Files
neuron/tools/retrieval-eval/legs.py
T
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  associative 66.7%, superseded 2/3 all UNCHANGED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.
2026-08-07 16:58:34 -05:00

118 lines
4.5 KiB
Python

import json,pickle,os,math,urllib.request,numpy as np
S='/private/tmp/claude-501/-Users-timlingo/82369039-a20e-4b5a-8a5e-28234a57b996/scratchpad/sim/'
C=pickle.load(open(S+'corpus.pkl','rb'))
NODES=C['nodes']; EDGES=C['edges']; N=len(NODES)
E=np.load(S+'emb.npy'); HAVE=np.load(S+'have.npy')
En=E/np.maximum(np.linalg.norm(E,axis=1,keepdims=True),1e-12)
LAYERS={int(l['layer_id']):l for l in (C['layers'] or [])} if C['layers'] else {}
TRANS=set(i for i,l in LAYERS.items() if l.get('transparent'))
def addressable(s):
if not s: return False
return all(0x20<=ord(ch)<=0x7e for ch in s)
ADDR=np.array([addressable(n['id']) for n in NODES])
OK=np.array([ (n['layer_id'] not in TRANS) and ADDR[i] for i,n in enumerate(NODES)])
SAL=np.array([n['salience'] for n in NODES])
LOW=[ (n['content']+'\x00'+n['label']+'\x00'+n['tags']).lower() for n in NODES]
DL=np.array([float(len(n['content'])+len(n['label'])+len(n['tags'])) for n in NODES])
IDX={}
for i,n in enumerate(NODES):
IDX.setdefault(n['id'],i)
STRUCT={"identity","contains","superseded_by","references","embodies","demonstrated_by","canonical-self","depends_on","currently_holds","activates"}
ADJ_F=[[] for _ in range(N)]; ADJ_T=[[] for _ in range(N)]
for e in EDGES:
a=IDX.get(e['from']); b=IDX.get(e['to'])
if a is None or b is None: continue
ADJ_F[a].append((e,b)); ADJ_T[b].append((e,a))
EXCL=np.array([n['node_type'] in ('Tag','InternalStateEvent') for n in NODES])
avgdl_all=None
def tokenize(q):
out=[]
for t in q.split():
if not any(t.lower()==x.lower() for x in out): out.append(t)
return out
_qcache={}
def qemb(q):
if q in _qcache: return _qcache[q]
body=json.dumps({"model":"nomic-embed-text","prompt":q}).encode()
r=urllib.request.urlopen(urllib.request.Request("http://127.0.0.1:11434/api/embeddings",data=body,headers={"Content-Type":"application/json"}),timeout=30)
v=np.array(json.loads(r.read())["embedding"],dtype=np.float32)
v=v/np.linalg.norm(v); _qcache[q]=v; return v
K1,B=1.2,0.75
SEED_MIN=0.60; SEED_K=8; ASSOC_SEEDS=3; DEPTH=2; FIRE=0.02; AMAX=64; DECAY=0.7
def legs(query):
toks=tokenize(query)
masks=[];
hit_idx=[]; hit_mask=[]
df=[0]*len(toks)
lt=[t.lower() for t in toks]
for i in range(N):
if not OK[i]: continue
s=LOW[i]; m=0
for t,tok in enumerate(lt):
if tok in s: m|=(1<<t)
if m:
hit_idx.append(i); hit_mask.append(m)
for t in range(len(toks)):
if m>>t&1: df[t]+=1
dl_n=int(OK.sum()); avgdl=float(DL[OK].sum()/max(dl_n,1))
idf=[math.log(1.0+((dl_n-d+0.5)/(d+0.5))) for d in df]
L=[]
for j,i in enumerate(hit_idx):
norm=1.0-B+B*(DL[i]/avgdl); w=0.0
for t in range(len(toks)):
if hit_mask[j]>>t&1: w+=idf[t]*(K1+1.0)/(1.0+K1*norm)
L.append((i,w,SAL[i]))
L.sort(key=lambda x:(-x[1],-x[2]))
qv=qemb(query)
cos=En@qv
cos=np.where(HAVE&OK,cos,-2.0)
order=np.argsort(-cos)
semfull=[(int(i),float(cos[i])) for i in order[:400]]
Sleg=[(i,(c-SEED_MIN)/(1-SEED_MIN)) for i,c in semfull if c>SEED_MIN]
semseed=[i for i,c in semfull[:SEED_K] if c>0.0]
# assoc
act={}; seen={}; qq=[]
for i,_,_ in L[:ASSOC_SEEDS]:
act[i]=1.0; seen[i]=2; qq.append((i,0))
for i in semseed:
if i in seen: continue
act[i]=1.0; seen[i]=2; qq.append((i,0))
qh=0
while qh<len(qq):
cur,h=qq[qh]; qh+=1
if h>=DEPTH: continue
parent=act[cur]
for e,oi in ADJ_F[cur]+ADJ_T[cur]:
if e['rel'] not in STRUCT: continue
if EXCL[oi]: continue
na=parent*e['w']*DECAY*SAL[oi]
if na<FIRE: continue
if seen.get(oi) and na<=act.get(oi,0): continue
act[oi]=na
if not seen.get(oi): seen[oi]=1
if len(qq)<AMAX*4: qq.append((oi,h+1))
A=[]
for i,st in seen.items():
if st!=1: continue
if not OK[i] or not HAVE[i]: continue
c=float(cos[i])
if c<=0.0: continue
A.append((i,c))
A.sort(key=lambda x:-x[1]); A=A[:AMAX]
return L,Sleg,A,cos
def interleave3(L,Sl,A,lim=10):
out=[]; li=si=ai=0
while len(out)<lim and (li<len(L) or si<len(Sl) or ai<len(A)):
if li<len(L):
if L[li][0] not in out: out.append(L[li][0])
li+=1
if len(out)>=lim: break
if si<len(Sl):
if Sl[si][0] not in out: out.append(Sl[si][0])
si+=1
if len(out)>=lim: break
if ai<len(A):
if A[ai][0] not in out: out.append(A[ai][0])
ai+=1
return out