P30: what-made leftover covering + lexical admit for first-person cause lines - #146
Merged
Merged
Conversation
Enjoy/participate paraphrases restated 'what made … easy/motivated' questions. Score leftover lines that add off-query evidence plus a reason token, and let that covering replace a restating hybrid. Co-authored-by: aria <aria@engramhq.com>
Causal leftover lines are often first-person and omit the queried name. AND/ILIKE on "made"/"part"/person floods recency overfetch so the cause line never enters the recall packet for leftover covering. Co-authored-by: aria <aria@engramhq.com>
cursor
Bot
changed the title
Short reason verbs ILIKE-flood and evidence-set covering was stuffing "stay connected" lines into top-k, crowding out first-person cause facts. Admit and covering now use the filtered lexical token set. Co-authored-by: aria <aria@engramhq.com>
Using the dropped lexical set for all queries stripped year from which-year evidence covering and returned an attended-health echo instead of 2021. Co-authored-by: aria <aria@engramhq.com>
Skip-ingest 180 on frozen diag-mh-135 is 114/180 (+1 / -0 vs P28). SH 61→62. Named recovery: first-person running-group cause line (push) on what-made. P29 111/180 dip is not a pin. 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
P30 is an isolated increment on frozen P28 (
113/180).conv-48-q138(What made being part of the running group easy for Deborah to stay motivated?) stores the gold first-person cause line (We help and push each other…) but hybrid restates enjoy/participate. Isolated leftover covering already prefers off-query evidence (push) when that line is in the packet.Live miss was retrieval, not covering: FTS ANDs
made/part/Deborah, ILIKE-OR recency overfetch is ~700 rows (gold at ~360 vs cap 120), and evidence-set covering then stuffed high-DFstaylines into top-30.This PR (1) prefers off-query evidence on what-made leftover covering, (2) drops what-made structure tokens (
made/part), the queried person, and short reason verbs (stay/easy) from lexical search, and (3) runs token-admit and evidence-set covering on that filtered token set so first-person cause lines can enter the packet. Generic English, not a LoCoMo-named product rule. Dest/main stay on P28 until a net-positive skip-ingest 180.Changes
internal/memory/recall.go:looksWhatMadeQuery, leftover covering scores off-query evidence (push/remind) over enjoy/participate restatement; hybrid replace gateleftoverCoveringWhatMadeMissesEvidence.internal/memory/service.go:searchLexicalQueryTokensdropsmade/part, capitalized person tokens, and short reason verbs on what-made queries; admit + covering use the filtered tokens.TestLeftoverCoveringWhatMadePrefersOffQueryEvidence,TestSearchLexicalTokensDropsWhatMadeStructureAndPerson.Test plan
go test ./internal/memory//recallanswers withpush(not enjoy/participate restatement)--run-prefix locomo-s0-diag-mh-135-p30vs P28545e82; pin only if net-positive vs 113/180Linear / tracking
Risk / follow-ups
made/partto global leftover-weak tokens. Do not switch FTS to OR-all-terms.1b5ab3e. No LME-500 claim.