The battery oddity was two holes, and both are closed

[punnerud]

The full-battery run recorded an oddity: unit_chain 0/2 while the units domain went 8/8 — "same task, different phrasing, n=2, no conclusion drawn." Chased properly, it was not noise. It was two distinct holes, one behind the other.

Hole 1: the synthesis does sums, and nothing checked them

A failing answer read 378 × 86,400 = 32,356,800 seconds — right expression, wrong product — and the prose gate only ever ran on step content. repair_sums() now runs on the synthesised answer: the claim regex catches the shape, the evaluator knows the exact value, and every rendering of the wrong number is replaced with it (digit boundaries respected — 1.42 is never touched by a repair of 42).

Measured: when it fires, it decides — 2 of 2 repaired runs correct.

Hole 2, the dominant one: bare assertion

On twelve seconds-in-N-weeks questions the model asserted a bare unchecked number eight times — "There are 1,612,800 seconds", no expression, no calc, no convert. A bare assertion gives every gate in the pipeline nothing to hold.

So a question that is itself one conversion is settled before the first model call: the exact value goes into the facts the synthesis prefers and the selection can offer, as a step-0 convert event. The reasoning runs as usual — around an anchor instead of a hope.

Measured

twelve identical questions, same seed	correct
before	4/12 (2 rescued by the repair)
with the anchor	12/12, zero repairs needed

And the original two battery failures, re-run: 0/2 → 2/2 with the machinery, still 0/2 without it — the mechanism, not luck.

The shape stays deliberately narrow: how many X in N Y and nothing else; question_conversion returns None for prose, arithmetic and word problems, tested.

Also: the bench filename now carries --group, after a two-question rerun overwrote the pinned twenty-question baseline — the second near-miss of that kind. Restored from git, claims re-checked.

11 new tests, 426 in all.