9717a4eeaf88141602820c78de3dce010aed473f
8 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Tim Lingo
|
9717a4eeaf |
measure: claim 24 unflooring is +4 (NOT-SHOWN); asymmetric embedding prefixes are -5 (discarded)
Measured on the 75-query extended gold set (iteration 8's held-out extension) against the certified stack baseline results-stack-ext.json, on the embedded corpus. Three runs of the candidate, zero drift. A. CLAIM 24 WITHOUT THE THRESHOLD - net +4, NOT-SHOWN, kept in the tree. fixed : q14, q25 (in-sample paraphrase), q43, q52, q63, q67 (held-out) broken : q15 (paraphrase), q28 (associative) 8 discordant, McNemar exact p = 0.2891, floor is 6. heldout_paraphrase 16.7% -> 30.0%, paraphrase 61.5% -> 69.2%. Every regression guard held: exact_rare 6/6, phrase 7/7, nonsense 10/10, superseded 2/3. Latency FLAT: p50 641 -> 632 ms. The in-sample half (+q14 +q25 -q15 -q28 = 0) was already on record in iteration 7's cmp-nogate.json, so only the held-out +4 is new. B. ASYMMETRIC TASK PREFIXES ON THE EMBEDDER - net -5, REVERTED in this commit. Rationale was sound and the prediction was wrong, which is why it was worth measuring: nomic-embed-text is an asymmetric retrieval encoder and this file embedded query and document bare on both sides. Prefixing does exactly what the model card implies for the far-away cases - it rescued q42 (gold at GLOBAL COSINE RANK 25,564) and q39 - but it re-ranks the whole space and broke more than it fixed: fixed : q24, q39, q42 broken : q18, q19, q22, q31, q43, q44, q52, q63 heldout_paraphrase 30.0% -> 23.3%, paraphrase 69.2% -> 53.8%. The corpus and the reproducer are kept (embed-corpus-prefixed.py, snapshot-pre-repair-20260806-embedded-prefixed.json) so nobody re-runs it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Tim Lingo
|
9790d9342d |
feat(engram): claim 24 without a threshold, and an embedding substrate that knows query from document
Two changes to the ONE leg that generalises. Iteration 8 measured that out of sample the semantic leg contributes 100% of the stack's gain and the graph leg contributes nothing, so this is where the remaining headroom is. 1. THE 0.60 FLOOR IS A PER-QUERY LOTTERY, AND CLAIM 24 HAS NO THRESHOLD IN IT. 06-claims.md l.148: "respond to embedding search queries by returning the node records whose embedding vectors have the HIGHEST COSINE SIMILARITY to a query vector, independently of the spreading activation traversal." A ranking. ENGRAM_EMBED_SEED_MIN is defined at el_runtime.c l.6094 as the HippoRAG seed-JOIN threshold and l.6102 admits the read-path leg merely "reuses" it. Measured on the 30 held-out paraphrases: the query's own top-1 cosine ranges 0.564-0.680, so the constant keeps a rank-1 answer for one query and discards a rank-1 answer for the next. Six golds sit at global cosine rank 1-2 scoring 0.564-0.589 - discarded by nothing but the constant. What holds the nonsense controls is the corpus-vocabulary gate (nhits == 0), not this floor. Cosine clamped to [0,1] per 05-detailed-description l.69. 2. THE VECTORS THEMSELVES ANSWER THE WRONG QUESTION. EL_EMBED_MODEL defaults to nomic-embed-text, an ASYMMETRIC retrieval encoder trained with task prefixes. Embedding query and document bare - as this file did on both sides - measures topical similarity rather than answer-hood. eg_embed_fetch now takes the task prefix: EL_EMBED_QUERY_PREFIX on the three query call sites, EL_EMBED_DOC_PREFIX on the two backfill sites. Restores no claim, and says so: Will specifies only "computed by an embedding model over the node's content" (l.17), so the model is his and its correct use is ours. It is the substrate under claim 24 - the index is only as good as the vectors in it. Reproducer for the derived corpus: tools/retrieval-eval/embed-corpus-prefixed.py. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Tim Lingo
|
4eb4c9e287 |
feat(engram): word-start match primitive + corpus-vocabulary gate on recall
