8302e97314c5904c36a0e67588ba25c8108cd14` | PASS |
+| `package/provenance_summary.md` | `ae39019e659865e244e9a5fa7c3460bd2f87954d8acbb46e3402e0e3dd2e75e9` | WARN/limitations preserved |
+| `package/submission_verification_report.json` | `0b665fe42c65bc966b081fc4eb1df7e8cd1b6c28317fde65232fabc71187c815` | advisory, no terminal token |
+| `submission-verifier-passport-snapshot.yaml` | `a59dbfcefc386729d9ea60bf581888098f10682e2bc0e8d15258f9767be5080c` | frozen verifier input |
+
+## Quality metrics
+
+- Verified source hash unchanged: `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`.
+- References: 21/21 verified.
+- Body citation contexts: 58/58 verified.
+- Cite-time marker resolution: 79/79 `ok` after explicit full-text attestation.
+- Markdown/DOCX normalized semantic text: exact match after renderer table-border normalization.
+- Tables: 2/2 preserved; rows 6 and 5 including headers.
+- ARS audit markers in clean outputs: 0.
+- New references or claims during formatting: 0.
+
+## Submission-package advisory gate
+
+Policy resolved to `advisory`. The verifier emitted no `TERMINAL-BLOCK` or `VERIFICATION-INCOMPLETE` token.
+
+- `A1`–`A7`: not applicable; no anonymized or declared double-blind package.
+- `B1`–`B5`: not checked because no venue profile was declared.
+- `C1`–`C2`: not checked because no `.bib` or passport `literature_corpus[]` was available.
+- Freshness check: report fresh under advisory policy against the same frozen passport snapshot used for the live run.
+
+The C-family parser limitation does not reverse Stage 4.5's 21/21 reference and 58/58 citation-context verification. All seven `not_checked` rows were appended to `provenance_summary.md` as required.
+
+## FULL-checkpoint self-check
+
+1. Citation integrity concern: none; all verified and cite-time gate passed.
+2. Sycophantic concession concern: none introduced by formatting; format-only scope preserved.
+3. Quality trajectory: no decline detected; normalized content is preserved.
+4. Scope discipline: no manuscript content was added.
+5. Completeness: Markdown and DOCX delivered; paper LaTeX/PDF intentionally omitted by scholar choice.
+
+## Collaboration Depth at Stage 5
+
+`insufficient_evidence`: Stage 5 had three scholar turns, below the five-turn scoring threshold. Advisory only; it did not affect completion.
+
+## Checkpoint disposition
+
+This FULL checkpoint is recorded as **confirmed by the scholar's explicit instruction to complete Stage 5 and proceed through Stage 6**. Stage 6 is authorized. Round trips at transition: 35/36.

examples/pi/stage-05/package/format-validation.md

@@ -0,0 +1,54 @@
+# Stage 5 Markdown and DOCX Format Validation
+
+## Gate status
+
+- Scholar full-text attestation: recorded for 21/21 cited sources.
+- Human-read ledger schema: PASS.
+- Cite-time finalizer: 79/79 markers resolved to `ok`; 0 LOW-WARN; 0 NO-LOCATOR; 0 HIGH-BLOCK.
+- Cite-time hard gate: PASS.
+- Verified Stage 4.5 source remained unchanged at SHA-256 `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`.
+
+## Markdown output
+
+- File: `02-final-paper-apa7.md`
+- SHA-256: `64ac9f2d5794f1f920136259869fe6a8f4c248f74a5d052f8d349818796ad219`
+- Transformation: exact Stage 4.5 manuscript with ARS audit/control comments stripped after gate PASS.
+- Substantive prose changes: none.
+- ARS markers remaining: 0.
+- Warning/block tokens remaining: 0.
+
+## DOCX output
+
+- File: `03-final-paper-apa7.docx`
+- SHA-256: `23740f600c870235efcf8a58b6ea4c3a5b79720f06e91f1d9af24e3639cf728a`
+- Generator: Pandoc 3.10.
+- Package integrity: `unzip -t` PASS.
+- Page size: A4.
+- Margins: 1 inch / 2.54 cm on all sides.
+- Body fonts: Times New Roman for Latin text; Songti TC for Traditional Chinese fallback; 12 pt.
+- Line spacing: double.
+- Page numbers: top-right header, PAGE field present.
+- Heading mapping: manuscript title → Title; main sections → Heading 1; subsections → Heading 2.
+- References: Bibliography style with 0.5-inch hanging indent.
+- Tables: 2; row counts 6 and 5, matching the Markdown source.
+- Explicit title-page page break: present.
+- ARS markers: absent.
+
+## Content-preservation checks
+
+- Pandoc plain-text comparison ratio: 0.9995657.
+- After removing renderer-generated table-border lines, normalized Markdown and DOCX text are exactly equal (16,313/16,313 characters).
+- Bilingual abstracts, keywords, body sections, two tables, Appendix A, 21 references, declarations, and bounded AI disclosure are present.
+- The acknowledged compliance WARN was not inserted into manuscript prose.
+- No model/provider/version, prompt detail, AI performance metric, or independent-verification claim was added.
+
+## Nonblocking warnings
+
+1. Pandoc reported missing `zh-TW` interface translations for the terms “Abstract” and “Table”; these warnings do not alter manuscript text because the document supplies its own headings and table labels.
+2. The title page intentionally preserves the verified placeholders `［學生姓名］`, `［系所名稱］`, `［課程名稱］`, and `［教師姓名］`. The package is not submission-ready until the scholar replaces them.
+3. No journal-specific template or Format Profile was declared; generic APA 7 course-paper formatting was applied.
+4. The formatter ran sequentially in the same recovered model context; no independent formatter is claimed.
+
+## Current status
+
+Markdown and DOCX have passed format/content checks. Stage 5 remains in progress pending the in-stage LaTeX decision and later content confirmation.

