2 Commits

Author	SHA1	Message	Date
Edward Cheng-I Wu	eb9e7bed17	feat(#127): v3.10 triangulation policy layer (PR-B) (#288) ... Ships the contamination-triangulation policy layer deferred by v3.9.0 (#102): opt-in strict modes that promote the advisory k=3 triangulation signal to a non-acknowledgeable terminal HIGH-BLOCK at the citation-emission boundary. Built on PR-A's canonical firm-rules + sync-lint base. Default behavior (absent / all-advisory terminal_policies) stays byte-equivalent to v3.9.0. Schema: - New standalone shared/contracts/passport/terminal_policies.schema.json (passport-level, NOT entry-level). contamination_triangulation enum {advisory, strict, strict_articles_only}; temporal_integrity accepts only advisory (forward-reserved namespace). - literature_corpus_entry.schema.json gains venue_type (closed enum incl. explicit unknown), venue_type_provenance (no _inferred values), venue_type_source (required iff trusted_source_declared). Bidirectional type/provenance pair dependency + one-way unknown-type rule. - check_literature_corpus_schema.py validates a passport-level terminal_policies block before iterating entries. Finalizer / formatter (agent prompts): - pipeline_orchestrator_agent.md: sole policy evaluator. Stamps a fully-encoded policy_hash slug ONLY under a non-advisory passport; co-emits a TERMINAL-BLOCK token under strict (k=3) alongside the advisory suffix. strict_articles_only is a precision mode (DOI + journal/conference venue + declared provenance); DOI-less / unknown-venue stays advisory by design. - formatter_agent.md: generic rule-11 HIGH-BLOCK refusal + STAMP-ONLY two-gate freshness check; never re-evaluates policy logic. Firm rule / migration / adapters / lint: - R-L3-2-A reworded to the broad default-advisory + opt-in-strict form (canonical single-source); contradiction guard added to check_firm_rules_sync.py. - migrate_literature_corpus_to_v3_10.py deep-merge-seeds terminal_policies (idempotent, dry-run, no venue backfill). - folder_scan / zotero / obsidian adapters declare venue_type. - New check_v3_10_policy.py (alongside the v3.9.0 lint) + CI wiring in spec-consistency.yml and _ci_pytest_manifest.toml. Docs: CHANGELOG Unreleased v3.10 entry; CLAUDE.md v3.10 Key Additions; v3.9.0 / v3.9.4 spec superseded-framing notes. Tests: 1857 passed / 3 skipped (+67 vs PR-A baseline of 1790, 0 regression). Spec: docs/design/2026-05-31-ars-v3.10-policy-layer-rescope-spec.md Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-05-31 23:20:57 +08:00
Edward Cheng-I Wu	8261cfee4b	refactor(#127): disambiguate R-L3-2/R-CIM rule IDs + canonical firm-rules sync lint (#285) ... PR-A of the v3.10 triangulation-policy-layer split: a zero-behavior-change base that removes the R-L3-2-A/B/C ID overload between the contamination advisory rules (original holder, v3.7.3 §3.2) and the claim-intent-manifest emission rules (borrowed the same ID in three writer-stage prompts). - Rename the claim-manifest rules R-L3-2-A/B/C -> R-CIM-A/B/C in synthesis_agent, report_compiler_agent, draft_writer_agent, and the claim_intent_manifest schema description. Wording byte-identical except the ID token. - Add shared/references/firm_rules.md as the canonical single source for the contamination R-L3-2-A..E blocks and the claim-manifest R-CIM-A..C blocks. The note for the contamination mirrors states plainly that their wording is NOT sync-checked yet (prose references, not full-block copies) — only their IDs are guarded; full wording-sync lands with the PR-B reword. - Add scripts/check_firm_rules_sync.py: sync-checks the R-CIM mirrors against the canonical blocks (modulo the agent self-reference noun) and guards the disambiguation. The collision guard seals the full R-L3-2-A..E namespace out of claim-manifest surfaces, and _extract_section is fence-aware so a forbidden ID hidden after a fake heading inside a fenced example cannot escape the scan. Wired into CI via spec-consistency.yml + the pytest manifest. 19 tests. - Spec: docs/design/2026-05-31-ars-v3.10-policy-layer-rescope-spec.md. Part of #127 (policy layer lands in PR-B on this base). No runtime behavior change; full pytest unchanged at 1790 passed / 3 skipped. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-05-31 16:30:16 +08:00