The agent judges identity by written rules, and a labeled set says whether it did (0025 decision 10)

[WaylandYang]

Follows #446. Adds decision 10 to record 0025 (Governance reads the ledger before it decides); the status line, the new decision with its table, and the README row are updated.

The labeled set and the harness

• scripts/bench/truth/ai-timeline.duplicates.json: every pair the resolver raised on one governed import of eight interlocking articles (615 pairs, 411 distinct name pairs), labeled by hand from each side's facts and source passages: same, different, or unknown where no reader could tell either. Keyed by the two names, either order, case-insensitive.
• scripts/bench/govern.mjs: --fresh imports and scores; --kb <id> resets a base's agent decisions without going through the API (the API would leave merge.revert in the ledger, which the next run would read as precedent), runs the governor, and scores: decided on its own, agreed with the labels, wrong merges, wrong keeps, left for people, moot after a merge. About twenty-five minutes per run on the bench base. Documented in scripts/bench/README.md.

What the labels changed

• IDENTITY_RULES, one text read by both prompts: a version, tier, division, subsidiary or legal entity added at the end of a name is a different thing; a phrase containing a name is not the name; a list is not its members; a common noun is not a proper name; a front qualifier dropped, a corporate suffix, a surname for a full name, a parenthetical acronym are the same thing; a different type label, a missing fact, a different partner or role is not a contradiction.
• Two rules the model cannot argue with. name_shape (identical, abbreviation, suffix, version, extension, phrase, unrelated): a same verdict on a version or a phrase is never applied on its own. type_family folds the extractor's labels into person, organization, place and event; only a difference across families is a conflict, so Organization against Corporation no longer blocks a merge and Person against Organization still does.
• Identical names whose labels fall in one family are shown to the model with one label; a different verdict on identical, compatible names gets the tool loop's second look before it is applied.

Measured on the same 589 pairs, reset and re-run after each change

run	decided on its own	agreed with the labels	wrong merges	wrong keeps	left for people	wall clock · loop calls
before (gate of 2026-09-06)	67%	91%	1	35	186	alongside the 63 min import · 301
rules, shapes and families	86%	95.3%	6	18	14	26 min · 42
acronyms, surnames, two-word tails	98.5%	96.3%	6	12	8	22 min · 12
one label for identical names, second look on a split	98.2%	97.3%	4	9	9	22 min · 15
one label only within one family (shipped)	97.9%	96.9%	4	11	11	25 min · 46

Wall clock is for 589 pairs with the model capped at three concurrent calls; about twenty-five pairs a minute, one batch call per cluster of up to twelve pairs (roughly sixty per run) plus the loop calls shown.

The last two differ by the model's run-to-run noise; the shipped version keeps the extractor's labels wherever they carry information, because the run before it merged a periodical into a service. Of the four wrong merges, two follow a precedent left by a simulated person accepting proposals blindly, which is the ledger working as designed on a bad decision. What still disagrees is mostly judgment (whether "Claude Mythos 5" is "Claude Mythos"), and the share left for people does not go to zero on purpose: the version and phrase rules hold back a few right merges to stop dozens of wrong ones, and a record with no facts cannot be settled by anyone.

Verified

Unit tests for the shapes, the families, and the gate with them; the prompt-shape tests; the governor's real-database test. Four full runs on the bench base as above.

🤖 Generated with Claude Code