Tokenizer rewritten in closed form (strict thousands grouping, dotted
tokens as a whole) and frozen: quotes, ISO dates, and specific number
forms are the documented candidate set; new prose forms extend the
docstring, not the regexes. Quote pairing no longer skips short pairs
into phantom quotes. Raw links parse only from the fenced-stripped
header. Skip-files applies only to root-level index/log. Body
blockquotes join the quote channel. Fence openers are indent-limited.
Examples now honor their own invariant: two fabricated quotes restored
to the raw wording, the market-data date moved into the raw body, and
the index example rewritten as a real global-index excerpt. A smoke
test assembles the examples and demands zero suspects, so the samples
stay under their own contract. Archive pages get a defined index date
(their Archived field) for lint cross-checks.
- Sentence-final punctuation in the raw no longer blocks matching;
the boundary only rejects continuations that form a different value.
This was a regression from the boundary fix itself and would have
flooded reports with verbatim facts.
- Raw links must resolve inside raw/; anything else is an evidence
error, since the invariant's permanence rests on raw immutability.
- No-material suppression is anchored to ingest entry headings, so
prose mentions elsewhere in the log cannot hide backlog.
- Metadata semantics are scoped to the header block after the H1:
an Archived line in the body no longer grants exemption, and
metadata-shaped lines in a raw file's body stay evidence.
- Status blocks are skipped wholesale, including their explanation
lines, which legitimately cite newer sources outside the article's
own Raw links.
- Fence stripping handles tilde and 3+-backtick fences via a state
machine; candidate tokens are whitespace-normalized before dedup;
index/log args are skipped even when passed explicitly; articles
are checked once, not twice.
- No-material log entries record the full raw path and the script
matches on it, so same-named files elsewhere are no longer wrongly
exempted from the backlog report.
- The invariant drops incremental lint state: the script re-checks the
whole wiki in seconds, which also removes the touched-set
reconstruction problem and the vanishing-unresolved-suspect hole.
- Status block checks move to Judgment Reports; nothing consumes them
mechanically yet, so a mechanical sweep would guard a hypothetical.
- Research is gated to explicit research/ingest requests, resolving
the routing contradiction with read-only Query.
- Mechanical Reports invoke the script via python3 (works without an
exec bit) and name the argument <project-root>.
- Boundary matching: '42K' no longer passes against '142K' in the raw,
closing the substring false-pass on the exact transcription-error
class the checker exists for.
- Coverage: decimals and 4+ digit numbers (incl. bare years) are now
candidates; small plain integers stay out of scope by design and the
boundary is documented in the docstring.
- Fenced code is stripped before extraction (line-by-line processing
made the multiline regex useless), and Status/Archived metadata
lines no longer produce marker-date false suspects.
- Raw files are matched without their collection-metadata header, so a
Collected date cannot false-pass an article's claim about that year.
- Evidence errors are reported instead of silently passing: ordinary
articles without a Raw field, and Raw links that do not resolve.
- Article path arguments resolve against the project root, and missing
files are warnings, not tracebacks.
Every load-bearing fact (number, date, direct quote) must exist
verbatim in the raw files an article links. Compile establishes the
invariant (locate before write, exact literals, derived values show
components); lint verifies it with scripts/check_evidence.py, which
greps candidate literals against linked raws and reports misses plus
unreferenced raw files. Lint is now three tiers: safe auto-fixes,
mechanical reports (the script), judgment reports. Also moves adding
See Also cross-references out of auto-fix, where judgment never
belonged.
Yuhan Lei