battery provenance: a recorded pool, a prereg that prices the extraction term, and three limits on the existing reads

[fathomlab]

Supersedes #53 — identical content plus the ledger rebuild, on a fathom-lab branch so CI runs where it can be merged. #53 was opened cross-fork because the authoring account has pull-only access here; that is worth fixing separately.

E is measurable prospectively

POOL 547,373 → ELIGIBLE 5,199 → BATTERY 462, survival ratio 0.00084.

The 462-concept battery the flagship reads are scored over is a hand-authored string literal — 465 words typed into run_g0clear.py lines 31-65, with exactly one filter (an order-preserving dedup removing chicken, orange, mushroom). No tokenization, vocabulary, frequency, POS or quality filter exists anywhere in that 188-line file. So the extraction term for those reads is UNCHECKABLE retrospectively: the pool never existed as an artifact, and there is no rejection log, no source corpus and no seed to replay.

This builds the pool that should have been there. Source wikitext-103-raw-v1 pinned at an HF revision, CC BY-SA 3.0, resolved from cache — no weights loaded, no GPU. Eight filters, each recording rule, justification, cost, removal count and a sha256 of its survivors. Two determinism leaks found and fixed that would have tripped the prereg's own VOID condition off Windows.

Three limits on the existing reads

These bound the published reads; they do not overturn them.

1. Selection provenance is unrecorded. Confirmed exactly. The parent bank is 121 words, not the 110 the docstring and parent_baseline.N both claim; 341 words are net new and how they were chosen is recorded nowhere.

2. The held-out set's held-out-ness has been spent — and it is larger than first reported: nine committed experiments score the seed-0 FIN-70, not seven. Never fitted on for the map or either headline read, but three write-side scripts select a hyperparameter on a FIN slice, one commented verbatim "held-out concepts used to pick the steer-optimal layer" — always on native target directions, so not map leakage, but not untouched either. And SEL_dirs is 60-of-69 inside b34v3's training set with 9 inside its held-out set, so "sits inside the training set" is true of 60 and false of 9.

3. The "open-vocabulary" readout is a closed list — with a caveat that reduces its weight. run_b35c.py:46 is vocab = fin + tr, the same 462. But the docstring and the prereg both state that outright, and b35c_result.json returned INVALID__null_artifact, so no open-vocabulary claim is published from that run. The overclaim is in the name, not in a standing number.

An independent reproduction of the split returned the 13/70 overlap matching b34v3_fresh_split_addendum.json exactly, which is what licensed the rest of the split arithmetic.

Disclosed weakness, named in advance

The pooled battery is broader in word class than the hand-authored one, because no offline POS tagger was available. The prereg names this in advance as the leading explanation if the reproduction gate fails, so it cannot be produced later as a convenient excuse.

Ledger

papers/LEDGER.md rebuilt: 380 → 381 preregs, 40 → 41 gated, 9 → 10 declaring a power basis, 8/8 → 9/9 frozen since the field existed. The new prereg declares one, which is why the second ratio stays whole.

🤖 Generated with Claude Code