[tenphi]

Summary

The automatic-retention context witness was rejecting two kinds of legitimate references as context_witness_invalid. Each rejection held the candidate as discourse_uncertain. While measuring #241, these were the largest single loss channel.

1. Speakerless authors. Assistant turns, and user turns without a speaker, have no author label. The verifier's first-person reference (source_name: "assistant", explicit, own frame) could therefore never be grounded.
2. Earlier-turn antecedents. The pattern is "I, Ada Marlow, declined the optional Zephyr QX-100 cable-protection extension." in one turn, then "Had I accepted that extension …" in the next. The rule from fix: retain source context through immutable antecedent witnesses #226 that name origins come only from the same item rejected every such binding.

What changes:

• retentionItemAuthor gives an item its speaker, or its assistant/user role when it has no speaker. The existing own-frame first-person rule then applies unchanged.
• retentionContextWitness takes the source's item order. An unambiguous_antecedent may now anchor on an exact name in an earlier item with the identical author label.
• The verifier contract describes both. RETAIN_VERIFIER_VERSION advances to v54, as fix: retain source context through immutable antecedent witnesses #226 did when it changed this contract.

Still rejected, with a test for each:

• another author's turn;
• a later turn;
• unknown source order;
• items without an author label;
• explicit across frames;
• a role label used by another frame;
• a preserved role or name that is missing from the record.

Every reference keeps its semantic-role, preservation and completeness checks. The inner-reporter attribution audit keeps its same-item rule.

Measurement: an exact replay with identical model responses

I replayed all 224 live gpt-6-luna retention runs recorded for #241 through runRetain. A model stub returned each recorded response, so every local check ran fresh under the old and the new grounding. 221 of 224 runs replayed completely; 3 revision calls had no recorded counterpart and were treated as unavailable.

	Old grounding	New grounding
Written memories	216	251 (+35)
context_witness_invalid holds	87	34
Previously written memories lost	—	0
Shipped config (single extraction + #241 revision): required recall	56.1%	62.8%
Shipped config: written / held	105 / 93	120 / 78
Shipped config: judged faithful precision	90.5%	88.3%

I hand-audited the 35 newly admitted records against their sources, and so did a blinded gpt-6-luna judge.

• All 35 have supported, correctly qualified text, and none shows an identity misbinding.
• The judge flagged 7 for labels that come from extraction:
  • 5 are assistant statements where the judge prefers self_attested or a filled source_speaker;
  • 1 counterfactual polarity;
  • 1 rejected offer labelled a decision rather than a plan.

What the 34 remaining holds are:

• the verifier calling the author of a plain-text note "user", which has no structure to ground;
• 5 antecedents across different authors, left out on purpose;
• name anchors on frames that do not contain the name, plus dates or common nouns listed as names. These are verifier mistakes.

Notes for the release

• RETAIN_VERIFIER_VERSION v54 invalidates timeline-source extraction proofs, so timeline sources re-verify on upgrade. This matches earlier contract changes.
• The cross-author case is the obvious next extension. The assistant mentions a product and Ada says "it" in the following turn; that was 5 cases here. It needs a deliberate decision about fix: retain source context through immutable antecedent witnesses #226's caution and is not included.

Testing

• pnpm check:repository, build, lint, knip, smoke and docs:doctor pass.
• pnpm test passes: 5024 passed in the full run. Three ingest.test.ts cases timed out at 30 s under full-suite load, and that file passes 70/70 on its own.
• format:check flags only packages/core/src/ops/context.ts, which is unchanged from main.

🤖 Generated with Claude Code