[gedaiu]

Part of #1293. The record before code, as you suggested there; no code here.

Thank you for agreeing the shape on #1293 and for the three questions, which made the record a lot sharper. It proposes a per-source {"names": "identity"} on a statements source, settled together with #1248, and writes down the agreed answers. Your three points have their own place:

1. Exact-name attach and the single-candidate split (decision 2). An identity source attaches to the exact-name entity whatever the profile score, so the split never happens for it. What stays deferred is containment for identity names and the split for every other source (under "Not doing").
2. Switching the flag (decision 6). Turning it on deletes the stored vectors of names that have identity evidence, in the same update, because leaving them and skipping them would put an evidence join into every nearest query. Turning it off deletes nothing: those names become pending, and the next extraction in the base backfills them.
3. "Identity evidence wins" (decision 5), stated explicitly, with what it costs prose: a variant of that name no longer finds the entity through its vector. The test the code PR adds for it is named in the record.

0054 gets a dated revision line pointing here, and the index lists 0071 under identity. If you'd like any of the six decisions shaped differently, I'm happy to revise before the code PR.

🤖 Generated with Claude Code

[WaylandYang]

Thank you, this is exactly the short record I hoped for. Decision 2 settles attach vs. split clearly, decision 6 says what happens to existing vectors in both directions, and decision 5 states "identity evidence wins" with its cost and a named test. The link from 0054 and the index row are in place, and 0071 is yours: #1213 has moved to 0072. One small thought for the code PR: when pending learns to skip identity evidence, please keep #1318's plan test green. Merging.