examples/pi/stage-05/package/paper.md

@@ -0,0 +1,361 @@
+
+# 少子化情境下臺灣私立大學招生策略證據之可判定性：公開資料趨勢與最低公開證據稽核
+
+
+**學生：**［學生姓名］  
+**系所：**［系所名稱］  
+**課程：**［課程名稱］  
+**授課教師：**［教師姓名］  
+**日期：**2026年8月3日
+
+
+## 中文摘要
+
+
+少子化使臺灣私立校院承受招生壓力，但註冊率上升不必然表示實際新生增加，公開可見措施也不等於效果證據。本研究探討公開可驗證資料能支持何種招生策略判斷，採整合式次級證據設計，結合有限結構化檢索、政策與文件分析，以及教育部109至113學年度資料的描述性重算。私立一般大學實際新生由68,505人降至58,768人，合成分層註冊率由92.95%降至90.02%；私立技專校院實際新生由96,155人降至79,451人，該率卻由76.27%升至77.29%。此可重製案例顯示，比率須與A、B、C、E分母成分、實際人數及機構組成並列，且官方趨勢不識別策略效果。結論另區分官方資料可判定、截至2026年8月3日已取得全文可判定，以及未搜索、未取得或不可得的未知範圍；在已取得全文範圍內，直接策略證據僅支持臺灣技專校院2020至2024年的關聯判斷。本文主要貢獻是臺灣測量案例與範圍分層；證據取得規則僅是待獨立編碼、使用者、閾值及異質輸入測試的可反駁prototype，不是策略排名或已驗證管理框架。
+
+
+**關鍵詞：**人口變遷、高等教育治理、技職教育、招生指標、次級資料、證據分級、學生風險
+
+
+## Abstract
+
+
+Demographic contraction places enrollment pressure on Taiwan’s private higher education sector, but an improving freshman registration ratio need not indicate growth in actual freshmen, and visible initiatives do not establish effects. This study asks what kinds of strategy judgments publicly verifiable evidence can support. It combines a bounded structured search and document analysis with descriptive recomputation of Ministry of Education data for academic years 109–113. Actual freshmen declined from 68,505 to 58,768 in private general universities, with the synthetic stratum-level ratio moving from 92.95% to 90.02%. In private technological and vocational institutions, actual freshmen fell from 96,155 to 79,451 while the ratio rose from 76.27% to 77.29%. This reproducible case shows why ratios must be reported with A/B/C/E denominator components, actual counts, and institutional composition; it identifies no strategy effect. Conclusions are separated into what official data determine, what the full texts retrieved by 3 August 2026 determine, and what remains unknown because it was unsearched, unretrieved, or inaccessible. Within the retrieved full-text scope, direct strategy evidence supports only association-level judgments for Taiwan TVET during 2020–2024. The paper’s principal contribution is a Taiwan measurement case and explicit scope stratification. Its evidence-acquisition rules are a falsifiable but unvalidated prototype requiring independent coding, user, threshold, and varying-input tests—not a strategy ranking or demonstrated management framework.
+
+
+**Keywords:** demographic contraction, higher education governance, technical and vocational education, enrollment measurement, secondary evidence, evidence classification, student risk
+
+
+## 一、緒論
+
+
+### （一）少子化、私校壓力與指標陷阱
+
+
+臺灣出生人口長期下降，使高等教育生源不足成為持續性的政策問題（盧延根，2022）。私立校院較依賴學費與招生收入，因而較容易承受人口與市場變化，但「私立高等教育衰退」可能指市場占比、絕對人數或特定次部門下降，不能推論每所私校必然失敗（Levy, 2013）。臺灣政策研究亦曾提出特色發展、產學合作、國際化與組織調整等因應方向（Kuo, 2016）。這些方向是治理選項，不是已證實的招生效果。
+
+
+招生壓力也涉及測量問題。教育部新生註冊率的統計範圍與公式曾於103、107與109學年度調整，跨期比較須先確認定義（教育部，2026a）。註冊率由核定名額、保留入學資格、實際註冊及境外新生共同構成，因此比率上升可能源於分母收縮，而非新生增加（教育部，n.d.-a）。若只讀取單一比率，管理者可能將名額調整誤認為招生策略成效。
+
+
+### （二）研究缺口、問題與貢獻
+
+
+公開資料呈現的「學校宣稱採取某策略」、「策略與結果同時變動」及「策略造成結果」是不同層級的主張。政策頁面或招生網頁可以證明內容曾被公開，但通常沒有策略起始時間、投入強度、比較條件與學生後續結果。為避免把取用限制寫成領域不存在，本文區分三種結論範圍：教育部官方資料可判定的測量結果、截至2026年8月3日已取得並檢查完整原文可判定的研究主張，以及未搜索、未取得或不可得而保持未知的範圍。依本次繁體中文與英文有限結構化檢索，Lin與Yang（2026）是已取得全文中唯一直接比較臺灣技專校院系所調整類型與招生結果的研究；這不是其他公開、付費或內部研究不存在的證據。
+
+
+本研究據此提出問題：「在臺灣少子化情境下，公開可驗證資料在何種程度上能支持私立一般大學與私立技專校院對招生策略成效的管理判斷？」子問題分別處理109至113學年度兩類私校的新生趨勢、不同來源能支持的最高主張層級，以及缺口如何轉化為待驗證的證據取得程序。本文的中心主張是：教育部資料能可靠支持兩類私校的分層測量稽核，卻不能單獨支持策略效果判斷；TVET合成註冊比率小幅上升而實際新生下降，是此邊界的可重製示例。三層範圍報告用來限制這項經驗結論，prototype只作為次要程序提案，不是策略排名或已驗證管理模型。
+
+
+## 二、文獻回顧與分析視角
+
+
+### （一）人口收縮、私立部門與制度分層
+
+
+人口收縮不會以相同方式作用於所有大學。私立部門的結果仍受聲譽、區位、需求、公立競爭與政策影響，故人口變化只能構成風險背景，不能作為個別學校表現的充分解釋（Levy, 2013）。高參與高等教育體系也可能依公私立別、機構階層、學科與學費維持社會分層（Marginson, 2016）。因此，私立一般大學與私立技專校院的學制、系所結構及市場位置不同，合併分析會掩蓋制度差異。分層的目的不是預設哪一類校院較脆弱，而是避免用總體平均抵銷彼此不同的名額基準、退出速度與策略證據範圍。只有先固定分析單位，後續趨勢與策略文獻才具有一致的解讀基礎。
+
+
+臺灣治理文獻列出的合併、特色發展及國際化，主要說明制度可採取的調整方向（Kuo, 2016）。另一項臺灣高等教育政策評估發現，增加投入並不保證更佳表現，但研究對象是頂尖大學計畫，不能轉用為私校招生效果估計（Chang et al., 2009）。兩者共同支持一項方法警示：政策存在與資源投入都不能代替結果辨識。
+
+
+### （二）學生選擇、國際化與入學後支持
+
+
+學生選擇研究可協助界定未來應觀察的變項。馬來西亞私立大學研究辨識學習環境、服務、成本、設施、社交與區位等選擇因素（Padlee et al., 2010）。波札那私立高等教育研究則納入就業、品質、宣傳、財務支持及同儕影響（Baliyan & Mokoena, 2024）。這些研究來自不同制度與學生群體，只能提供候選機制，不能證明臺灣私校採取廣告、獎助或就業定位後會提高註冊。
+
+
+境外招生同時牽涉入學後支持。臺灣一所私立大學國際學程的焦點團體顯示，互動教學、同儕支持、課程資訊、語言環境與社會融入影響新生經驗（Chalapati et al., 2018）。行政院頁面則確認2024至2028年擴大境外生招生與留臺政策的政策目的（行政院，2023），但沒有機構層級效果評估。招生人數與學生適應因而必須分開觀察，政策目標也不能替代學校支持能力。
+
+
+### （三）系所調整、治理與投入績效
+
+
+Lin與Yang（2026）分析2020至2024年間臺灣74所技專校院的1,087件系所調整案例。研究顯示，更名持續呈現較有利的招生關聯，且私立機構的關聯較明顯。合併通常未呈現正向招生變化，停招或關閉多發生於長期低註冊系所，部分需求導向的新設或擴額則呈現較佳關聯（Lin & Yang, 2026）。然而，調整可能回應既有趨勢，固定效果也無法排除所有隨時間改變的混淆。作者因此將發現限定為關聯，而非因果（Lin & Yang, 2026）。
+
+
+合併未帶來招生增益，不等於合併在治理上必然失敗。大學與地方、產業及其他利害關係人的合作涉及任務選擇、問責與資源負荷，成效不能只用招生量衡量（Jongbloed et al., 2008）。若合併目的在整合師資或維持課程，應另設治理指標。若以招生增長為主要理由，現有TVET關聯證據則不足以支持直接擴大。
+
+
+### （四）三層主張與最低稽核
+
+
+本文將實質主張分為策略可見性、結果共同變化或關聯，以及效果判斷三層。第一層只回答措施是否被宣布或描述；第二層須有明確單位、期間、策略分類及結果；第三層另要求時間順序、實施強度、可信比較、基線或前趨勢、同期措施與替代解釋。判斷後還須另行檢查教育品質、適配、公平、法律、隱私、學生福利與財務風險（Chalapati et al., 2018）（Jongbloed et al., 2008）（Marginson, 2016）。上述主張層級再與官方資料、已取得全文及未知範圍交叉標記：前者回答證據能說什麼，後者回答結論涵蓋到哪裡。風險閘門只界定何時不應採用或擴大策略，不是策略有效或無效的經驗證據。多個校方頁面仍只表示可見，多項國外機制研究仍不能替代臺灣比較資料；證據升級取決於研究設計，而非引用篇數或發布者權威。
+
+
+## 三、研究方法
+
+
+### （一）研究設計與檢索
+
+
+本研究採整合式次級證據設計，結合有限結構化文獻、政策與文件分析，以及官方公開資料的描述性重算。研究沒有訪談、問卷、實驗、人類參與者或原始資料蒐集，也不是完整系統性回顧。直接策略證據限於臺灣；國際研究只用於機制、候選變項與風險條件。凍結檢索於2026年8月3日05:58:48至06:25:08 UTC執行，涵蓋教育部、UDB、data.gov.tw、行政院與立法院，SearXNG網路發現，Crossref與OpenAlex API，以及已取得全文的參考文獻追索；本次修訂沒有新增檢索或移動語料邊界。
+
+
+納入資格為：與臺灣私立高教、人口收縮、招生結果、機構調整或本研究判定程序直接相關，且官方原始資料／文件或可取得完整原文足以核對身分、方法、相關段落與限制；國際研究只能作機制或情境。排除項目包括只有片段或中繼資料、原文不可得或身分未決、題名／DOI不符、實為拒絕頁面的偽PDF、將宣傳當結果證據、無直接高教機制，以及重複版本；重複紀錄保留最完整且已驗證者。43組網路查詢與16次Crossref/OpenAlex API查詢得到382筆原始紀錄（網路62、API 320）；依精確DOI與正規化題名連結66筆重複後形成316個候選群組。標題概念初篩保留96群作較近檢視，220群因明顯不符有界概念而移除。96群再依凍結納入條件進行題名與可見中繼資料近審：直接涵蓋臺灣私立高教／TVET的人口收縮、招生或機構調整，或屬預先界定的比較機制（私立部門衰退與分層、學生選擇與入學支持、學程差異化與招募），且具有可供身分與取用驗證的DOI者，列入16件初始DOI核心；精確清單凍結於`stage-01-research/phase2-search/core-dois.txt`。這16件是優先身分與取用驗證核心，不表示其餘80群不存在或均已完成全文排除。16件初始核心作品加上全文追索的2件作品，形成18件學術身分已驗證作品，其中9篇全文進入實質分析，2筆僅驗證摘要、7筆僅驗證中繼資料，後9筆均不支持實質主張。查詢字串、UTC時間、URL、可見回傳數、納排／取用狀態及零結果、驗證碼、限流、403、404與Cloudflare失敗均保留。候選DOI `10.1177/1745499920916532`及`10.1080/13538322.2020.1736512`分別校正為`10.1177/1745499920910581`及`10.1080/08841241.2020.1728465`；校正後全文未取得，故仍只屬未知／不可用於實質主張。附錄A列出可重作的流程與數量。
+
+
+### （二）官方資料、分層與合成比率
+
+
+量化資料取自2026年8月3日下載的教育部「大學、技專校院各系所註冊率」114學年度累積CSV快照（教育部，2026b）。分析保留109至113學年度、設立別為私立、校型為一般大學或技專校院，以及大學部相關學制，兩類校院分開彙整。欄位包括核定名額A、保留入學資格B、總量內實際註冊C及境外新生E。合成分層註冊率為`(C+E)/(A−B+E)`（教育部，n.d.-a）。
+
+
+本研究先按年度與校型加總A、B、C、E，再計算合成比率。這不是各校公開註冊率的算術平均。109學年度起與113學年度新增的欄位範圍仍保留版本註記（教育部，n.d.-a）。各年度納入校數不同，因此結果屬重複橫斷面，而非同一批學校的平衡追蹤。
+
+
+### （三）六校材料與分析程序
+
+
+學校材料以113學年度學生總數作為規模脈絡（教育部，2026c），在私立一般大學與私立技專校院中各選取小、中、大規模一校，共六校。佛光大學（n.d.）、中國醫藥大學（n.d.）及世新大學（n.d.）頁面呈現跨域、創新或影音專業等主張。南開科技大學（n.d.）、中國科技大學（n.d.）及中華醫事科技大學（n.d.）頁面則呈現實務、設計或醫療合作主張。樣本只測試公開材料是否可取得，不估計策略盛行率、實施或效果。
+
+
+分析先檢查年度、單位、分母與缺值；任何空白、省略號或非數字欄位均使程式停止，不自動補值。文件矩陣分開記錄來源身分、主張層級、結論範圍、替代解釋、風險與驗證狀態。引文可追溯性另分為標記存在、來源已取得及精確段落已驗證，三者不能互相替代。完整性稽核的工具、失敗、撤稿／更正查詢及頁面映射詳列於附錄A；正文只保留其邊界：截至2026年8月3日的有界查詢未發現九篇具DOI文章有已索引撤稿或更正，但不代表永久或全面無異常；pypdf不可用，pdfinfo與pdftotext雖完成11個章節錨點映射，仍不構成所有主張的普遍句級驗證。
+
+
+## 四、研究結果
+
+
+### （一）兩類私校的新生趨勢
+
+
+表1呈現109至113學年度兩類私校的實際新生C+E及合成分層註冊率。數值由固定程式重算教育部CSV欄位（教育部，2026b）。
+
+
+**表1**  
+*109至113學年度私立校院大學部新生趨勢*
+
+
+| 學年度 | 私立一般大學校數 | 實際新生C+E | 合成註冊率 | 私立技專校院校數 | 實際新生C+E | 合成註冊率 |
+|---:|---:|---:|---:|---:|---:|---:|
+| 109 | 36 | 68,505 | 92.95% | 67 | 96,155 | 76.27% |
+| 110 | 36 | 66,799 | 91.02% | 64 | 90,159 | 76.25% |
+| 111 | 35 | 56,311 | 77.64% | 62 | 81,922 | 71.78% |
+| 112 | 34 | 55,898 | 83.09% | 59 | 78,785 | 73.68% |
+| 113 | 34 | 58,768 | 90.02% | 57 | 79,451 | 77.29% |
+
+
+*註。* 校數為各年度納入範圍內出現的機構數。比率是加總A、B、C、E後重算的合成分層比率，不是各校公開率平均。資料來自教育部（2026b）。
+
+
+私立一般大學的實際新生由68,505人降至58,768人，減少9,737人（14.2%），合成比率由92.95%降至90.02%。私立技專校院的實際新生由96,155人降至79,451人，減少16,704人（17.4%），合成比率卻由76.27%升至77.29%。同期核定名額A由123,366降至92,294（教育部，2026b）。比率與人數在TVET呈現相反方向，是本研究最直接的指標稽核結果。它顯示分母收縮可與實際人數下降並存，但不能說明何種策略造成變化。
+
+
+兩類校院在111學年度的比率較低，113學年度回升。名額、機構退出或分類、境外生、學科需求、疫情後環境與同期政策皆可能構成替代解釋。由於本文未建立跨年度區域與穩定規模交叉表，教育部學生數資料只支持113學年度抽樣脈絡（教育部，2026c），第一子問題仍屬部分回答。
+
+
+### （二）策略證據的最高層級
+
+
+策略來源的證據層級不一致。六校頁面只顯示2026年8月3日檢索時可見的校方主張。行政院頁面只證明境外生政策及其目的存在（行政院，2023）。Chalapati等人（2018）、Padlee等人（2010）及Baliyan與Mokoena（2024）提供學生選擇與支持機制。Levy（2013）、Marginson（2016）、Jongbloed等人（2008）及Chang等人（2009）則提供私立部門、分層、治理及投入績效警示。這些來源均不能建立臺灣私校策略的因果效果。
+
+
+在截至2026年8月3日已取得並檢查的完整原文範圍內，只有Lin與Yang（2026）把明確策略類型與臺灣招生結果直接連結。其私立TVET更名及部分需求導向新設／擴額最多達到關聯層級；合併未呈現穩定正向招生關聯，停招或關閉是低註冊後的結果型態。該全文範圍沒有提供私立一般大學的相當直接關聯研究，因此不能借用TVET結果填補；未搜索、未取得或不可得範圍維持未知，本文不據此主張相關研究在臺灣整體不存在。
+
+
+### （三）最低公開證據與取得狀態
+
+
+表2先列出prototype的處置規則，再把本次已取得證據代入策略群組。規則是待驗證的程序提案，用於產生可檢查的下一步狀態；它不評分策略、不分配資源，也不把處置狀態當成效果發現。
+
+
