The spare-key note scored 0.832 to 0.867 against a spare passport, a blue
shirt, a blue document box and a car key. Score and margin cannot separate
those: the right note runs 0.817 to 0.892 and the silent cases 0.787 to
0.874, so the ranges overlap and structure has to decide.
RecallAllowed now takes two structural facts from the router. A locative
question must corroborate every identity term against the candidate's
subject, read up to its first dictionary-proven verb, so a location object
in the note cannot answer for the thing being located. A turn that is not
question-shaped needs a named shared topic even when it ends in '?', which
is what "я отменил напоминание про молоко" lacked when it recalled an
unrelated note at 0.825 with no runner-up to fail the margin.
query_min_score moves 0.55 to 0.80 for tokenizer rev 2. The held-out
fixture answers 14/27 real recalls and 0/14 false ones.
LocativeAnswerVerifier is the resident-model second opinion, kept behind
the deterministic gate and wired into nothing. The measurement that says
why is docs/evals/2026-08-15-locative-answerability-verifier.md.
--no-verify: master is the working branch this session by the owner's call.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every case is scored over its own notes plus the whole filler set, and the
two stores disagree about a repeated id: sqlite upserts on it, the in-memory
store appends. So one collision makes a case score differently on the two
backends, and it reads as an embedder or gate difference — the one thing this
harness exists to measure. It was dodged by hand during #373 by renaming two
ids.
The check sits in Load rather than in TestLoadFixture, so it covers every
caller of the fixture and not only the one that remembers to look.
queryMemory and queryNotes both gate on score alone, so both needed it. The
eval keeps its own copy of bestRecall — package main is not importable — and a
fixture that measures a weaker gate than the daemon runs flatters it, so the copy
moves in step and its test pins the new rule.
Measured on the held-out recall fixture with the real embedder: 17/32 cases pass
→ 22/32, false recall 1/5 → 0/5, answered after gate 18/27 → 17/27. The one true
recall lost is en-hard-024, an English question against a Russian note, where no
lexical test can help.
The memory pass ran only after the notes-only gate had already rejected
the same note at the same score. Notes and facts share one vector index,
so a note that failed there failed again — the branch could only ever
return a fact.
Now the memory pass runs first: one search over everything Maven
remembers, one gate, and the memory that clearly matches best answers
(a note gets phrased, a fact is read back). The notes-only pass stays
behind it for notes the vector index does not hold. No threshold moved,
so the set of questions answered is unchanged — only which memory
answers them.
Fixture gained two mixed note+fact cases, so the answerable count goes
25 -> 27: hash recall@1 36.0% -> 37.0% (ratchet 0.32 unchanged, comment
updated), e5 recall@1 72.0% -> 70.4%, false recall still 1/5.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The e5 embedder puts every cosine in one narrow band (0.79-0.89), so the
absolute query_min_score gate cannot tell a real hit from a made-up
question: any value under the band answers everything, any value above it
answers nothing. False recall was 5/5.
New gate asks whether one note is clearly the best instead: top1 - top2 >
delta. New query_min_margin config knob, default 0.008, read off the sweep
in the recall harness. The absolute floor stays as a second check.
On the recall fixture with e5: answered 72% -> 68%, false recall 5/5 -> 1/5.
Vikunja #359
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The old model was a symmetric paraphrase model, so it scored "do these
look alike" instead of "does this note answer this question". Also fixes
the file mismatch: the Makefile, the deploy config and both evals now all
name the same quantized file, and the quantized one is what gets measured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Real recall is 48% after the gate, and one must-be-silent query gets an
answer anyway. Review finding 2 (the score distributions overlap, so no
gate separates a real recall from a false one) and finding 4 (the memStore
branch at voice.go:776 is unreachable for notes). Adds an embedder cache
so the gate sweep does not re-embed the fixture nine times.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Measures whether Maven can find the right note again from a paraphrased
question. Review internal/memory/recalleval/recalleval.go's Score for how
rank, gate and false recall are kept as three separate numbers, and the
fixture's filler list for why recall@3 is not free.
Fixture JSON is generated data and does not count toward the diff limit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
claude
kami