docs: bind published counts to the command that re-derives them

[ManSio]

What

Four published numbers had drifted from the code they describe:

number	stated	live (measured on this tree)
README test badge	1965 passed	2007 collected / 2001 passed
README architecture row	~1889 tests	2007 collected
WISDOM intel census	intel_*=14	20
WISDOM test census	tests=1180	2007 collected

This is a named class, not sloppiness: benchmark decay — a public static
number whose meaning erodes while the number stays put.

Why a guard and not just an edit

A manual edit fixes it once. Each claim now carries the command that re-derives
it, and disagreement exits rc=3, so the next drift is caught by
un_all.py
instead of by a reader noticing.

The comparison is symmetric. The first cut used live <= stated, which only
catches understatement; its own selftest caught an inflated badge passing through.
Two of the four synthetic cases exist to be rejected, one per direction.

History is superseded, not rewritten

The dated 2026-08-12 snapshot (intel_*=14, tests=1180) was correct when
written. It is now labelled as a snapshot and left intact; the live census is a
separate section. Published history gets a new record that supersedes it, not a
silent numeric edit.

Also

run_all.py reported the protocol guards' false-positive share as UNMEASURED
after it had been measured. Re-derived from raw: 8 reported, 5 false positives
= 62.5%, 3 actionable. Quoting the raw count overstated defects by 2.7x.

Verified on this tree

• pytest tests/ → 2001 passed, 6 skipped, exit 0
• tools/verification/run_all.py → 14/14 provability steps OK, exit 0
• verify_public_claims.py --selftest → 4/4, including 2 that must be rejected
• verify_public_claims.py → 4/4 published numbers reproduce
• pre-commit: 9/9 hooks OK

The guard was confirmed failing on the known-bad state (4/4 WRONG) before the fix
and passing after — held-out on the original defect, not generalised.

Sources

• GTM-Bench, "Keeping a Benchmark Honest" (2026-09-06) — benchmark decay;
  "every reported score names the version it was run against; a number without a
  version is uninterpretable".
• READU (arXiv 2607.15780) — an alert judge to remove false positives is part
  of the construction, not an optional extra.
• driftmd — "badge versions" as its own named check.

Not in this PR

• G5 denominator n_reviewed is still 0. The tri-state claim classification
  (CLAIM / NOT_A_CLAIM / DERIVED with machine-checked reason codes) is not done
  and is not claimed here.
• 7 tracked files outside the personal-path guard's declared scope still contain a
  local username (.local/*.py, docs/archive/*.md,
  scripts/reconstruct_judge_cot.py). The guard's boundary is declared in its
  docstring, so this is unreported territory rather than a hole in the guard —
  left for a separate decision, and archives must not be rewritten.

🤖 Generated with opencode

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 480abcc2-0320-4e70-b81b-0c1fd1906754

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}