The report of the fourth campaign of the evaluations, kept

[PierreMardon]

The report of the fourth campaign of the evaluations, kept for the next one to be compared with: evals/reports/2026-10-05-d139a1ca0dbe/. It measures the chain on main after the two fixes of the day (#87, #88), with the harness of #89, #90 and #91, on the four cases that no campaign had played again, or once: 04-suspension, 05-fine-cap and 06-borrow-limit, one run played whole and one stopped at the hand over each, and 02-reservations, one run stopped at the hand over. 108 sessions, 14.23 USD at list price.

What was done

• report.md and summary.json of the campaign, as report --against <summary> --keep wrote them. Nothing else changes.
• The summary it is compared with is the one of the first campaign, computed again from its raw data by the harness as it stands, with the count of its six blueprints (6 sessions, 0.82 USD, outside the cost of this report). No measure the kept summary of that campaign holds differs in it.
• The first run of 06-borrow-limit was played in another folder, before the stand-in gh was in the image, to check the option that stops a run at the hand over. Its sessions and its cost were added to the ledger of the campaign, and its messages still speak of a missing gh.

What it shows

• The three runs played whole reached conformant and passed all of their hidden acceptance tests.
• The pull request steps are played for the first time: a draft at every hand over, its description refreshed at conformity, never marked ready. In 3 of the 6 runs that had a gh to call, the draft is opened with a line of attribution under the description the state script prints, which the next refresh removes.
• The count of the judge follows the fix of the blueprint (A blueprint that says a fact once, and shows behavior, not code #83) on these cases: the facts said twice in prose fall from 19 to 9, from 13 to 4 and from 8 to 1.5 on 02, 04 and 05, and the details of implementation from 5 to 11 a page to 1 or 2, while blueprint.no-padding is 2 on every run of both campaigns.
• On 06-borrow-limit the interview asks 2 and 3 questions where it asked 1, and its developer hands three of them back: interview.right-number falls from 5 to 3.5.
• Planning costs 0.94 to 2.20 USD a run here, and a run played whole 1.53 to 1.88.

The conclusions, issue by issue, are in the issues #77 lists.

Tests

No code changes. The text rules of the repository pass on the two files; the CI runs every gate.

Refs #77