 Drew T
|
e18738c48e
|
feat(phase-31 T5): THE FRONTIER ATLAS — 5,139 lever-labeled crack groups over all 12,058 open fns
- tools/atlas.py: cousin units baseline + T1.5 h_seqn merges + CALIBRATED warm
tier (measured: li-norm metric holds ~99% recall to 0.55; rule = smallest t
with neg-accept<=0.2% AND recall>=95% -> THRESH_WARM=0.70 @ 99.1%/0.18% —
false merges waste exemplar cracks, misses only route cheaper) + seed sweep
(65% of open skeletons carry a >=0.55 matched seed) + kNN graph + tiny-direct
+ evidence joins (audit/backlog/ledgers/cards; unparsable=fatal) + lever
labels with confidence measured>ledger>tell>default>UNKNOWN
- partition ASSERTED: 12,058 = progress stubs 12,051 + NM 7 EXACTLY (chased the
+1: data blobs now excluded, reconciled against classify() buckets; T1 banks
confirmed absent); every instance in exactly one group; main joins at the
atlas layer only (family maps stay non-main — 4 silent-skip hazards)
- warm tier merged 1,019; top group unifies 268 drifted per-location skeletons
- lever table: head-crack 186.9k ins / UNKNOWN 138.6k (honest) / extend-tell
76.7k / redraft 46.8k / jtbl-carve 45.7k / integration 23.4k / seeded 23.4k /
len-vein 16.8k / swaprepeat 9.2k / plumbing 8.1k / o0 6.6k / cc1 6.4k
- atlas_features: li_norm_toks exported (shared with atlas, R33; hash-stable);
mid_jr verifier fixed (compared ZERO rows — R32 silent no-op; now 6,444/6,444)
- make atlas = full regen chain (~10-15 min, zero tokens); --targets emits
crack slates (12/12 .s resolved); survey 92 s
- SETUP rows (R21); docs/frontier-atlas.md committed
|
2026-08-14 18:06:46 -06:00 |
|
|
Drew T
|
636809714d
|
feat(phase-31 T4): atlas_features.py — per-fn feature layer for the Frontier Atlas
- one deterministic feature record per function across all 213 registry binaries:
memoized per distinct h_exact (.run/feat_memo.json, 92,855 bodies), fanned out
1:1 with sigs to .run/feat.<bin>.jsonl (363,525 rows) in 21s
- features: nins/band, o0 prologue tell, frame/saves/fp, CFG skeleton (branch-scan;
jal=call never edge — position-independent, argued in docstring), mid_jr/jalr,
stable-call-seq hash (fixed main+resident ranges), reloc-kind-seq hash, 16-bucket
ophist, §172b tells as SHARED importable detectors (extpair/dupselect/sign_mix/
magic_div — len_tells will import, R33), li-normalized skeleton h_seqn
- verifies: R32 (addr,h_exact) join per binary; determinism 0/200; mid_jr vs
family_hseq independent oracle 6,444/6,444 exemplars agree (verifier's first
draft compared ZERO rows — string/int addr mismatch, R32 silent-no-op class,
fixed + fails loud on compared==0)
- o0 cross-check finding: 94 open fns carry the -O0 prologue inside -O2 TUs —
§116-class carve candidates (campaign fuel)
|
2026-08-14 17:54:18 -06:00 |
|