The README's generated numbers block still publishes M4; M5 measured the thing the README leads with

[MongLong0214]

Found while putting M5 into the README for 0.7.0.

What is stale

bench/report.ts's README_SOURCES declares one file:

bench/results/t702-m4-final.jsonl   112 rows   model: not recorded

The block it generates is what scripts/check-readme-numbers.mjs regenerates and byte-compares in CI. It is the only place the README is allowed to carry a measured statistic — and it publishes M4, while the README's behaviour claim now rests on M5.

Why this was not fixed in the release

README_SOURCES is deliberate machinery, not a config value. Its own comment explains why a glob would be dangerous, why M1/M1-b/M2 are not pooled, and why the ablation log is excluded. Repointing it at M5 means declaring seven shards with a status note covering supersession (80 cells superseded by re-run), the excluded error row, and the per-arm truncation split — the same care the M4 note took.

Doing that between a CI failure and a tag would have been the wrong moment. The release ships with the statistics living in bench/VERDICT-M5.md and the README linking there.

What the gate did right

It rejected the first attempt, which had p = 0.0000, −16.1pp and Fisher exact typed into the prose. That is exactly the failure it exists to prevent: a statistic retyped into a paragraph drifts from the log that produced it, and the generated block below is worth nothing if the paragraph above it carries a hand-written one.

Scope

• Declare the seven M5 shards in README_SOURCES with a status note in the M4 note's register: supersession rule, exclusions, the single harness and dist digest, and what the number does not cover ([claim]-tier delivery, one model, ten constructed fixtures).
• Decide whether M4 stays declared beside M5 or is retired to historical — they are different matrices with different arms, and the existing note already says pooling them would be wrong.
• Regenerate all four READMEs with --write.

Acceptance

• The generated block reports M5's rows, run ids, driver, model and matrix
• node scripts/check-readme-numbers.mjs passes with no --write
• The status note names the supersession rule and the excluded row
• Whatever happens to M4 is stated in the note rather than implied by its absence

[MongLong0214]

Closed by #489, merged to dev.

The seven M5 shards are declared with a status note in the register the M4 one set: why three are re-runs, where the supersession rule lives, the excluded error row, the unequal truncation and which direction it pushes, and that every record rendered [claim] so this is not the [directive] path 0.7.1 made reachable.

Repointing found a second thing worth more than the repoint: the generator counted rows on disk, so the block reported 1,240 where the registered analysis reports 1,160. It now counts task-arm-seed cells the way bench/m5-analysis.ts does and names both figures, so the published line and the registered analysis cannot disagree without one of them changing.