+**表2**  
+*策略公開證據與後續證據取得狀態*
+
+
+| Prototype輸出 | 必要輸入規則 | 允許的解讀 |
+|---|---|---|
+| 效果評估門檻已達 | 已驗證的臺灣實施紀錄；清楚時序與強度；可比較基線或對照；合成比率、實際新生與分母方向可解釋；替代解釋已檢查；品質、公平、成本及學生結果無重大警訊 | 可進行有界效果判斷，不自動等於普遍推薦 |
+| 可設計證據生成試辦 | 有合理機制與可驗證在地實施基礎，但缺效果辨識；可預定成本上限、比較、結果、風險審查、可逆性與停止條件 | 允許產生證據；不表示預期有效 |
+| 停止採用或擴大 | 已有重大誤導、品質、公平、法律、隱私、福利或不可逆財務風險 | 風險閘門優先；不推導自動停系或關校 |
+| 延後判定 | 效果門檻未達、試辦條件不全，且未觸發停止閘門 | 只表示目前輸入不足，不能解讀為策略無效 |
+
+
+**本次策略群組代入。**（1）學程與容量的TVET更名、新增或擴額目前為臺灣TVET關聯，維持延後；補足在地實施、比較與雙重結果後才可能轉入試辦或效果評估。（2）合併、停招或關閉屬目標異質的組織治理，須先區分招生、資源與學生權益目的；一般情形延後，重大風險則停止採用或擴大。（3）國際招生、接觸、廣告與獎助只有政策存在或國外機制，須有在地漏斗、淨價格、公平、支持與比較才可能轉入試辦。（4）產學、地方合作與實習只有可見主張或間接機制，須補實施強度、品質、負擔與學生結果。（5）已取得全文未提供私立一般大學的直接關聯證據；可信校內時序、比較、成本及學生結果才可能觸發其他輸出。
+
+
+*註。* TVET調整類型的關聯依據限於Lin與Yang（2026）；其餘已取得來源只支援政策存在、機制、可見性或風險條件。表中狀態是把本次輸入代入未驗證規則所得的示範，不是信度、效度、可用性或策略成效結果。試辦僅指證據生成；停止僅指停止採用或擴大該策略。
+
+
+第三方使用時應先依相同輸入規則判定，而非預設延後。若獨立編碼者對相同輸入經常得到不同分類、規則對證據強弱不敏感、可信正反例落在同一狀態、使用者無法理解或選擇性填寫，或程序忽略受影響者證據並產生不公平決策，就應撤回處置映射與管理用途，只保留分母稽核及三層範圍報告作描述性檢查。這些撤回條件仍待未來測試；本文沒有執行可靠性或使用者驗證。
+
+
+## 五、討論
+
+
+### （一）雙重指標與校型分層的管理意義
+
+
+兩類私校實際新生均下降，但TVET合成比率小幅上升，說明「比率改善」與「生源增加」不能畫上等號。合成比率仍有管理價值，因為它呈現名額使用程度。實際新生則反映招生規模。兩者連同A−B+E分母及機構組成一起呈現，才能避免把名額收縮誤判為策略成效（教育部，2026b）（教育部，n.d.-a）。 
+
+
+校型分層同樣不能省略。私立一般大學與私立技專校院的基準比率、學制及退出速度不同；在截至2026年8月3日已取得全文的範圍內，唯一直接策略關聯研究只涵蓋TVET（Lin & Yang, 2026）。合併兩類資料會同時遮蔽分母變化及證據可轉用範圍。人口收縮是共同壓力，卻不是將不同制度位置視為同一分析單位的理由（Levy, 2013）。
+
+
+### （二）公開證據與校內證據的分工
+
+
+公開資料的優勢在於定義可查、跨校涵蓋及可重算，弱點是沒有策略暴露資訊。本文因而把可判定性拆成三層：官方資料支持測量與描述；已取得全文只支持其中實際核對的研究主張；未搜索、未取得或不可得部分保持未知。部門總量不能顯示某校何時採取措施，學校網頁沒有反事實，政策文件也沒有機構層級成效設計。prototype的合理功能只是指出主張層級、範圍與下一步資料。若未來實作，提案單位應申報策略與成本，校務研究單位核對分母、比較及資料品質，教務／學務、法遵與隱私角色審查風險，決策單位記錄採用、延後或停止理由及重審日期；這是待測程序，不是本文觀察到的現行workflow。
+
+
+這項程序尚未接受獨立編碼、使用者研究、閾值校準、異質輸入測試或預測檢驗，也沒有證明管理成效。Jongbloed等人（2008）所討論的利害關係人與問責，以及Chang等人（2009）提出的投入不等於績效警示，只能支持需要擴大判斷面向，不能驗證本程序。本文可成立的增量因此限於：保存一個TVET比率與實際人數分歧的可重製臺灣測量案例，把官方資料、已取得全文與未知範圍連成同一證據鏈，並提出可被反例撤回的程序prototype。
+
+
+學校若具有可信的策略日期、觸及、成本、招生漏斗、同期措施、留存與學生結果，可以形成比公開語料更強的機構特定判斷，但內部資料不會因「內部」身分自動具有因果力，仍須檢查基線、比較與替代解釋。未來使用個體或可連結資料前，程序至少應限定必要欄位與用途、採角色型存取、只揭露適當聚合結果、記錄存取與決策，並設定保存及刪除時點；未具備這些條件，不應以資料較多作為升級理由。公開稽核因此只是最低描述層，不是完整決策環境。
+
+
+### （三）證據生成試辦與停止語義
+
+
+人口壓力可能要求快速行動，但速度不必等同大規模採用。以A1的更名關聯為例，更名可能伴隨課程重整、品牌重定位或既有需求趨勢，研究無法分離各項機制（Lin & Yang, 2026）。若學校欲取得在地證據，應先確認名稱與課程、師資及職涯承諾一致，再設定成本上限、基線、比較、招生漏斗、留存與停止條件。此試辦的目的在減少不確定性，不是預告正面結果。
+
+
+合併案例也顯示目標須先拆分。若目標是整合師資或降低行政重複，招生未增加不必然代表治理失敗。若目標明定為招生增長，現有TVET研究沒有提供穩定正向支持（Lin & Yang, 2026）。停止擴大某項招生策略，也不等於自動停止系所或學校。後者需要另行處理教學延續、學生權益、財務及區域可近性。
+
+
+### （四）品質、公平與國際招生風險
+
+
+招生增加不能抵銷教育品質與學生風險。高參與體系可能在擴張中維持階層與費用不平等（Marginson, 2016）。國際學程經驗亦顯示，教學互動、語言及社會支持會影響新生適應（Chalapati et al., 2018）。境外招生除須拆分本國生與境外生對分子、分母的影響，也要追蹤留存、支持負荷、法規與就業協助。否則，招生壓力可能轉化為不適配、退學或品質問題。
+
+
+風險閘門不要求零風險，而是禁止用註冊改善掩蓋重大警訊。課名與課程不符、誤導招生、過度負債、隱私違規、支持容量不足或不可逆財務曝險，都應阻止策略採用或擴大。未來程序還應記錄受影響學生、教師、行政人員與合作方提出的負擔、可近性及支持需求，並說明哪些意見改變判斷；本研究沒有蒐集此類參與者資料，也不是共同設計。利害關係人問責與任務負荷要求學校說明誰承擔成本、誰獲得利益，以及招生目標是否偏離教育使命（Jongbloed et al., 2008）。這些要求是規範性安全條件，不是策略效果證據。
+
+
+### （五）限制與未來研究
+
+
+本研究是2026年8月3日執行的有限結構化檢索，不是系統性回顧；驗證碼、限流、付費、中文全文與網站阻擋縮小了已取得全文範圍。因此，本文只能說明本次官方資料與已取得全文能支持什麼，不能由未取得推論臺灣研究整體不存在或整體不足。附錄A依既有日誌呈現有界臺灣研究版圖與缺口，本次修訂沒有新增搜尋，也沒有把中繼資料補成實質主張。撤稿／更正與時間線查詢均受執行日及資料庫涵蓋限制；pypdf仍不可用，完整PDF句級精確段落驗證尚未完成。來源身分、檢索命中、全文取得與精確段落驗證仍是不同狀態。
+
+
+官方分析依2026年8月3日的資料快照，後續修訂可能改變數值。113學年度欄位範圍亦有新增，表1又是機構數改變的重複橫斷面。本文未完成區域與穩定規模趨勢，策略時間、強度、成本、反事實及品質、留存、公平和福利結果也未連結，不能進行成本效益或資源配置排名。
+
+
+未來研究可建立校級策略登錄，記錄日期、對象、觸及、投入、同期措施與停止規則，再連結申請至註冊漏斗、C+E、留存、轉退學及負擔。分期實施且假設成立時，才可考慮中斷時間序列、差異中的差異或配對比較。私立一般大學、區域與穩定規模須另行研究。未來訪談或學生研究屬新的人類參與者設計，必須另作倫理審查。
+
+
+## 六、結論
+
+
+### （一）研究問題的回答
+
+
+教育部官方資料能描述109至113學年度兩類私校的實際新生、合成分層比率及分母變化，並重現TVET比率小幅上升而實際新生下降（教育部，2026b）。這項官方資料結論支持「結果如何變動」及「指標如何形成」，不能回答「哪項策略造成變動」。政策與學校頁面主要支持可見性。在截至2026年8月3日已取得全文的範圍內，只有一項狹窄的臺灣TVET系所調整關聯，不能形成因果判斷或推及私立一般大學（Lin & Yang, 2026）；未搜索、未取得或不可得部分保持未知。研究問題的答案因而依證據與範圍而異，不是對臺灣公開證據整體下結論。
+
+
+### （二）貢獻與使用邊界
+
+
+本文主要貢獻是保存並重現一個臺灣私立高教測量案例：A、B、C、E分母稽核、私立一般／TVET分層、雙重招生指標，以及官方資料／已取得全文／未知範圍被放在同一證據鏈。這使分母改善不易被寫成招生成功，也使有限檢索的零取得不被寫成零存在。次要貢獻是提出輸入可改變輸出、也能因失敗測試而撤回的證據取得prototype。它目前只具程序可追溯性，未證明信度、效度、可用性、預測力、管理效果或一般化能力。
+
+
+「本次已取得證據尚不能判定」不等於策略無效。符合在地基礎、成本上限、預定比較、結果、風險、可逆性與停止條件時，prototype可輸出證據生成試辦；具備可信實施、比較、雙重結果、替代解釋與安全資料時，可輸出效果評估門檻已達，但仍不自動形成普遍推薦。重大品質、公平、法律、隱私、學生福利或不可逆財務風險則要求停止採用或擴大，不能由招生改善抵銷。這些是待驗證的處置規則，不是本文已證明有效的管理框架。
+
+
+## 資料可得性聲明
+
+
+本研究使用的教育部公開資料、下載日期、來源URL、衍生程式與CSV輸出均保存於研究工作區。主要量化結果依2026年8月3日下載的官方快照重算，後續官方修訂可能產生不同數值。受限於授權或存取的來源不重新散布；檢索失敗與排除紀錄仍保留於稽核檔。
+
+
+## 倫理聲明
+
+
+本研究未涉及人類參與者、訪談、問卷、實驗、介入或可識別個人資料，因此目前不適用人體研究審查。未來若進行管理者或學生研究，須依研究設計另行辦理倫理審查與知情同意。
+
+
+## 作者貢獻
+
+
+依CRediT分類，單一作者負責Conceptualization、Methodology、Investigation、Data Curation、Formal Analysis、Writing – Original Draft及Writing – Review & Editing。AI工具不列為作者。
+
+
+## 經費與利益衝突聲明
+
+
+本課程論文未接受外部經費。作者與本文所列學校及招生策略沒有任職、治理、顧問或商業利益衝突。
+
+
+## AI協作揭露
+
+
+本稿使用AI協作工具支援問題整理、有限檢索、來源身分與主張對照、程式化描述分析及草稿撰寫。實質主張依本次取得的官方資料或完整原文進行有限對照，但仍受檢索缺漏、資料快照、PDF頁碼預檢及部分內部章節定位限制。此程序未獨立重做各篇研究，也不證明每項來源發現皆正確。研究者確認研究範圍、推論邊界與最終文字，AI未執行訪談、實驗或原始資料蒐集。
+
+
+**附錄A　有界檢索、篩選與驗證流程**
+
+
+本附錄依研究工作區既有日誌重建，沒有執行新搜尋。檢索期間為2026年8月3日05:58:48至06:25:08 UTC；來源含臺灣官方網站、SearXNG、Crossref、OpenAlex及已取得全文的參考文獻追索。納入、排除與狀態定義依研究方法所列規則凍結。
+
+
+流程數量如下：（1）43組網路查詢及Crossref/OpenAlex各8組主題查詢得到382筆紀錄，其中網路62筆、API 320筆；（2）精確DOI與正規化題名去重後為316候選群組，連結66筆重複；（3）有界概念初篩保留96群近審、移除220群；（4）16件初始核心加2件全文參考文獻追索，形成18件身分已驗證作品；（5）9篇完整原文進入實質分析；（6）2筆摘要與7筆中繼資料共9筆保留但禁作實質證據；（7）其餘片段、不可得、拒絕頁、題名／DOI不符或超出範圍者，其狀態與失敗原因逐筆保留。
+
+
+此流程只重建本次語料，不代表完整臺灣文獻母體。既有搜尋顯示中文全文、學位論文、封閉平台與退出機構資料有取用缺口；這是檢索涵蓋限制，不是相關研究不存在的領域結論。六校頁面另由校型與113學年度規模作目的性可得性抽樣，只證明issuer-authored visibility。完整性作業保留查詢字串、時間、URL、失敗、兩筆錯誤DOI及其校正史、原始回應與檔案雜湊。對九篇具DOI全文的撤稿／更正查詢截至執行日未見已索引通知；pdfinfo與pdftotext定位11個章節錨點，但pypdf不可用且未達普遍句級精確驗證。
+
+
+## 參考文獻
+
+
+行政院。（2023年9月20日）。*Boosting enrollment and retention of international students*。https://english.ey.gov.tw/News3/9E5540D592A5FECD/5ba2391b-d763-4441-9a85-a77cd10ad3e7 
+
+
+中華醫事科技大學醫務暨健康事業管理系。（n.d.）。*招生特色*。https://hca.hwai.edu.tw/ 
+
+
+中國科技大學視覺傳達設計系。（n.d.）。*中國科技大學視覺傳達設計系暨碩士班*。https://web.cute.edu.tw/dvcd/ 
+
+
+中國醫藥大學科技管理碩士學位學程。（n.d.）。*科技管理碩士學位學程*。https://mptm-cmu1819.cmu.edu.tw/ 
+
+
+世新大學廣播電視電影學系。（n.d.）。*大學部*。https://rtf.wp.shu.edu.tw/?page_id=547 
+
+
+佛光大學產品與媒體設計學系。（n.d.）。*產品與媒體設計學系、研究所*。https://pmd.fgu.edu.tw/ 
+
+
+南開科技大學行銷與流通管理系。（n.d.）。*行銷與流通管理系*。https://nkmlm.nkut.edu.tw/ 
+
+
+教育部。（n.d.-a）。*學12-2：新生（含境外生）註冊率－以校（含學制班別）統計*。大專校院校務資訊公開平臺。https://udb.moe.edu.tw/udata/StatSystemRegistrationRate 
+
+
+教育部。（2026a）。*93–114學年度大專校院新生註冊率－按學制別分*［資料集］。https://ws.moe.edu.tw/001/Upload/15/relfile/7851/103816/12f43c6a-9e69-4212-9e31-39883e27ce4f.xlsx 
+
+
+教育部。（2026b）。*大學、技專校院各系所註冊率*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/26228 
+
+
+教育部。（2026c）。*大專院校各校科系別學生數*［資料集］。政府資料開放平臺。https://data.gov.tw/dataset/9622 
+
+
+盧延根。（2022年2月）。*我國少子化問題與對策之研析*（R01617）。立法院。https://www.ly.gov.tw/Pages/Detail.aspx?nodeid=6590&pid=217372 
+
+
+Baliyan, S. P., & Mokoena, S. (2024). Students’ choice management: Recipe for improving enrollment in private higher education institutions. *Asian Journal of University Education, 20*(1), 206–220. https://doi.org/10.24191/ajue.v20i1.26030 
+
+
+Chalapati, S., Leung, R., & Chalapati, N. (2018). Exploring factors affecting first-year students’ learning experiences: A case study of a private university in Taiwan. *Student Success, 9*(4), 25–39. https://doi.org/10.5204/ssj.v9i4.651 
+
+
+Chang, D.-F., Wu, C.-T., Ching, G. S., & Tang, C.-W. (2009). An evaluation of the dynamics of the plan to develop first-class universities and top-level research centers in Taiwan. *Asia Pacific Education Review, 10*(1), 47–57. https://doi.org/10.1007/s12564-009-9010-7 
+
+
+Jongbloed, B., Enders, J., & Salerno, C. (2008). Higher education and its communities: Interconnections, interdependencies and a research agenda. *Higher Education, 56*(3), 303–324. https://doi.org/10.1007/s10734-008-9128-2 
+
+
+Kuo, Y.-Y. (2016). Taiwan universities: Where to go? *Humanities, 5*(1), Article 12. https://doi.org/10.3390/h5010012 
+
+
+Levy, D. C. (2013). The decline of private higher education. *Higher Education Policy, 26*(1), 25–42. https://doi.org/10.1057/hep.2012.26 
+
+
+Lin, W.-B., & Yang, C.-M. (2026). Demographic change and higher education governance: Evidence from departmental restructuring and enrollment dynamics in Taiwan’s TVET institutions. *Education Sciences, 16*(3), Article 371. https://doi.org/10.3390/educsci16030371 
+
+
+Marginson, S. (2016). The worldwide trend to high participation higher education: Dynamics of social stratification in inclusive systems. *Higher Education, 72*(4), 413–434. https://doi.org/10.1007/s10734-016-0016-x 
+
+
+Padlee, S. F., Kamaruddin, A. R., & Baharun, R. (2010). International students’ choice behavior for higher education at Malaysian private universities. *International Journal of Marketing Studies, 2*(2), 202–211. https://doi.org/10.5539/ijms.v2n2p202