The retrieval match test is a raw substring scan, so a query token matches anywhere INSIDE a corpus word: "throom" matches "bathroom". Measured over the 38-query gold set on this corpus that is not a rare accident - q28's lexical leg is 36,954 records of which only 13 contain a query token at a word start (99.96% mid-word noise), six further queries carry ~20,500 mid-word-only records each, and the nonsense control q35 returns 7 records ALL of which match only mid-word. istr_contains_wordstart() anchors a token to a word start (preceding char not alphanumeric) while still matching suffixes, so "value" still hits "values". That empties the lexical leg for gibberish, and the nhits==0 corpus-vocabulary gate (iteration 6's mechanism, feat/claim24-unfloored-semantic) then makes the whole query decline rather than let the semantic leg answer it. Measured vs feat/bm25-lexical-leg on the embedded corpus, 2 runs each, 0 queries of run-to-run drift on both sides: net +1 (nonsense:q35), 0 losses, McNemar p=1.0 -> NOT-SHOWN (floor is 6) nonsense clean 2/3 -> 3/3; exact_rare 100%, phrase 100%, paraphrase 61.5%, associative 66.7%, superseded 2/3 all UNCHANGED latency p50 1184 -> 543 ms (0.46x) Iteration 6 called q35 "a DEFECTIVE CONTROL ... cannot be cleaned without breaking the lexical leg". It can: the defect was the match primitive, and cleaning it cost nothing. Also committed: results-wsclaim24.json + cmp-nogate.json, a measured negative for bundling the claim-24 unfloored semantic leg on top (gains q14/q25, breaks q15/q28/q33/q34, net -2) - it independently reproduces iteration 6's q15/q28 losses and shows unflooring REQUIRES the vocabulary gate. Reproducers: legs.py (leg-level replica, reproduces baseline hit@5 exactly on all 38 queries), policy2.py, ceiling.py, wb2.py. |
||
|
Tim Lingo
|
55f9ee3cb0 |
measure: BM25 lexical leg vs semseed baseline - net +2 (q10,q11), NOT-SHOWN
hit@5 68.6% -> 74.3%, phrase 71.4% -> 100%, MRR@10 0.461 -> 0.502, latency p50 0.97x. Zero losses, zero run-to-run drift on both sides. 2 queries moved against a 6-query noise floor: NO MEASURABLE DIFFERENCE by the harness's own test (McNemar exact p=0.50). Unaddressable records in returned slots: 57 -> 0. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Tim Lingo
|
6f3a048f36 |
feat(engram): semantically seed the graph leg (Will's HippoRAG pass, SEED_K=8)
engram_assoc_leg previously took its seeds only from the top-3 LEXICAL hits. For a paraphrase query the lexical hits are noise by construction, so the walk never reached the neighbourhood that holds the answer. This adds the seeding pass Will documents at el_runtime.c l.6082 — "Semantic seeding (HippoRAG pattern, use similarity twice): the query is embedded, the top-K nodes by cosine join the seed set" — using his own ENGRAM_EMBED_SEED_K (8). Similarity is now used twice, coherently: cosine picks where to STAND in the graph, the structural-relation walk decides what is REACHABLE, and cosine orders what was reached (iteration 2's finding, unchanged). The seed list is deliberately NOT floored at ENGRAM_EMBED_SEED_MIN. Measured over all 38 gold queries: true paraphrase targets score cosine 0.46-0.66 and the three nonsense controls' own nearest neighbours score 0.55/0.60/0.62 — the distributions OVERLAP, so no absolute cosine floor separates signal from gibberish. The gate that works is reachability: gibberish's nearest neighbours carry no structural edge, so its graph leg is empty and the controls hold. The raw top-K is selected inside the existing scoring pass, so the cosine is computed exactly once per node: no extra corpus pass, no extra embed round-trip, latency flat (p50 1220 -> 1227 ms, 1.01x). Measured vs the certified baseline feat/hybrid-semantic-recall, embedded corpus, 2 runs each, zero run-to-run drift on both sides: hit@5 51.4% -> 68.6% MRR@10 0.387 -> 0.461 paraphrase 38.5% -> 61.5% associative 0% -> 66.7% exact_rare 100% held, nonsense 2/3 held, superseded 2/3 held phrase 85.7% -> 71.4% (q11, the known rank-5 rotation tax) net +6 queries (7 fixed / 1 broken), McNemar p=0.0703 |
