Correct twelve documented facts that drifted from the files

[aaddrick]

Every gate passed while all twelve were wrong. No check reads prose.

The expensive one

evals/README.md still quoted the isolation prompt as "You cannot run commands or read files in this conversation." 375b7b4 removed that sentence from evals/runners.example.json, and tests/test_run_evals.py lists it in CONTAMINATED_WORDINGS. A reader who copied the documented prompt into evals/runners.json would repeat the run 002 contamination: 6 of 21 blockers on agent-owned-edit alone, at about $11 and two hours per run. The section now quotes the shipped text and tells readers to copy from the config, not from the doc.

Case count: 20 -> 24

04b16f7 added four uncertainty cases. Seven places still said 20:

• README.md and all five translations
• both attribution counts in evals/README.md
• the position-balance claim, twice

Both frozen blind-key.jsonl files measure 36 of 72, not 30 of 60.

NOTICE.md

Three attribution errors:

1. The shape layer holds 11 rules, not 10. Rule 2 "Do the work you own" was added in 04b16f7, so crediting it upstream overstates what the source contributed. It is now listed as new here.
2. "ships as a Claude Code output style, not a skill" is false. Six harnesses get it as a skill.
3. Six new cases is now ten: two language, two verbatim, one decision, five uncertainty. evals/README.md already named that set.

14 of the cases was correct and stayed correct.

The rest

• The freeze example used run 002's id with run 001's judge cost and a note the ledger contradicts. It now shows an unclaimed 003.
• The two 120 responses isolation numbers predate the ledger and the 24-case catalog. They now say so, and the blocker counts name their runs.
• CI runs a fifth gate. CLAUDE.md and CONTRIBUTING.md both undercounted it, so a maintainer running the documented set could still fail on the ledger diff.
• CONTRIBUTING.md named four generated files. sync_style.py writes five, and AGENTS.md comes from CLAUDE.md, not from the style.

Also here

A "Connect on LinkedIn!" line in all six README headers, translated in the five non-English files.

Gates

sync_style.py --check     All 5 generated files match their sources.
check_configs.py          All 8 shipped config files parse.
run_evals.py validate     Evaluation cases are valid.
unittest discover         Ran 75 tests ... OK
ledger.py index           no diff

No frozen run changed. AGENTS.md is regenerated, not hand-edited.