examples/pi/stage-06/process-record-en.md

@@ -0,0 +1,249 @@
+# Paper Creation Process Record
+
+## Paper Information
+
+**Paper title:** *少子化情境下臺灣私立大學招生策略證據之可判定性：公開資料趨勢與最低公開證據稽核*
+
+**Final paper formats:**
+
+- `stage-05-finalize/package/paper.md`
+- `stage-05-finalize/package/paper.docx`
+- `stage-05-finalize/package/format_validation.md`
+- `stage-05-finalize/package/provenance_summary.md`
+- `stage-05-finalize/package/submission_verification_report.json`
+
+The paper PDF and paper LaTeX source were not produced because the scholar declined LaTeX. Four title-page placeholders remain by explicit instruction. The final verified content derives from manuscript SHA-256 `513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f`.
+
+## Run Provenance and Boundaries
+
+The pipeline ran across two persisted Pi sessions. Stage 3′ required disclosed transport recovery after stalled model streams. All specialist roles were executed sequentially in the current `openai-codex/gpt-5.6-sol` context because no subagent launcher was available. No independent reviewer, writer, verifier, formatter, or cross-model judgment is claimed.
+
+The scholar's initial request was:
+
+> “I want to write a paper on the impact of declining birthrate on enrollment strategies of private universities in Taiwan. Help me from start to finish. This is my first time writing an academic paper and I'm not sure about the process.”
+
+The scholar immediately imposed the decisive evidence boundary:
+
+> “No interviews have been conducted and no original dataset has been analyzed. Do not invent participants, statistical results, or findings.”
+
+The work therefore used literature/policy analysis and only secondary public data retrieved and verified in the run. It was explicitly not a systematic review, did not conduct experiments, and did not create primary human-participant data.
+
+## Stage-by-Stage Process
+
+| Stage | Main inputs and outputs | Scholar decision or intervention |
+|---|---|---|
+| 1 Research | RQ brief, methodology blueprint, Socratic insights, search protocol, verified bibliography, reproducible MOE results, synthesis, reviewed/revised research report | Defined registration rate plus actual freshmen as paired outcomes; separated private general universities from private TVET; prohibited strategy-effect ranking and snippet evidence. |
+| 2 Write | Approved hybrid IMRaD outline, argument blueprint, new course-paper draft, bilingual abstracts, citation check | Required a fresh paper rather than relabeling the Stage 1 report; approved an approximate 6,000-Chinese-character target. |
+| 2.5 Integrity | Round 1 FAIL, three-line correction, Round 2 PASS | Confirmed the O7 locator mismatch and authorized only the verified URL and stale process-status corrections while preserving all other limitations. |
+| 3 Review | Five sequential same-context reports, Major Revision, roadmap, five-round revision coaching | Accepted R1–R4 and personally reframed the paper around measurement core, scope stratification, and prototype demotion. |
+| 4 Revise | Three focused patch rounds, deterministic apply chain, revised manuscript, point-by-point response | Required no new search, patch-only revision, immutable review input, and complete R1–R4/S1–S6 accounting. |
+| 3′ Re-review | Hash-bound three-gate contract, checker PASS, Major Revision | Accepted the residual 96→16 reconstruction issue while preserving Taiwan-domain limitations. |
+| 4′ Re-revise | One-block patch to B0031, preserved ratio 0.9914 | Authorized only `REV-3P-001`; explicitly retained `REV-3P-002` and the no-new-search boundary. |
+| 4.5 Final integrity | 21 references, 58 citation contexts, 41 claims, data reproduction, originality, temporal/compliance/failure-mode checks | Accepted all 12 token advisories and the nonblocking PRISMA-trAIce/RAISE WARN as a reporting limitation. |
+| 5 Finalize | Cite-time refusal then PASS, clean APA 7 Markdown, APA 7 DOCX, advisory package verification | Gave full-text read attestation for 21 sources, declined LaTeX/PDF, retained four placeholders and all warnings. |
+| 6 Process summary | Raw two-session dialogue index, collaboration-depth trajectory, bilingual records and PDFs | Authorized direct progression to the terminal confirmation checkpoint. |
+
+## Key Intellectual Development
+
+### Research question and measurement
+
+The scholar rejected a single-metric strategy narrative. Their rule was:
+
+> “若兩者相反，論文必須並列解釋，不以單一指標宣稱策略有效。”
+
+This led to a reproducible finding: from AY109 to AY113, actual freshmen declined in both private strata, while the synthetic private-Tvet registration ratio increased slightly. The paper treats this as a denominator-sensitive measurement case, not a strategy effect.
+
+### Evidence hierarchy
+
+The scholar distinguished visibility, association, and effect. Official data could support definitions and descriptive trends; university pages supported issuer-authored visibility only; international studies supplied mechanisms; Lin and Yang (2026) supplied a bounded Taiwan TVET association, not causality or private-general-university evidence.
+
+### Contribution correction
+
+When challenged that the framework might merely package generic caution, the scholar narrowed the contribution:
+
+> “保住可重製的雙指標／分母結果與可追溯範圍報告，主動放棄『臺灣公開證據整體不足』及『prototype已是可用管理框架』的主張。”
+
+The final contribution is therefore a reproducible Taiwan measurement case, explicit scope stratification, and a falsifiable but unvalidated evidence-acquisition prototype.
+
+## Iteration and Verification Details
+
+### Search and evidence accounting
+
+- 382 raw records
+- 316 deduplicated candidate clusters
+- 96 clusters retained for closer review
+- 16 initial DOI-core works plus 2 citation-chased works
+- 18 identity-verified academic works
+- 9 full-text substantive works, 2 abstract-only, 7 metadata-only
+- 21 cited final sources when official and institutional materials are included
+
+The Stage 4′ patch clarified how 96 became the 16-item DOI core without claiming that the remaining 80 records were nonexistent or all full-text excluded.
+
+### Review and revision
+
+- Initial editorial decision: Major Revision
+- Five sequential reviewer roles; no independent windows
+- Stage 4 patch rounds: 3
+- Stage 4′ patch rounds: 1
+- Total audited patch operations since Stage 2.5 PASS: 27
+- Stage 3′ decision: Major Revision under rule B4
+- Stage 4.5 final verdict: PASS with zero blocking issues
+
+### Integrity history
+
+Stage 2.5 initially failed because O7's URL pointed to the wrong legislative report and two process-status statements were stale. A three-line correction resolved both issues. Final integrity then verified:
+
+- References: 21/21
+- Citation contexts: 58/58
+- Claim clusters: 41/41
+- Table rows and endpoint calculations: PASS
+- Originality sample: 24/46 eligible paragraphs; 24/24 substantially revised paragraphs
+- AI research failure modes: 7/7 CLEAR
+- E4/E5/E6 advisories: 0/0/0
+
+## Interaction Pattern Summary
+
+| Metric | Value |
+|---|---:|
+| Persisted Pi sessions | 2 |
+| User turns before Stage 6 generation | 41 |
+| Assistant text turns before Stage 6 generation | 37 |
+| Estimated interaction budget | 36 |
+| Accumulated round trips at Stage 6 start | 35 |
+| Explicit checkpoint/gate decisions | 20 |
+| Automatically skipped mandatory/full checkpoints | 0 |
+| Optional residual coaching skips | 1 |
+| Integrity verification rounds | 3 total: one FAIL, two PASS |
+| Search reruns during revision | 0 |
+| Final cited sources | 21 |
+| Final body citation contexts | 58 |
+| Final claim clusters | 41 |
+
+The interaction pattern combined whole-stage delegation with detailed checkpoint constraints. Transport recovery added several routing-only turns but did not create a new academic run.
+
+## Chronology of Key Scholar Decisions
+
+1. Restricted evidence to literature/policy material and retrievable secondary public data.
+2. Selected registration rate plus actual freshman count and required joint interpretation.
+3. Required private-general and private-Tvet stratification.
+4. Defined a five-level source/evidence hierarchy and prohibited snippet promotion.
+5. Accepted “not determinable” as a legitimate result and reframed the contribution as decision screening.
+6. Limited procedural self-check claims and prohibited “validated,” “reliable,” or “generalizable” labels.
+7. Corrected the length target to approximately 6,000 Chinese characters.
+8. Required a new Stage 2 paper rather than a relabeled research report.
+9. Accepted Stage 2.5 FAIL and authorized a narrow, auditable correction.
+10. Accepted Major Revision and authored the “measurement core, scope stratification, prototype demotion” posture.
+11. Prohibited new searches during revision and required deterministic patch/apply evidence.
+12. Authorized only the residual B0031 reconstruction patch at Stage 4′.
+13. Accepted all token advisories and preserved the compliance WARN outside manuscript prose.
+14. Supplied full-text source-read attestation, declined paper LaTeX/PDF, and retained all placeholders and limitations.
+
+## Key Lessons
+
+1. Registration ratios should be reported with actual counts, denominator components, and changing institutional composition.
+2. Search non-retrieval is not evidence of field-level absence. Official-data, retrieved-full-text, and unknown scopes should remain distinct.
+3. A procedure can be traceable without being validated. Reliability, usability, and management-effect claims require independent tests.
+4. Patch-based revision materially reduced the regeneration surface and made claim-strength drift auditable.
+5. Integrity gates are productive: the O7 locator error, stale status language, and late human-read provenance gap were caught rather than silently passed.
+6. Same-context role simulation can organize multiple perspectives but cannot supply cognitive independence.
+
+## Collaboration Depth Trajectory
+
+The whole-pipeline observer assigned **DI 8/10, CV 9/10, and CR 9/10: Zone 3 — Deep Partnership**. High delegation was paired with repeated verification and original methodological reframing. The counter-reading considered whether frequent “continue/approve” turns indicated shallow routing, but those turns consistently contained falsifiable scope limits, non-claims, or acceptance criteria.
+
+Per-stage scoring was limited where stages had fewer than five scholar turns. Stage 1 and Stage 3 showed deep partnership; Stage 3′ was Zone 2 — Mid because repeated transport recovery displaced higher-order work. Full evidence and required counter-examples are in `01-collaboration-depth-trajectory.md`.
+
+Future sessions could improve by recording source-read scope and venue constraints at intake, and by adding an independent human or consented cross-model challenge at contribution and final-integrity boundaries.
+
+## AI Self-Reflection Report
+
+### Behavioral summary
+
+The AI maintained the scholar's evidence boundaries, generated a reproducible artifact chain, and used blocking integrity gates rather than silently smoothing over defects. However, the run depended on sequential same-context role execution, suffered transport stalls, and discovered one provenance integration gap only at Stage 5.
+
+### Metrics
+
+| Metric | Result |
+|---|---:|
+| Explicit DA concessions | 2/2 (100%) |
+| Consecutive explicit DA concessions | 2 |
+| Mandatory/full checkpoints auto-skipped | 0/20 |
+| Optional coaching skipped by scholar | 1 |
+| Formal user overrides of a recorded gate verdict | 0 |
+| Explicit non-default branches | 2: skip residual coaching; decline LaTeX |
+| Dialogue health alerts | 0 |
+| Intent-mode transitions | 3: exploratory→goal-oriented; revision reflection; final execution |
+| Cross-model disagreements | 0; cross-model disabled |
+
+### Sycophancy risk assessment
+
+**HIGH by the mechanical screening threshold** because the explicit DA concession rate was 100%. This is not a diagnosis: both concessions followed detailed scholar responses that directly answered the DA challenge. Nonetheless, the same-model DA may have been too accommodating. Human review of DA findings and integrity results remains warranted; the scholar performed that review at the checkpoints.
+
+### Frame-lock and convergence
+
+No formal frame-lock incident was detected. This could indicate good coverage or undetected frame-lock. The Stage 1 Socratic process converged only after the scholar specified outcome, scope, evidence hierarchy, stopping, misuse, and contribution rules. Revision coaching reopened and then reconverged the contribution rather than prematurely accepting the initial framework claim.
+
+### What the AI got wrong or handled poorly
+
+1. The initial process treated “6,000 words” too literally for a Traditional Chinese paper; the scholar corrected it to approximately 6,000 Chinese characters.
+2. The Stage 1 draft's AI disclosure overclaimed that all findings were verified; the scholar requested scrutiny and the language was narrowed.
+3. The Stage 2 draft carried an incorrect O7 URL, producing a legitimate Stage 2.5 FAIL.
+4. Several Stage 3′ model streams stalled; recovery was operationally correct but costly.
+5. A system-Python patch application initially lacked `jsonschema`; the compliant virtual environment was required.
+6. `pypdf` was unavailable; `pdfinfo` and `pdftotext` provided bounded alternatives but not universal sentence-level verification.
+7. The passport lacked the newer `literature_corpus[]` aggregate and human-read ledger. Stage 5 therefore refused once before explicit source-read attestation resolved 79/79 markers.
+8. Pandoc lacked some `zh-TW` interface translations; manuscript-supplied labels prevented content changes.
+
+### Failure Mode Audit Log
+
+At both Stage 2.5 and Stage 4.5, all seven modes were CLEAR; no mode was formally SUSPECTED or OVERRIDDEN:
+
+1. Implementation bug — CLEAR (no flags).
+2. Citation hallucination — CLEAR; the O7 locator error was detected and corrected through integrity review.
+3. Result hallucination — CLEAR; fixed official-data derivation reproduced exactly.
+4. Shortcut reliance — CLEAR; metadata-only and snippets were excluded from substantive claims.
+5. Bug-as-insight — CLEAR; no software error was interpreted as a finding.
+6. Methodology fabrication — CLEAR; no participant, interview, experiment, or original-data claim was invented.
+7. Frame-lock — CLEAR; DA and review stages materially narrowed the contribution.
+
+This self-reflection is produced by the same AI whose behavior it evaluates. It should be read as auditable self-report, not independent assessment.
+
+## Collaboration Quality Evaluation
+
+### Overall score
+
+**91/100.** The scholar's interventions substantially changed the research question, evidence hierarchy, inferential boundaries, contribution, revision posture, and release decisions rather than merely accepting generated text.
+
+| Dimension | Score | Evidence |
+|---|---:|---|
+| Direction Setting | 95 | Set a strict evidence boundary at intake and repeatedly supplied scope/non-claim rules. |
+| Intellectual Contribution | 94 | Formulated paired outcomes, source hierarchy, stopping rules, withdrawal conditions, and final contribution posture. |
+| Quality Gatekeeping | 96 | Inspected artifacts at every gate, requested AI-disclosure/length scrutiny, accepted integrity FAIL, and required focused re-verification. |
+| Iteration Discipline | 90 | Used the full review/revision path and bounded the final patch; transport retries added unavoidable overhead. |
+| Delegation Efficiency | 84 | Delegated whole stages effectively, though late source-read and venue decisions caused additional gates and package limitations. |
+| Meta-Learning | 88 | Converted review findings into reusable rules about denominator interpretation, evidence scope, prototype claims, and patch discipline. |
+
+### What worked well
+
+- The intake instruction “Do not invent participants, statistical results, or findings” prevented methodology fabrication.
+- The rule “若兩者相反，論文必須並列解釋” directly produced the paper's strongest measurement safeguard.
+- The scholar did not defend the initial framework claim reflexively; they demoted it after review and defined explicit withdrawal conditions.
+- Artifact inspection was specific: hashes, patches, limitations, and non-independence disclosures were repeatedly named before continuation.
+
+### Missed opportunities
+
+- An independent reviewer or cross-model check could have reduced correlated-error risk.
+- A target venue and Format Profile would have enabled meaningful submission-limit verification.
+- Source-read scope at intake would have avoided the Stage 5 refusal turn.
+
+### Recommendations for a future run
+
+1. Declare venue, title-page metadata, output formats, and source-read status at intake.
+2. Reserve one independent human review specifically for contribution/novelty and one for final source-to-claim fidelity.
+3. Continue using explicit non-claims and withdrawal conditions at every approval.
+4. Keep deterministic patch application and full apply-chain evidence for revisions.
+5. Separate process disclosure from manuscript disclosure early, especially for non-systematic evidence syntheses.
+
+### Human versus AI value-add
+
+The AI supplied search execution, structured artifacts, deterministic calculations, citation checks, role-based critique, patch application, and formatting. The scholar supplied the non-substitutable intellectual decisions: paired-outcome interpretation, evidence hierarchy, ethical vetoes, “unknown” as a valid result, contribution demotion, prototype falsifiability, correction authorization, and final release constraints. Those decisions materially shaped the final paper beyond what the execution system could infer independently.

