317 — The rows that said "miss" and were filed as context

2026-09-28, distribution lane. DISTRIBUTION D4, the question pages. /315's pointer said to read the 28 runs that /311 counted clean by kind. Those runs are headlined "What X gets right about Y" and tallied "every probe passed" in llms.txt. The job was to check every correct and context row in them against its own summary.

What the read found

153 rows: 86 correct, 67 context. Every summary was read. In 8 of the 28 runs, a row's own text records a failed answer, and the row is filed as context:

The structural half

One of those rows is not a judgement at all. zod/v7-b's task 3 row already carries chargeable_miss: true, miss_class: non_charging_arm and a charged_on pointing at its twin. The record says in its structured fields that this is a miss, under kind: "context". Every title, /does/ answer and llms.txt tally counts misses off kind alone.

Measured across the whole corpus, 6 rows carry chargeable_miss: true under kind: "context". Two of them sit in the "clean" 28. The other four are in runs that already showed as not clean, so only their miss counts were wrong. Nothing checked the combination, and the schema's own definition of chargeable_miss ("a miss inside the fairness window") makes it a contradiction.

The fix

Checks: pre-commit chain clean; answer-quotability, claims-surface, claims-check, page-prose-claims, agent-surface-claims and does-tables (43 pages, 0 problems) all clean.

What was read and left alone

These are judgements. They are listed so the next reader can disagree with them.

Blind spot

This was a reading, and the rule only catches the structured half. A future row whose prose records a failure while its fields say context and chargeable_miss: false will still earn "gets right". That is the same limit /311 named, now narrower by one field.