fix(recall): cover they-were observation leftover on how-react - #156
Merged
Merged
Conversation
How-react leftover covering admits observational leftover of the form "they were/looked/seemed (so) ADJ" when another packet line covers the object tokens. Dislike/hate restatements do not cover. Session expand admits the observation leftover from the same session as the FTS seed. Do not add a dog/snow dictionary. Do not steal dancers/NYC/injury. Co-authored-by: aria <aria@engramhq.com>
They-were/looked leftover only covers how-react when the line itself covers an object token (snow via snowy, dogs). Session-neighbor "they looked so cute" leftover does not cover. Rank floor/boost and episode keep use the same gate. Co-authored-by: aria <aria@engramhq.com>
How-react queries match the generic list cue (content token "dogs") and NeedsMultiHop, so selectEvidenceSetCovering can fill top-k with restatement rows that already cover object tokens and drop they-were leftover. Re-insert object-linked observation leftovers into the capped results. Co-authored-by: aria <aria@engramhq.com>
List-corpus candidates span many sessions. Map iteration spent the 8-session fetch budget before the leftover session, so they-were observation was admitted only some of the time. Seed from ranked FTS hits instead. Co-authored-by: aria <aria@engramhq.com>
Skip-ingest product /recall on diag-mh-135 is 124/180 (net +1 vs P39). Unique gain is conv-44-q107 dogs/snow they-were confused. Do not replace README 11.4% / 70% 1x30. Co-authored-by: aria <aria@engramhq.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
P39 pinned product
/recallat 123/180. The isolated ranking miss with gold in store wasconv-44-q107(How do Audrey's dogs react to snow?, goldConfused). Live P39/recallreturned the compiler restatementAudrey's dogs dislike snow.while the observational leftoverthey were confusedsat in the same session.This PR extends leftover covering with a generic English how-react path: observational leftover of the form
they were/looked/seemed (so) ADJbeats dislike/hate restatement when the observation line itself covers a query object token (e.g.snowycoverssnow). Session expand seeds from ranked FTS hits (not the candidate map) so the leftover session is not dropped from the 8-session budget. How-react questions also trip the generic list cue; the evidence-set cap re-inserts object-linked observation leftovers.Skip-ingest 180 on frozen
diag-mh-135is 124/180 (0.689) — net +1 vs P39, unique gainconv-44-q107, 0 losses. Pin: locomo-s0-diag-mh-135-p40-20260825.md. Product SHA72f7d97; pin docs SHA on this branch. Goal remains 162/180. Do not merge #133. Do not replace README 11.4% / 70% 1×30.Changes
looksHowReactQuery—how do/does/did+react*orrespond*followed byto. Does not matchhow should I respond.leftoverCoveringReactionObservationLine— last wordsthey were/looked/seemed (so) ADJ; ADJ lowercase alphabetic len≥4, not stopword/weak/month; 6–24 words; not they-evaluative / chat / interrogative / caption.react*/respond*only on how-react when tokens remain; floor 0.9; rank boost 0.75; keep leftover episodes; FTS-seeded session expand + evidence-set cap keep for object-linked observation.leftoverCoverRareTokensminLen 4 on how-react so object tokens likesnow/dogsare kept.docs/benchmarks/README.md, dated pin, summary JSON, failure ledger, cycle-closeout, handover unique Next = dog-owners group (conv-44-q117).Test plan
gofmton touched filesunset BRAINY_DATABASE_URL && go vet ./internal/memory ./cmd/apiunset BRAINY_DATABASE_URL && go test ./internal/memory(pass)/recallholds after restart: dogs/snowconfused12/12; Tim injury doctor-said; NYCIt's got; dancers graceful; Joanna writing; turtle-care; electronic; dinner; Gina advice; veterans; push; smartwatch / P28 22/22--run-prefix locomo-s0-diag-mh-135-p40vs 123/180: 124/180, +1/−0, uniqueconv-44-q107Linear / tracking
Risk / follow-ups
conv-44-q117joined a dog owners group.reactglobally. Do not dropturtles/care. Do not dropsay/motivate/spread/advice. Do not revive P29./recallis still 11.4% on SHA1b5ab3e.