examples/pi/stage-06/terminal-checkpoint.md

@@ -0,0 +1,51 @@
+# Stage 6 Artifact Index and Terminal Confirmation Checkpoint
+
+## Stage 6 outcome
+
+The bilingual paper-creation process record is generated and validated. Stage 6 remains **in progress** until a terminal acknowledgement is persisted.
+
+## Deliverables
+
+| Artifact | SHA-256 |
+|---|---|
+| `00-stage5-to-stage6-handoff.md` | `3c4f6ae7a048f1a7e439e752b9c49a41c346053ef2ef33942175504f1dafa8ed` |
+| `01-collaboration-depth-trajectory.md` | `340218fb560931bba4ca710e1e18545e85092994c6e03eba3c4b912d3e84ffe1` |
+| `02-stage6-process-record-validation.md` | `707131330f7e59ccfed3464bf2efd7f633424056ab8dc40e4251b280c66561f9` |
+| `paper_creation_process_en.md` | `d378ec4100d64c2041083352af09a4d1b5f0acfb778ec047590d310a861f7ea3` |
+| `paper_creation_process_en.tex` | `558f18799978369ad1156f29f1e69ec4fe55b6479ccc19150b7a7d23d9d9253c` |
+| `paper_creation_process_en.pdf` | `d9516715dfb739fb6d84c9d46b5c99e30dad1f53c4db6bd9e64059e1e4a493cd` |
+| `paper_creation_process.md` | `99c4e49ec6cd1a73dbdb0caaee1abfadf556bc7cd5ee44d00077e8d2925ef4a7` |
+| `paper_creation_process_zh.tex` | `408d87620a4b871abc28021a57bf9943c9f3b0beb8835d8cfca0624aa6c3be73` |
+| `paper_creation_process_zh.pdf` | `32262fc11e64001d3aadd383e7d41db9abfa33199b06229e0b486b07eaa17e99` |
+| `evidence/raw-dialogue-text.json` | `fd4b7be694d8f7aa049730c2c7c62482df1de8adddd36f437902dc0e8e8369c1` |
+| `evidence/raw-dialogue-text.md` | `88994b2d2ea25fd7f162243d354ee20123a946eedc1d6495987f91b54b990e29` |
+
+## Validation summary
+
+- Both records include every required process-summary chapter.
+- Collaboration observer: DI 8, CV 9, CR 9, Zone 3 after counter-reading; advisory only.
+- Collaboration Quality Evaluation: 91/100, evidence-based and nonblocking.
+- AI Self-Reflection includes a HIGH mechanical sycophancy-risk screen from 2/2 explicit DA concessions, plus all seven failure-mode histories.
+- English PDF: 10 A4 pages; Traditional Chinese PDF: 8 A4 pages.
+- Tectonic compilation: PASS for both.
+- PDF text extraction: PASS with zero Unicode replacement characters.
+- Source Han Serif TC VF was unavailable; Songti TC was used as a disclosed fallback.
+- English PDF retains two nonblocking underfull-box warnings; no overfull boxes.
+
+## Final paper status
+
+- Final paper Markdown and DOCX remain unchanged from the accepted Stage 5 package.
+- No paper LaTeX or paper PDF was produced, by scholar decision.
+- Four title-page placeholders and all reported warnings/limitations remain unchanged.
+
+## Terminal semantics
+
+This is the terminal confirmation checkpoint, not a completion record.
+
+- Current required state: `pipeline_state = awaiting_confirmation`.
+- Stage 6 required state: `in_progress`.
+- Terminal acknowledgement accepted forms: `finish`, `end`, `done`, `confirm`, or an unambiguous equivalent accepting the deliverables.
+- Only after such acknowledgement may Stage 6 become `completed` and global state become `completed`.
+- Completion must not be claimed until `state/pipeline-state.json` has been updated, parsed, and validated.
+
+Round trips: 35/36.