||
|
Tim Lingo
|
059ce02003 |
feat(engram): an associative leg on the recall path (claim 10 typed relations)
The recall route had no way to reach a node that shares no token and no embedding neighbourhood with the query. The design reserves that case for the graph, and nothing on the read path consulted an edge. This adds a third ranked leg beside the lexical and semantic ones: expand the top 3 lexical hits along STRUCTURAL relations only (claim 10 — identity, contains, superseded_by, references, ...), two hops, both directions, pruned at the same 0.02 firing threshold engram_activate uses; order what was reached by query similarity. Merged by strict rotation, never by score blending. Not PR #135. That wired recall wholesale to engram_activate and lost 57 points of phrase accuracy. The failure there was RANK, not reach — a 2-hop associate at strength 0.06 cannot outrank thousands of 1-hop neighbours of strong lexical seeds. Here the lexical leg is untouched and the associative list is empty for most queries, because a node whose only edges are `tagged` and `related` expands to nothing. MEASURED, hybrid-semantic baseline -> this, 38-query gold set, embedded corpus: associative 0.0% -> 66.7% (first non-zero ever recorded on that category) hit@5 51.4% -> 62.9% exact_rare, phrase, paraphrase, nonsense, superseded: all unchanged latency p50 1.01x 4 queries moved, all gains, 0 losses, McNemar p=0.125 deterministic: two runs of the same binary differ on 0 of 38 rows VERDICT: NOT-SHOWN. The harness needs 6 queries to clear p<0.05 and the whole associative category is only 6 queries, so even 4/6 fixed cannot reach the floor. The mechanism is confirmed to work; the gold set cannot certify it. |
||
|
Neuron
|
635453b936 |
feat(engram): rank-interleave the semantic leg into recall; embed the corpus
Replaces the score-fusion first cut with rank fusion, which is what the data called for. nomic's cosine scale is compressed (true matches 0.55-0.70, unrelated pairs 0.35-0.50), so an additive blend of cosine onto token-coverage is dominated by whichever leg has the wider spread. Alternation is invariant to both scales: L1, S1, L2, S2, ... deduped, capped at limit Lexical ranking is left byte-identical; the semantic ranking is computed beside it and admitted only above ENGRAM_EMBED_SEED_MIN (0.60) — Will's existing seed floor, no new tuning constant. That floor is what keeps the nonsense controls clean: a query with no real match must not be answered with its neighbours. embed-corpus.py / merge-corpus.py produce the derived corpus the semantic leg needs (76,986 vectors, nomic-embed-text, 0 failures, 11 min). Zero of 78,791 nodes carried an embedding before this; the field round-tripped through the snapshot but nothing ever wrote it. MEASURED, 38-query gold set, paired against the SAME derived corpus so the comparison isolates the code change: hit@5 34.3% -> 51.4% paraphrase 0.0% -> 38.5% MRR@10 0.294 -> 0.387 superseded 1/3 -> 2/3 outranks recall@10 33.3% -> 50.5% latency p50 1146 -> 1220ms (1.06x) exact_rare 100% -> 100% phrase 85.7% -> 85.7% nonsense 2/3 -> 2/3 6 queries fixed, 0 broken, McNemar exact p=0.0312, 0 drift across repeats. Regression guards all held. Contrast PR #135, which swapped the read path to spreading activation wholesale: phrase 85.7 -> 28.6, latency 2.81x. Correct mechanism, wrong substrate. The substrate is now present. Restores engram claim 24 (previously 0% honoured). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Neuron
|
cf41d12d22 |
test(retrieval): a measurement harness for memory recall, and its first verdict
Nothing else on the memory roadmap should be built until a change can be shown
to help. Right now we judge by feel, and the benchmark literature is full of
systems that felt better and measured worse. This is the missing gate.