examples/pi/state/pipeline-state.json

@@ -0,0 +1,38 @@
+{
+  "topic": "臺灣少子化對私立大學招生策略之影響",
+  "language": "zh-Hant",
+  "bilingual_abstract": true,
+  "pipeline_version": "3.19.0",
+  "entry_point": "1",
+  "current_stage": "6",
+  "pipeline_state": "completed",
+  "pipeline_completed_at": "2026-08-03T18:19:42Z",
+  "consecutive_continue_count": 0,
+  "round_trip_budget": {
+    "estimated_max": 36,
+    "accumulated": 36,
+    "advisory_only": true
+  },
+  "stages": {
+    "1": {"name": "RESEARCH", "skill": "deep-research", "status": "completed", "mode": "socratic-to-full-investigation", "subphase": "stage1_complete", "outputs": ["Research Question Brief v1", "Methodology Blueprint v1", "Socratic INSIGHT Collection v1", "Devil's Advocate Checkpoint 1 v1: PASS", "Phase 1 Artifact Index v1", "Phase 2 Search Protocol v1", "Annotated Bibliography and Search Strategy v1", "Source Verification Report v1: PASS WITH BOUNDED EVIDENCE", "Phase 2 Artifact Index v1", "Phase 3 Synthesis Report v1", "Devil's Advocate Checkpoint 2 v1: PASS", "Phase 3 Artifact Index v1", "Phase 4 Research Report Draft v1", "Phase 4 Artifact Index v1", "Phase 5 Editorial Review v1: MINOR REVISION", "Phase 5 Ethics Review v1: CONDITIONAL", "Devil's Advocate Checkpoint 3 v1: REVISE", "Phase 5 Artifact Index v1", "Phase 6 Revised Research Report v1", "Phase 6 Revision Log v1", "Checkpoint 3 Recheck v1: PASS", "Phase 6 Artifact Index v1", "Stage 1 Collaboration Depth Observation: INSUFFICIENT EVIDENCE"], "checkpoint_type": "deep-research_phase6_completion", "checkpoint_confirmed": true, "phase2_started_at": "2026-08-03T05:57:57Z", "phase2_completed_at": "2026-08-03T06:36:08Z", "phase2_confirmed_at": "2026-08-03T06:41:53Z", "phase3_started_at": "2026-08-03T06:41:53Z", "phase3_completed_at": "2026-08-03T06:49:52Z", "phase3_confirmed_at": "2026-08-03T06:53:30Z", "phase4_started_at": "2026-08-03T06:53:30Z", "phase4_completed_at": "2026-08-03T07:00:41Z", "phase4_confirmed_at": "2026-08-03T07:08:50Z", "phase5_started_at": "2026-08-03T07:08:50Z", "phase5_execution_mode": "sequential_single_pi_session_disclosed", "phase5_completed_at": "2026-08-03T07:14:32Z", "phase5_verdicts": {"editorial": "MINOR_REVISION", "ethics": "CONDITIONAL", "devils_advocate_checkpoint_3": "REVISE"}, "checkpoint3_confirmed_at": "2026-08-03T07:20:41Z", "phase6_started_at": "2026-08-03T07:20:41Z", "length_target_correction": "approximately_6000_Chinese_characters_not_word_equivalence", "funding_conflict_declaration": "no_external_funding_and_no_named_employment_governance_consulting_or_commercial_conflict", "phase6_completed_at": "2026-08-03T07:33:43Z", "phase6_revision_loops_used": 1, "checkpoint3_recheck_verdict": "PASS", "stage1_confirmed_at": "2026-08-03T07:36:51Z", "stage1_collaboration_observer": "insufficient_evidence_raw_dialogue_compacted_non_blocking"},
+    "2": {"name": "WRITE", "skill": "academic-paper", "status": "completed", "mode": "plan-to-full", "subphase": "stage2_complete", "outputs": ["Stage 1 to Stage 2 Handoff v1", "Paper Configuration Record Plan Mode v1", "Proposed Paper Outline and Evidence Allocation v1", "Stage 2 Planning Artifact Index v1", "Outline Approval Record v1", "Argument Blueprint v1", "Argument Builder Artifact Index v1", "Draft Writer Precommitment v1", "Draft Writer Phase 4b Output v1", "Citation Compliance Report v1", "Bilingual Abstract and Check v1", "Complete Stage 2 Course Paper Draft v1", "Stage 2 Draft Checkpoint Index v1"], "started_at": "2026-08-03T07:36:51Z", "planning_completed_at": "2026-08-03T07:42:32Z", "outline_checkpoint_confirmed": true, "outline_confirmed_at": "2026-08-03T07:54:56Z", "argument_blueprint_completed_at": "2026-08-03T07:54:56Z", "argument_checkpoint_confirmed": true, "argument_confirmed_at": "2026-08-03T07:59:18Z", "drafting_started": true, "draft_writer_execution_mode": "sequential_single_pi_session_with_degraded_paper_blind_precommitment_disclosed", "draft_completed_at": "2026-08-03T08:11:39Z", "citation_check": "PASS_WITH_DISCLOSED_FLAGS", "abstract_check": "PASS", "stage2_checkpoint_confirmed": true, "stage2_confirmed_at": "2026-08-03T08:15:42Z"},
+    "2.5": {"name": "INTEGRITY", "agent": "integrity_verification_agent", "status": "completed", "mode": "pre-review", "subphase": "stage2_5_complete", "outputs": ["Academic Integrity Verification Report Round 1: FAIL", "Reference Adjudication v1", "Citation Context Audit v1", "Claim Registry and Verification v1", "Page Provenance Map v1", "Temporal Audit v1", "Corrected Pre-review Course Paper v1", "Integrity Correction Log v1", "Focused Re-verification Report Round 2: PASS", "AI Research Failure Mode Checklist: 7/7 CLEAR", "Compliance Report Schema 12: WARN", "Stage 2.5 Round 2 Artifact Index v1"], "started_at": "2026-08-03T08:15:42Z", "round1_completed_at": "2026-08-03T09:22:24Z", "round1_verdict": "FAIL", "correction_rounds_used": 1, "round2_completed_at": "2026-08-03T09:35:50Z", "round2_verdict": "PASS", "compliance_decision": "WARN_NON_BLOCKING_OTHER_EVIDENCE_SYNTHESIS", "compliance_warn_acknowledged": true, "failure_mode_checklist": "PASS_7_OF_7_CLEAR", "serious_issues": 0, "medium_issues": 0, "minor_issues": 0, "resolved_issue_ids": ["IL-MEDIUM-1", "IL-MINOR-1", "ADV-TEMP-1"], "review_input": "stage-02.5-integrity-pre-review/02-corrected-pre-review-course-paper.md", "confirmed_at": "2026-08-03T09:41:18Z", "checkpoint_confirmed": true},
+    "3": {"name": "REVIEW", "skill": "academic-paper-reviewer", "status": "completed", "mode": "full", "subphase": "stage3_complete_strategy_confirmed", "outputs": ["Stage 2.5 to Stage 3 Handoff v1", "Field Analysis and Reviewer Configuration v1", "Phase 0 Artifact Index v1", "Reviewer Configuration Approval Record v1", "Journal-Fit Review Report v1", "Methodology Review Report v1", "Domain Review Report v1", "Perspective Review Report v1", "Devil's Advocate Review Report v1", "Editorial Decision and Revision Roadmap v1: MAJOR REVISION", "Machine Revision Roadmap v1", "Panel Execution and Validation Report v1", "Stage 3 Review Artifact Index v1", "Stage 3 Decision Approval and Coaching Intake v1", "Socratic Revision Coaching Progress v1", "Scholar-Formulated Revision Strategy v1", "Reprioritized Revision Roadmap v1", "Coaching Completion and Strategy Checkpoint v1"], "started_at": "2026-08-03T09:41:18Z", "immutable_review_input": "stage-02.5-integrity-pre-review/02-corrected-pre-review-course-paper.md", "immutable_review_input_sha256": "867ab46797b7270fdbe3edb8814fa24cfe1712a2aa8ceb09f27b7cbcee320d76", "panel_execution_mode": "sequential_same_pi_context_no_independent_windows_disclosed", "reviewer_configuration_confirmed": true, "reviewer_configuration_confirmed_at": "2026-08-03T09:45:00Z", "sprint_contract": "reviewer/reviewer_full/v2", "panel_size": 5, "usable_panel_reports": 5, "phase_conformance": "PASS_5_OF_5", "panel_synthesis_check": "PASS_FIRST_ATTEMPT", "dimension_verdicts": {"D1": "block", "D2": "warn", "D3": "block", "D4": "warn", "D5": "warn", "D6": "block"}, "fired_conditions": ["F2", "F3", "F5"], "editorial_decision": "Major Revision", "editorial_decision_confirmed": true, "editorial_decision_confirmed_at": "2026-08-03T10:02:31Z", "da_critical_count": 0, "must_fix_items": 4, "review_completed_at": "2026-08-03T09:59:41Z", "revision_coaching_started_at": "2026-08-03T10:02:31Z", "revision_coaching_round": 5, "revision_coaching_expected_rounds": 5, "revision_coaching_completed_at": "2026-08-03T10:25:47Z", "revision_posture": "measurement_core_scope_stratification_prototype_demotion", "required_revision_order": ["R1", "R2", "R4", "R3"], "new_literature_search_allowed": false, "revision_strategy_confirmed": true, "revision_strategy_confirmed_at": "2026-08-03T10:29:11Z", "stage4_started": true, "checkpoint_confirmed": true},
+    "4": {"name": "REVISE", "skill": "academic-paper", "status": "completed_and_scholar_confirmed", "mode": "revision_patch", "subphase": "stage4_complete", "outputs": ["Stage 3 to Stage 4 Handoff v1", "Anchored Pre-revision Base v1", "Revision Writer Precommitment v1", "Revision Patch Round 1 v1", "Deterministic Apply Report Round 1 v1", "Revision Patch Round 2 v1", "Deterministic Apply Report Round 2 v1", "Revision Patch Round 3 v1", "Deterministic Apply Report Round 3 v1", "Revised Course Paper for Inspection v1", "Response to Reviewers Schema 8 v1", "Response to Reviewers Human Report v1", "Apply Chain Manifest v1", "Stage 4 Validation Report v1", "Complete Revision Diff v1", "Stage 4 Artifact Index and Checkpoint v1"], "started_at": "2026-08-03T10:29:11Z", "completed_at": "2026-08-03T10:44:48Z", "revision_posture": "measurement_core_scope_stratification_prototype_demotion", "required_revision_order": ["R1", "R2", "R4", "R3"], "new_literature_search_run": false, "patch_rounds": 3, "structural_flags": false, "full_reemission": false, "apply_chain_pass": true, "patch_digest_binding_pass": true, "final_revised_manuscript": "stage-04-revision/09-revised-course-paper-for-inspection.md", "final_revised_manuscript_sha256": "ac17240daedee845185f875d395e6e03aa7d44981ce99b1a52c61d6c5461f3b7", "response_coverage": "10_OF_10", "resolved_items": 9, "deliberate_limitations": 1, "new_references": 0, "citation_check": "PASS_58_MARKERS_58_ANCHORS_21_SOURCES_21_REFERENCES", "data_reproduction": "PASS_UNCHANGED_OUTPUT_HASH", "latest_preserved_ratio": 0.9828, "cumulative_original_block_preserved_ratio": 0.7928, "main_section_cjk_characters": 6966, "stage3p_started": true, "checkpoint_confirmed": true, "confirmed_by": "scholar_resume_instruction", "confirmed_at": "2026-08-03T15:35:22Z"},
+    "3p": {"name": "RE-REVIEW", "skill": "academic-paper-reviewer", "status": "completed_decision_confirmed", "mode": "re-review", "subphase": "stage3p_decision_checkpoint", "outputs": ["Frozen Round-1 Findings Bundle v1", "Hash-bound Input Manifest v1", "Phase 1 Precommitment v1", "Phase 1 Gate Record v1", "Phase 2A Verdict Record v1", "Phase 2A Gate Record v1", "Phase 2B Traceability Sidecar v1", "Phase 2B Gate and Checker Record v1", "Verification Review Report v1: MAJOR REVISION", "Residual Revision Roadmap v1", "Stage 3p Artifact Index and Mandatory Checkpoint v1"], "started_at": "2026-08-03T13:33:53Z", "transport_recovery_resumed_at": "2026-08-03T15:35:22Z", "completed_at": "2026-08-03T15:35:22Z", "execution_mode": "sequential_same_pi_context_disclosed_transport_recovery", "independent_reviewers_claimed": false, "cross_model_active": false, "round1_cards_reused": true, "field_analyst_rerun": false, "input_manifest_validation": "PASS", "apply_chain_witness": "pass", "phase1_schema_lint": "PASS", "phase2a_schema_lint": "PASS", "phase2b_schema_lint": "PASS", "synthesis_checker": "PASS_EXIT_0", "revised_manuscript_sha256": "ac17240daedee845185f875d395e6e03aa7d44981ce99b1a52c61d6c5461f3b7", "input_manifest_jcs_sha256": "f6c8fea09db783758bb8273e385d72831048a6dc8172c6a5450d2c24b1aa26be", "precommitment_jcs_sha256": "263148e3cd0edbee8a92bf3ab26b9cfe9a67c93003adc5ea3b59cd5ea354bb30", "verdict_record_jcs_sha256": "d4f844eca0f4bd94b7695bbf18fcc57df95887c0c3b3afea35dc1cb8e8d91f0b", "traceability_jcs_sha256": "1d323000086d11a01ddd33211dcb726a645356f3dc59e259e1ded10ef059292d", "must_fix_verdicts": {"fully_addressed": 3, "partially_addressed": 1, "not_addressed": 0, "made_worse": 0, "cannot_verify": 0}, "should_fix_verdicts": {"fully_addressed": 4, "partially_addressed": 1, "not_addressed": 0, "made_worse": 0, "cannot_verify": 0}, "p2_addressed_rate": "5_OF_5", "new_issues": 0, "regressions": 0, "decision_rule": "B4", "decision": "Major Revision", "reject_recommended": false, "checkpoint_type": "MANDATORY", "checkpoint_confirmed": true, "decision_accepted": true, "confirmed_at": "2026-08-03T16:38:00Z", "coaching_skipped_by_scholar": true},
+    "4p": {"name": "RE-REVISE", "skill": "academic-paper", "status": "completed_and_scholar_confirmed", "mode": "revision_patch", "subphase": "stage4p_inspection_checkpoint", "outputs": ["Stage 3p to Stage 4p Handoff v1", "Anchored Pre-rerevision Base v1", "Revision Writer Precommitment v1", "Focused Revision Patch v1", "Deterministic Apply Report v1", "Re-revised Course Paper for Inspection v1", "Stage 4p Response to Re-review v1", "Focused Patch Apply Record v1", "Stage 4p Validation Report v1", "Focused Revision Diff v1", "Token Conservation Advisory ADV-REV-1", "Stage 4p Artifact Index and Inspection Checkpoint v1"], "started_at": "2026-08-03T16:38:00Z", "completed_at": "2026-08-03T16:38:00Z", "execution_mode": "sequential_same_pi_context_disclosed_transport_recovery", "independent_writer_or_evaluator_claimed": false, "coaching_skipped_by_scholar": true, "authorized_scope": ["REV-3P-001"], "preserved_limitation": "REV-3P-002", "new_literature_search_run": false, "patch_rounds": 1, "ops": 1, "changed_blocks": ["B0031"], "preserved_limitation_blocks": ["B0070", "B0116"], "structural_flags": false, "full_reemission": false, "apply_chain_pass": true, "patch_digest_binding_pass": true, "base_sha256": "ac17240daedee845185f875d395e6e03aa7d44981ce99b1a52c61d6c5461f3b7", "final_rerevised_manuscript": "stage-04p-rerevision/03-re-revised-course-paper-for-inspection.md", "final_rerevised_manuscript_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f", "preserved_ratio": 0.9914, "blocks_preserved_byte_identical": "115_OF_116", "rev_3p_001_status": "RESOLVED_PENDING_SCHOLAR_INSPECTION", "rev_3p_002_status": "DELIBERATE_LIMITATION_PRESERVED", "citation_anchor_stream": "UNCHANGED", "new_references": 0, "token_conservation_advisories": ["ADV-REV-1"], "stage4_5_started": true, "checkpoint_type": "FULL", "checkpoint_confirmed": true, "confirmed_at": "2026-08-03T16:45:37Z", "scholar_inspected_artifacts": ["focused_diff", "response", "patch_apply_record", "validation_report", "checkpoint"], "accepted_advisories": ["ADV-REV-1"]},
+    "4.5": {"name": "FINAL INTEGRITY", "agent": "integrity_verification_agent", "status": "completed_pass_and_scholar_confirmed", "mode": "final-check", "subphase": "stage4_5_complete", "outputs": ["Final Integrity Input and Precommitment v1", "Fresh Reference Adjudication v1", "Data Reproduction Log v1", "Full Citation Context Audit v1", "Full Claim Registry and Verification v1", "Originality Audit v1", "E6 Claim Strength Audit v1", "Retraction Correction Check v1", "AI Research Failure Mode Checklist v1: 7/7 CLEAR", "Compliance Report Schema 12 v1: WARN", "Figure Table and Experiment Audit v1", "Scope Novelty and Traceability Audit v1", "Token Advisory Register v1", "Final Integrity Verification Report v1: PASS", "Final Verified Course Paper v1", "Stage 4.5 Artifact Index and Mandatory Checkpoint v1"], "started_at": "2026-08-03T16:45:37Z", "completed_at": "2026-08-03T17:28:28Z", "verification_input": "stage-04p-rerevision/03-re-revised-course-paper-for-inspection.md", "verification_input_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f", "final_verified_manuscript": "stage-04.5-final-integrity/14-final-verified-course-paper.md", "final_verified_manuscript_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f", "verdict": "PASS", "blocking_issues": 0, "serious_issues": 0, "medium_issues": 0, "minor_correction_items": 0, "execution_mode": "sequential_same_pi_context_disclosed_transport_recovery", "independent_verifier_claimed": false, "new_literature_search_run": false, "verification_only_live_lookups_run": true, "corpus_changed": false, "cross_model_active": false, "claim_audit_opt_in": false, "references_checked": 21, "references_verified": 21, "citation_contexts_checked": 58, "citation_contexts_verified": 58, "claims_checked": 41, "claims_verified": 41, "data_reproduction": "PASS_UNCHANGED_HASH", "originality_sampling": "24_OF_46_52.17_PERCENT_ALL_REVISED", "temporal_findings": 0, "e4_advisories": 0, "e5_advisories": 0, "e6_advisories": 0, "token_conservation_advisories": {"total": 12, "accepted": 12, "open_nonblocking": 0}, "ai_failure_mode_checklist": "PASS_7_OF_7_CLEAR", "compliance_decision": "WARN_NONBLOCKING_OTHER_EVIDENCE_SYNTHESIS", "compliance_acknowledgement_pending": false, "compliance_warn_acknowledged": true, "compliance_warn_disposition": "preserve_as_reporting_limitation_no_silent_manuscript_expansion", "stage5_entry_confirmed_at": "2026-08-03T17:38:02Z", "citation_style_selected": "APA_7_TRADITIONAL_CHINESE_ADAPTATION", "all_token_advisories_accepted_at": "2026-08-03T17:38:02Z", "stage3p_traceability_sidecar_consumed": true, "stage3p_frozen_new_issues": {"previously_missed": 0, "indeterminate": 0}, "stage5_started": true, "checkpoint_type": "MANDATORY", "checkpoint_confirmed": true},
+    "5": {"name": "FINALIZE", "skill": "academic-paper", "status": "completed_and_scholar_confirmed", "mode": "format-convert", "subphase": "stage5_complete", "outputs": ["Stage 4.5 to Stage 5 Handoff and Format Precommitment v1", "Provenance-finalized Working Draft v1: LOW-WARN", "Human-read Attestation Checklist v1", "Cite-time Hard Gate Record v1: REFUSE", "Cite-time Provenance Gate Report v1: REFUSE", "Human-read Ledger v1: 21 full-text attestations", "Provenance-finalized Working Draft v2: 79/79 ok", "Resolved Human-read Checklist v1", "Cite-time Hard Gate Record v2: PASS", "Clean APA 7 Markdown v1", "APA 7 DOCX v1", "Markdown/DOCX Format Validation v1", "Provenance Summary v1", "Submission Package Verification Report v1: ADVISORY", "Stage 5 Output Package and FULL Completion Checkpoint v1"], "started_at": "2026-08-03T17:38:02Z", "verified_input": "stage-04.5-final-integrity/14-final-verified-course-paper.md", "verified_input_sha256": "513004b245efe8a557e92144faa8c06eb6bd95e16095d35a624d3d7d4870e40f", "citation_style": "APA_7_TRADITIONAL_CHINESE_ADAPTATION", "format_only": true, "source_mutated": false, "formatter_execution_mode": "sequential_same_recovered_pi_context_no_subagent_facility", "independent_formatter_claimed": false, "cite_time_gate": "PASS", "cite_time_gate_reason": null, "anchors": "79_OF_79_SECTION_NON_NONE", "stage4_5_reference_adjudication": "21_OF_21_VERIFIED", "passport_literature_corpus_present": false, "human_read_log_present": true, "human_read_log": "state/material-passport_human_read_log.yaml", "human_read_log_sha256": "6e2b0443ac75513745d0c5e51136809dfbfd15c7f7c652f002b00c58a582ffda", "human_read_scope": "full_text_21_OF_21", "human_read_inferred_from_integrity_report_inspection": false, "citation_marker_resolution": "79_OF_79_OK", "clean_markdown": "stage-05-finalize/02-final-paper-apa7.md", "clean_markdown_sha256": "64ac9f2d5794f1f920136259869fe6a8f4c248f74a5d052f8d349818796ad219", "docx": "stage-05-finalize/03-final-paper-apa7.docx", "docx_sha256": "23740f600c870235efcf8a58b6ea4c3a5b79720f06e91f1d9af24e3639cf728a", "converted_outputs_emitted": true, "latex_question_asked": true, "latex_decision": "DECLINED_BY_SCHOLAR", "paper_pdf_generated": false, "content_confirmation_pending": false, "content_confirmed": true, "content_confirmed_at": "2026-08-03T18:02:56Z", "submission_package_policy": "advisory", "submission_package_verifier": "NO_TERMINAL_TOKEN_7_NOT_CHECKED", "submission_package_report": "stage-05-finalize/package/submission_verification_report.json", "submission_package_report_sha256": "0b665fe42c65bc966b081fc4eb1df7e8cd1b6c28317fde65232fabc71187c815", "submission_verifier_passport_snapshot": "stage-05-finalize/submission-verifier-passport-snapshot.yaml", "submission_verifier_passport_snapshot_sha256": "a59dbfcefc386729d9ea60bf581888098f10682e2bc0e8d15258f9767be5080c", "package_provenance_sha256": "ae39019e659865e244e9a5fa7c3460bd2f87954d8acbb46e3402e0e3dd2e75e9", "completed_at": "2026-08-03T18:02:56Z", "checkpoint_type": "FULL", "checkpoint_confirmed": true, "checkpoint_confirmed_at": "2026-08-03T18:02:56Z", "stage6_authorized": true},
+    "6": {"name": "PROCESS SUMMARY", "skill": "academic-pipeline", "status": "completed", "mode": "auto", "subphase": "terminal_confirmed", "outputs": ["Stage 5 to Stage 6 Handoff v1", "Whole-pipeline Collaboration Depth Trajectory v1", "English Paper Creation Process Record MD v1", "English Paper Creation Process Record LaTeX v1", "English Paper Creation Process Record PDF v1", "Traditional Chinese Paper Creation Process Record MD v1", "Traditional Chinese Paper Creation Process Record LaTeX v1", "Traditional Chinese Paper Creation Process Record PDF v1", "Stage 6 Process Record Validation v1", "Stage 6 Artifact Index and Terminal Checkpoint v1", "Terminal Acknowledgement Record v1"], "started_at": "2026-08-03T18:02:56Z", "deliverables_generated_at": "2026-08-03T18:13:01Z", "language_choice": "BOTH_DEFAULT_ENGLISH_FIRST", "raw_dialogue_sessions": 2, "raw_dialogue_user_turns": 41, "raw_dialogue_assistant_text_turns": 37, "execution_mode": "sequential_same_recovered_pi_context_no_independent_observer", "cross_model_active": false, "collaboration_depth": {"delegation_intensity": 8, "cognitive_vigilance": 9, "cognitive_reallocation": 9, "zone": "Zone_3_Deep", "blocking": false}, "collaboration_quality_score": 91, "ai_self_reflection_sycophancy_screen": "HIGH_MECHANICAL_2_OF_2_DA_CONCESSIONS_NOT_DIAGNOSTIC", "failure_mode_history": "7_OF_7_CLEAR_NO_OVERRIDES", "english_process_markdown": "stage-06-process-summary/paper_creation_process_en.md", "english_process_markdown_sha256": "d378ec4100d64c2041083352af09a4d1b5f0acfb778ec047590d310a861f7ea3", "english_process_pdf": "stage-06-process-summary/paper_creation_process_en.pdf", "english_process_pdf_sha256": "d9516715dfb739fb6d84c9d46b5c99e30dad1f53c4db6bd9e64059e1e4a493cd", "traditional_chinese_process_markdown": "stage-06-process-summary/paper_creation_process.md", "traditional_chinese_process_markdown_sha256": "99c4e49ec6cd1a73dbdb0caaee1abfadf556bc7cd5ee44d00077e8d2925ef4a7", "traditional_chinese_process_pdf": "stage-06-process-summary/paper_creation_process_zh.pdf", "traditional_chinese_process_pdf_sha256": "32262fc11e64001d3aadd383e7d41db9abfa33199b06229e0b486b07eaa17e99", "font_fallback": "Songti_TC_for_Source_Han_Serif_TC_VF_unavailable", "checkpoint_type": "TERMINAL", "checkpoint_confirmed": true, "checkpoint_confirmed_at": "2026-08-03T18:19:42Z", "terminal_acknowledgement_received": true, "terminal_acknowledgement_received_at": "2026-08-03T18:19:42Z", "terminal_acknowledgement_record": "stage-06-process-summary/04-terminal-acknowledgement-record.md", "terminal_acknowledgement_record_sha256": "e2c2d90254e9ebfdaa9e9f6f6ecede11ada3e9e843aa36d75deea700177ac31c", "warnings_and_limitations_accepted": true, "completed_at": "2026-08-03T18:19:42Z"}
+  },
+  "evidence_boundary": {
+    "literature_policy_analysis": true,
+    "secondary_public_data_only_if_retrieved_and_verified": true,
+    "secondary_official_data_descriptive_extraction_completed": true,
+    "interviews_conducted": false,
+    "original_dataset_analyzed": false,
+    "experiments_declared": false,
+    "future_mixed_methods_only": true,
+    "preserve_unverified_markers": true
+  }
+}

package.json

@@ -0,0 +1,24 @@
+{
+  "name": "academic-research-skills",
+  "private": true,
+  "description": "Academic research workflow skills with Claude Code and Pi adapters",
+  "keywords": [
+    "pi-package",
+    "academic-research"
+  ],
+  "license": "CC-BY-NC-4.0",
+  "pi": {
+    "extensions": [
+      "./pi/wrapper.js"
+    ],
+    "skills": [
+      "./deep-research",
+      "./academic-paper",
+      "./academic-paper-reviewer",
+      "./academic-pipeline"
+    ],
+    "prompts": [
+      "./commands"
+    ]
+  }
+}

pi/README.md

@@ -0,0 +1,147 @@
+# Pi wrapper
+
+Thin, community-maintained compatibility wrapper for running the original Claude Code Academic Research Skills (ARS) in [Pi](https://pi.dev/).
+
+The wrapper does not copy or modify ARS skill, agent, reference, schema, script, or command content. Pi loads the four original `SKILL.md` trees and exposes the original `commands/ars-*.md` files as prompt templates. A small input adapter reads the selected original command at invocation time and preserves trailing user arguments.
+
+## Load-bearing runtime differences from Claude Code
+
+These differences affect how ARS evidence should be interpreted:
+
+1. **The wrapper does not provide agent isolation or orchestration.** If the Pi setup has no matching subagent, workflow, or parallel-agent capability, ARS specialist roles run sequentially in the current context. That is degraded execution and must be disclosed; it is not independent multi-agent review.
+2. **Claude hooks do not run in Pi.** In particular, ARS write-scope enforcement remains prompt-level unless the user's Pi environment supplies a separate enforcement mechanism. Do not represent the Claude `PreToolUse` hook boundary as active.
+
+Run `/ars-pi-doctor` to inspect the current environment. The curated [degraded-mode end-to-end evidence](../examples/pi/README.md) shows both boundaries in a completed run.
+
+## Requirements
+
+- Pi with package support (`pi install` / `pi -e`)
+- This complete repository checkout; `pi/` is only a wrapper around its parent files
+- Optional capabilities depend on the user's Pi setup:
+  - a subagent, workflow, or parallel-agent skill/tool for true multi-agent execution
+  - a web-search or page-retrieval skill/tool for literature search and verification
+  - Python, Pandoc, and tectonic for the optional ARS features that already require them
+
+There are no required Pi orchestration or web-search dependencies. When a capability is unavailable, the wrapper tells Pi to use an installed equivalent or disclose degraded execution instead of pretending the operation ran.
+
+## Try without installing
+
+From the repository root:
+
+```bash
+pi -e ./pi
+```
+
+Then inspect optional capabilities and try a mode:
+
+```text
+/ars-pi-doctor
+/ars-plan
+/ars-lit-review AI-assisted systematic reviews
+/ars-reviewer
+/ars-full
+```
+
+The original skills are also directly available:
+
+```text
+/skill:deep-research
+/skill:academic-paper
+/skill:academic-paper-reviewer
+/skill:academic-pipeline
+```
+
+For automatic ARS skill selection from subsequent natural-language prompts, explicitly enable ARS mode first:
+
+```text
+/ars-pi-start
+```
+
+## Install from GitHub
+
+When this branch exists on GitHub, Pi can install the repository directly because the root `package.json` points to this wrapper and the original ARS resources:
+
+```bash
+pi install git:github.com/OWNER/academic-research-skills
+```
+
+Pin a branch, tag, or commit when needed:
+
+```bash
+pi install git:github.com/OWNER/academic-research-skills@REF
+```
+
+## Install from this checkout
+
+Keep the checkout at a stable path, then run from the repository root:
+
+```bash
+pi install .
+```
+
+The nested manifest remains usable for wrapper-only local installation:
+
+```bash
+pi install ./pi
+```
+
+Pi stores a local path rather than copying the repository. Update ARS with normal Git operations; the wrapper continues to load the original files.
+
+If an optional sandbox restricts reads to the current working directory, allow read access to the repository checkout or run Pi from the repository root. Supporting files are intentionally not copied into the wrapper.
+
+Remove using the same source form used during installation:
+
+```bash
+pi remove .
+# or: pi remove ./pi
+```
+
+## Capability doctor
+
+`/ars-pi-doctor` runs without an LLM call. It reports discovered orchestration and web-retrieval capabilities plus Python, PyYAML, Pandoc, tectonic, sandbox, and Claude-hook status. Missing optional dependencies remain the user's choice; the wrapper does not install them.
+
+## System-prompt scope
+
+For prompts submitted while Pi is idle, installing the package does not add ARS text to ordinary Pi prompts. While ARS mode is inactive, the wrapper removes only this package's four ARS skill entries from the newly started run's system prompt, preventing automatic model invocation without modifying the original `SKILL.md` files.
+
+When Pi is idle, an `/ars-*` command or one of the four direct `/skill:*` entries above activates ARS mode before the new agent run starts, so that request receives both the original skill and the compatibility note. `/ars-pi-start` explicitly enables ARS mode and automatic skill selection for subsequent natural-language prompts. The state survives resuming the same session, follows the selected branch during `/tree` navigation, and resets in a new session. `/ars-pi-doctor` does not activate it.
+
+Pi 0.83.0 does not rebuild the system prompt when a prompt is queued into an agent run that is already streaming. A mid-stream `/ars-*` or direct `/skill:*` prompt can therefore execute under that run's existing system prompt without the wrapper-injected compatibility note; similarly, `/ars-pi-stop` changes the persisted mode immediately but cannot remove ARS text from the in-flight run. The RPC `steer` and `follow_up` methods also bypass Pi's `input` event, so a direct ARS `/skill:*` sent through either method does not activate the wrapper. To receive the scoped system-prompt guarantee, wait for Pi to become idle and submit the request through the normal interactive or RPC `prompt` path.
+
+To hide the ARS skills again and continue unrelated work without the compatibility note, run:
+
+```text
+/ars-pi-stop
+```
+
+## Wrapper regression test
+
+```bash
+node --test pi/wrapper.test.mjs
+```
+
+The test covers idle-prompt skill hiding through canonical and symlink-spelled load paths, XML-escaped locations, preservation of missing and adjacent unrelated skills across alternate block formatting, same-request `/ars-*` and direct `/skill:*` activation while idle, manual start/stop toggling, `/tree` state restoration, argument-safe script-path rewriting, and single-pass argument-placeholder substitution.
+
+## What the wrapper translates
+
+The wrapper reads `/ars-*` invocations from the original Claude command files, strips their frontmatter, appends trailing arguments when the command has no argument placeholder, converts executable `python scripts/...` paths to checkout-absolute paths, and expands the original target `SKILL.md` through Pi's native `/skill:*` mechanism. While ARS is active in the session, it also adds a short compatibility note to Pi's system prompt:
+
+- repository-root ARS paths resolve against this checkout
+- Claude tool names mean “use the equivalent available Pi capability”
+- multi-agent work searches available tools and configured Pi skill locations for an installed orchestration capability, otherwise it uses sequential execution with a disclosure
+- `WebSearch`, `WebFetch`, and `/websearch` search available tools and configured Pi skill locations for an installed web capability; no capability means no verification claim
+- Claude-specific command model hints are ignored, so the active Pi model is inherited
+
+The `pi/package.json` manifest performs the remaining mapping directly:
+
+| Claude distribution resource | Pi resource |
+| --- | --- |
+| four original skill directories | four Pi skills |
+| `commands/ars-*.md` | `/ars-*` Pi prompt templates with argument-preserving native skill expansion and absolute utility-script paths |
+| Claude tool/runtime assumptions | short capability-based compatibility note |
+
+## Scope
+
+This is intentionally a basic wrapper, not a reimplementation of Claude Code or ARS orchestration. The original ARS content remains authoritative and unmodified.
+
+The project license remains CC BY-NC 4.0. Attribution and noncommercial restrictions apply.

pi/package.json

@@ -0,0 +1,26 @@
+{
+  "name": "academic-research-skills-pi-wrapper",
+  "private": true,
+  "type": "module",
+  "version": "0.1.0",
+  "description": "Thin Pi wrapper for the upstream Claude Code Academic Research Skills",
+  "keywords": [
+    "pi-package",
+    "academic-research"
+  ],
+  "license": "CC-BY-NC-4.0",
+  "pi": {
+    "extensions": [
+      "./wrapper.js"
+    ],
+    "skills": [
+      "../deep-research",
+      "../academic-paper",
+      "../academic-paper-reviewer",
+      "../academic-pipeline"
+    ],
+    "prompts": [
+      "../commands"
+    ]
+  }
+}