WHAT IT MEASURES, AND WHY IT BOOTS A REAL SOUL
The subject is Will's designed retrieval — spreading activation over the
weighted directed graph, four-factor multiplicative scoring — not a proxy for
it. A Python re-implementation would measure my reading of the design, so the
harness compiles the actual soul.el amalgam from a git ref and asks it over
HTTP on /api/neuron/recall, exactly as the MCP wrapper and the app do.
BUILT ON WHAT WAS ALREADY HERE, NOT AROUND IT
docs/research/graphrag_eval/{collect,score}.py — per-query relevant-id
scoring and fixed-denominator precision@5 (kept verbatim: an empty result
should be punished like a page of junk).
docs/research-archive/p0-prototypes/eval_pinned_40q_20260715.py — the pinned
ground truth + --check winnability gate, so every run judges alike.
scripts/verify-soul-contract.sh — the isolation recipe, including the
non-obvious SOUL_ISE_URL pin without which an "isolated" soul silently
syncs the operator's live brain.
gen-soul-amalgam.sh + .gitea/workflows/ci.yaml — the build recipe and flags.
New here: ids rather than regexes as ground truth, an associative category
derived from real edges, a superseded category scored on ranking, a
machine-checked zero-lexical-overlap guarantee on paraphrases, paired
significance testing, and measurement of the real compiled soul rather than an
offline replica of one leg of it.
THE GOLD SET IS AUDITABLE, NOT VIBES
38 queries over the real 78,768-node corpus, each carrying a `derivation`
string, each re-validated by `build_gold_set.py --check`. exact_rare is mined
(document frequency 1). phrase is mined (verbatim scan; >25 matches rejected as
too diffuse). paraphrase is hand-selected then PROVEN to share zero content
words with its target — a leak fails the build, so the category cannot decay
into lexical matching. associative is derived from real hub edges with
lexically-reachable siblings dropped. nonsense is verified absent. superseded
pairs are kept only when both sides survive as distinct nodes.
HONEST ABOUT NOISE
Minimum detectable swing on 38 queries is 6: if every changed query moves the
same way, p = 2*0.5^n first clears 0.05 at n=6. Run-to-run drift is measured,
not assumed — activation is a stateful read, and it shows: main is fully
deterministic across 3 runs, the candidate drifts by 1 query. compare.py
reports "no measurable difference" for anything inside max(6, drift+1).
FIRST VERDICT — feat/recall-through-activation
hit@5 34.3% -> 22.9%, phrase 85.7% -> 28.6%, latency p50 2.81x. Five discordant
pairs, all five against the candidate, none for it; McNemar exact p = 0.0625,
so by the stated rule this is one query short of significant and is reported as
such rather than as a win for main. The latency regression is deterministic and
not in any noise band.
The benefit the branch was written for is absent: associative recall is 0/6 on
BOTH builds. Probed directly, the traversal returns the lexical seed at rank 8
and none of its 12 hub siblings. Two measured corpus facts explain it — only
4,060 of 78,768 nodes (5.2%) carry any edge, and no node has an embedding, so
the fourth factor of the four-factor product has nothing to compute from. The
mechanism runs; the corpus lacks the structure it needs.
SAFETY
Throwaway port, throwaway HOME, disposable per-run copy of the corpus; live
ports refused by name. Every soul started is killed AND confirmed dead by pid
probe, with the confirmation written into the results file; run_comparison.sh
sweeps for strays and exits non-zero if any survive. Nothing under ~/.neuron,
/Applications/Neuron*, or ~/neuron-dev-stack is read, written, or restarted.
Rung: E2E-VERIFIED — 6 full runs (3 per config) against the real compiled
binaries on the real corpus; numbers above are measured, not projected.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|