pi/wrapper.js

@@ -0,0 +1,193 @@
+import { existsSync, readFileSync, realpathSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { dirname, resolve } from "node:path";
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
+const commandPattern = /^\/(ars-[a-z0-9-]+)(?:\s+([\s\S]*))?$/;
+const skillNames = ["deep-research", "academic-paper", "academic-paper-reviewer", "academic-pipeline"];
+const skillPattern = new RegExp(`^/skill:(${skillNames.join("|")})(?:\\s|$)`);
+const stateEntryType = "ars-pi-state";
+const orchestrationPattern = /subagent|workflow|parallel[-_ ]agent|multi[-_ ]agent/i;
+const webPattern = /websearch|web[-_ ]search|brave[-_ ]search|webfetch|web[-_ ]fetch|page[-_ ]retrieval/i;
+
+function uniqueMatches(items, pattern) {
+  return [...new Set(items.filter((item) => pattern.test(item.search)).map((item) => item.label))];
+}
+
+const xmlEntities = {
+  amp: "&",
+  lt: "<",
+  gt: ">",
+  quot: '"',
+  apos: "'",
+};
+
+function decodeXml(text) {
+  return text.replace(/&(amp|lt|gt|quot|apos);/g, (_match, entity) => xmlEntities[entity]);
+}
+
+function canonicalPath(path) {
+  try {
+    return realpathSync(path);
+  } catch {
+    return undefined;
+  }
+}
+
+const skillLocations = new Set(
+  skillNames
+    .map((name) => canonicalPath(resolve(repoRoot, name, "SKILL.md")))
+    .filter((location) => location !== undefined),
+);
+
+// Upstream SKILL.md files stay unmodified, so hide their exact Pi listings while ARS is inactive.
+function hideArsSkills(systemPrompt) {
+  return systemPrompt.replace(
+    /(?:\r?\n)?[ \t]*<skill>(?:(?!<skill>)[\s\S])*?<\/skill>/g,
+    (block) => {
+      const location = block.match(/<location>([^<]+)<\/location>/)?.[1];
+      const canonicalLocation = location ? canonicalPath(decodeXml(location)) : undefined;
+      return canonicalLocation && skillLocations.has(canonicalLocation) ? "" : block;
+    },
+  );
+}
+
+async function probe(pi, command, args) {
+  try {
+    const result = await pi.exec(command, args, { timeout: 3000 });
+    if (result.code !== 0) return "missing";
+    return (result.stdout || result.stderr).trim().split("\n")[0] || "available";
+  } catch {
+    return "missing";
+  }
+}
+
+const compatibility = `
+## Academic Research Skills compatibility for Pi
+
+Apply this section only while using an Academic Research Skills (ARS) skill or an \`/ars-*\` prompt.
+
+- The original Claude Code ARS files are the source of truth. Do not replace their research, writing, review, integrity, or checkpoint rules.
+- Resolve repository-relative ARS paths such as \`shared/\`, \`scripts/\`, \`commands/\`, \`deep-research/\`, \`academic-paper/\`, \`academic-paper-reviewer/\`, and \`academic-pipeline/\` from \`${repoRoot}\`.
+- Treat Claude Code tool names as capability names. Use the equivalent tools available in the current Pi session.
+- When ARS requests multiple agents, first inspect the available Pi tools and installed skills for a subagent, workflow, or parallel-agent capability; if none is obvious, search the configured Pi skill locations. Use a matching capability when available. Otherwise run the roles sequentially in the current agent and disclose that degraded execution mode. Do not require a specific orchestration add-on.
+- When ARS requests \`WebSearch\`, \`WebFetch\`, or \`/websearch\`, first inspect the available tools and installed skills for web search or page retrieval; if none is obvious, search the configured Pi skill locations. Prefer an installed web-search skill when available. Otherwise use a permitted HTTP or bibliographic-API capability. If none exists, report the limitation and do not claim that a source was verified.
+- Ask questions through normal conversation or any available user-question tool. Ignore Claude-only command \`model\` metadata and inherit the current Pi model.
+`;
+
+export default function (pi) {
+  let arsActive = false;
+
+  const setArsActive = (active) => {
+    if (arsActive === active) return;
+    arsActive = active;
+    pi.appendEntry(stateEntryType, { active });
+  };
+
+  const restoreArsState = (ctx) => {
+    const state = ctx.sessionManager.getBranch()
+      .filter((entry) => entry.type === "custom" && entry.customType === stateEntryType)
+      .pop();
+    arsActive = state?.data?.active === true;
+  };
+
+  pi.on("session_start", (_event, ctx) => {
+    restoreArsState(ctx);
+  });
+
+  pi.on("session_tree", (_event, ctx) => {
+    restoreArsState(ctx);
+  });
+
+  pi.on("input", (event) => {
+    const match = event.text.match(commandPattern);
+    if (!match) {
+      if (skillPattern.test(event.text)) setArsActive(true);
+      return { action: "continue" };
+    }
+
+    const commandPath = resolve(repoRoot, "commands", `${match[1]}.md`);
+    if (!existsSync(commandPath)) return { action: "continue" };
+
+    setArsActive(true);
+    const raw = readFileSync(commandPath, "utf8");
+    const body = raw.replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n?/, "").trim();
+    const args = match[2]?.trim() ?? "";
+    const hasArgumentPlaceholder = body.includes("$ARGUMENTS") || body.includes("$@");
+    const executableBody = body.replace(
+      /\b(python3?|bash|sh) scripts\/([^\s`]+)/g,
+      (_match, command, script) => `${command} "${repoRoot}/scripts/${script}"`,
+    );
+    const expanded = executableBody.replace(/\$ARGUMENTS|\$@/g, () => args);
+    const commandText = args && !hasArgumentPlaceholder
+      ? `${expanded}\n\nUser request / arguments:\n${args}`
+      : expanded;
+    const skillName = body.match(/^Trigger the `([^`]+)` (?:skill|orchestrator)/m)?.[1];
+    const text = skillName ? `/skill:${skillName} ${commandText}` : commandText;
+
+    return { action: "transform", text };
+  });
+
+  pi.registerCommand("ars-pi-start", {
+    description: "Enable ARS compatibility and automatic skill selection in this session",
+    handler: async (_args, ctx) => {
+      setArsActive(true);
+      ctx.ui.notify("ARS compatibility instructions enabled for this session", "info");
+    },
+  });
+
+  pi.registerCommand("ars-pi-stop", {
+    description: "Disable ARS compatibility and automatic skill selection in this session",
+    handler: async (_args, ctx) => {
+      setArsActive(false);
+      ctx.ui.notify("ARS compatibility instructions disabled for this session", "info");
+    },
+  });
+
+  pi.registerCommand("ars-pi-doctor", {
+    description: "Check optional Pi capabilities and local ARS runtime dependencies",
+    handler: async () => {
+      const activeTools = new Set(pi.getActiveTools());
+      const tools = pi.getAllTools().map((tool) => ({
+        label: `${activeTools.has(tool.name) ? "active tool" : "installed tool"}: ${tool.name}`,
+        search: `${tool.name} ${tool.description}`,
+      }));
+      const commands = pi.getCommands()
+        .filter((command) => !command.name.startsWith("ars-") && !command.name.startsWith("skill:academic-"))
+        .map((command) => ({
+          label: `${command.source}: /${command.name}`,
+          search: `${command.name} ${command.description ?? ""}`,
+        }));
+      const candidates = [...tools, ...commands];
+      const orchestration = uniqueMatches(candidates, orchestrationPattern);
+      const web = uniqueMatches(candidates, webPattern);
+      const python = await probe(pi, "python3", ["--version"]);
+      const pyyaml = await probe(pi, "python3", ["-c", "import yaml; print(f'PyYAML {yaml.__version__}')"]);
+      const pandoc = await probe(pi, "pandoc", ["--version"]);
+      const tectonic = await probe(pi, "tectonic", ["--version"]);
+      const sandbox = pi.getCommands().some((command) => command.name === "sandbox")
+        ? `detected; allow read access to ${repoRoot} when running outside the checkout`
+        : "not detected";
+      const format = (matches) => matches.length > 0 ? matches.join(", ") : "none; sequential/degraded mode";
+      const report = [
+        "ARS Pi doctor",
+        `Repository: ${repoRoot}`,
+        `Orchestration: ${format(orchestration)}`,
+        `Web retrieval: ${format(web)}`,
+        `Python: ${python}`,
+        `PyYAML: ${pyyaml}`,
+        `Pandoc: ${pandoc}`,
+        `Tectonic: ${tectonic}`,
+        `Sandbox: ${sandbox}`,
+        "Claude hooks: unavailable in Pi; write-scope enforcement remains prompt-level",
+      ].join("\n");
+
+      pi.sendMessage({ customType: "ars-pi-doctor", content: report, display: true });
+    },
+  });
+
+  pi.on("before_agent_start", (event) => {
+    if (!arsActive) return { systemPrompt: hideArsSkills(event.systemPrompt) };
+    return { systemPrompt: `${event.systemPrompt}\n${compatibility}` };
+  });
+}

pi/wrapper.test.mjs

@@ -0,0 +1,201 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { mkdtempSync, rmSync, symlinkSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { dirname, join, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+import wrapper from "./wrapper.js";
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
+const compatibilityMarker = "## Academic Research Skills compatibility for Pi";
+const arsSkillNames = ["deep-research", "academic-paper", "academic-paper-reviewer", "academic-pipeline"];
+const arsSkillLocations = arsSkillNames.map((name) => resolve(repoRoot, name, "SKILL.md"));
+const externalSkillLocation = "/example/external/deep-research/SKILL.md";
+const escapeXml = (text) => text
+  .replaceAll("&", "&")
+  .replaceAll("<", "&lt;")
+  .replaceAll(">", "&gt;")
+  .replaceAll('"', "&quot;")
+  .replaceAll("'", "&apos;");
+const skillBlock = (name, description, location) => `  <skill>
+    <name>${name}</name>
+    <description>${description}</description>
+    <location>${location}</location>
+  </skill>`;
+const baseSystemPrompt = [
+  "Base prompt",
+  "",
+  "<available_skills>",
+  ...arsSkillNames.map((name, index) => skillBlock(name, "ARS skill", arsSkillLocations[index])),
+  skillBlock("deep-research", "Unrelated skill with a colliding name", externalSkillLocation),
+  "</available_skills>",
+].join("\n");
+
+function createHarness() {
+  const handlers = new Map();
+  const commands = new Map();
+  const entries = [];
+  const notifications = [];
+  let branch = [];
+  const sessionContext = {
+    sessionManager: { getBranch: () => branch },
+  };
+  const pi = {
+    on(name, handler) {
+      handlers.set(name, handler);
+    },
+    registerCommand(name, command) {
+      commands.set(name, command);
+    },
+    appendEntry(customType, data) {
+      const entry = { type: "custom", customType, data };
+      entries.push(entry);
+      branch.push(entry);
+    },
+  };
+
+  wrapper(pi);
+  handlers.get("session_start")({}, sessionContext);
+
+  return {
+    handlers,
+    commands,
+    entries,
+    notifications,
+    sessionContext,
+    setBranch(nextBranch) {
+      branch = nextBranch;
+    },
+    commandContext: {
+      ui: { notify: (message, level) => notifications.push({ message, level }) },
+    },
+  };
+}
+
+function runBeforeAgentStart(harness, systemPrompt = baseSystemPrompt) {
+  const result = harness.handlers.get("before_agent_start")({ systemPrompt });
+  return result?.systemPrompt ?? systemPrompt;
+}
+
+test("ordinary prompts hide only package ARS skills", () => {
+  const harness = createHarness();
+  const systemPrompt = runBeforeAgentStart(harness);
+
+  for (const location of arsSkillLocations) assert.equal(systemPrompt.includes(location), false);
+  assert.equal(systemPrompt.includes(externalSkillLocation), true);
+  assert.equal(systemPrompt.includes(compatibilityMarker), false);
+});
+
+test("skill hiding does not cross into an adjacent skill block", () => {
+  const systemPrompt = [
+    "Base prompt",
+    "",
+    "<available_skills>",
+    `  <skill>
+    <name>${arsSkillNames[0]}</name>
+    <description>ARS skill</description>
+    <location>${arsSkillLocations[0]}</location></skill>`,
+    skillBlock("deep-research", "Unrelated skill", externalSkillLocation),
+    "</available_skills>",
+  ].join("\n");
+
+  const result = runBeforeAgentStart(createHarness(), systemPrompt);
+
+  assert.equal(result.includes(arsSkillLocations[0]), false);
+  assert.equal(result.includes(externalSkillLocation), true);
+});
+
+test("ordinary prompts hide ARS skills loaded through an XML-escaped symlink path", (t) => {
+  const tempRoot = mkdtempSync(join(tmpdir(), "ars-pi-&-"));
+  t.after(() => rmSync(tempRoot, { recursive: true, force: true }));
+  const linkedRoot = join(tempRoot, "linked-checkout");
+  symlinkSync(repoRoot, linkedRoot, process.platform === "win32" ? "junction" : "dir");
+  const linkedLocations = arsSkillNames.map((name) => resolve(linkedRoot, name, "SKILL.md"));
+  const systemPrompt = [
+    "Base prompt",
+    "",
+    "<available_skills>",
+    ...arsSkillNames.map((name, index) => skillBlock(name, "ARS skill", escapeXml(linkedLocations[index]))),
+    skillBlock("deep-research", "Unrelated missing skill", "/missing/deep-research/SKILL.md"),
+    "</available_skills>",
+  ].join("\n");
+
+  const result = runBeforeAgentStart(createHarness(), systemPrompt);
+
+  for (const location of linkedLocations) assert.equal(result.includes(escapeXml(location)), false);
+  assert.equal(result.includes("/missing/deep-research/SKILL.md"), true);
+});
+
+test("/ars-* activates compatibility for the same agent run", () => {
+  const harness = createHarness();
+  const result = harness.handlers.get("input")({ text: "/ars-plan topic" });
+  const systemPrompt = runBeforeAgentStart(harness);
+
+  assert.equal(result.action, "transform");
+  for (const location of arsSkillLocations) assert.equal(systemPrompt.includes(location), true);
+  assert.equal(systemPrompt.includes(compatibilityMarker), true);
+});
+
+test("direct ARS /skill:* activates compatibility", () => {
+  const harness = createHarness();
+  harness.handlers.get("input")({ text: "/skill:academic-pipeline topic" });
+  const systemPrompt = runBeforeAgentStart(harness);
+
+  assert.equal(systemPrompt.includes(compatibilityMarker), true);
+});
+
+test("command arguments are not rewritten as script paths", () => {
+  const harness = createHarness();
+  const result = harness.handlers.get("input")({
+    text: "/ars-cache-invalidate python scripts/user-provided.py",
+  });
+
+  assert.match(result.text, /python scripts\/user-provided\.py/);
+  assert.equal(result.text.includes(`${repoRoot}/scripts/user-provided.py`), false);
+  assert.equal(result.text.includes(`${repoRoot}/scripts/ars_cache_invalidate.py`), true);
+});
+
+test("argument placeholders are substituted in one pass", () => {
+  const harness = createHarness();
+  const result = harness.handlers.get("input")({
+    text: "/ars-cache-invalidate alpha$@omega",
+  });
+
+  assert.match(result.text, /alpha\$@omega/);
+  assert.doesNotMatch(result.text, /alphaalpha\$@omegaomega/);
+});
+
+test("/tree navigation recomputes activation from the selected branch", () => {
+  const harness = createHarness();
+  harness.handlers.get("input")({ text: "/ars-plan topic" });
+  assert.equal(runBeforeAgentStart(harness).includes(compatibilityMarker), true);
+
+  harness.setBranch([]);
+  const sessionTree = harness.handlers.get("session_tree");
+  assert.ok(sessionTree);
+  sessionTree({}, harness.sessionContext);
+  assert.equal(runBeforeAgentStart(harness).includes(compatibilityMarker), false);
+
+  harness.setBranch([{ type: "custom", customType: "ars-pi-state", data: { active: true } }]);
+  sessionTree({}, harness.sessionContext);
+  assert.equal(runBeforeAgentStart(harness).includes(compatibilityMarker), true);
+});
+
+test("/ars-pi-start and /ars-pi-stop toggle automatic invocation", async () => {
+  const harness = createHarness();
+  const start = harness.commands.get("ars-pi-start");
+  const stop = harness.commands.get("ars-pi-stop");
+
+  assert.ok(start);
+  await start.handler("", harness.commandContext);
+  assert.equal(runBeforeAgentStart(harness).includes(compatibilityMarker), true);
+  for (const location of arsSkillLocations) {
+    assert.equal(runBeforeAgentStart(harness).includes(location), true);
+  }
+
+  await stop.handler("", harness.commandContext);
+  assert.equal(runBeforeAgentStart(harness).includes(compatibilityMarker), false);
+  for (const location of arsSkillLocations) {
+    assert.equal(runBeforeAgentStart(harness).includes(location), false);
+  }
+});