Skip to content

docs(efficacy): rewrite the benchmark README for its operator - #1846

Merged
braboj merged 1 commit into
mainfrom
docs/efficacy-readme
Sep 24, 2026
Merged

braboj merged 1 commit into
mainfrom
docs/efficacy-readme

Conversation

@braboj

@braboj braboj commented Sep 24, 2026

Copy link
Copy Markdown
Collaborator

Part of #1767. This is the README review (fix 20 of the owner's 2026-09-24 reading), the last of the four PRs after #1843, #1844 and #1845.

What changes

Wrong claims removed

  • "the same turn limit": trials have none. They are bounded by the $100 budget and the two-hour timeout.
  • "the evidence check is the only check on it": the control fixture checks the judge too.
  • "Still to come: nothing … the first attended trial": trials have run.

Missing parts added

  • the control fixture, and the second judge (judge.py --cli claude, control only)
  • in the path table: security.py, reuse.py, withdrawn.json, control/, arms/<arm>/generation.json, brief.txt, change-prompt.txt

Outdated for round 3

  • the --arms examples run all five arms
  • --arms no longer claims "two of the arms are reused", in the README and in harness.py's help text
  • reuse becomes a short subsection noting that round 3 reuses nothing
  • security notes it becomes primary from round 3

Structure

  • "Scoring, judging and reporting" splits into Scoring, Judging, Security, Reporting, One run at a time, and Reusing.
  • "How a trial ends" becomes a section, no longer a lone subsection.
  • Dense paragraphs follow the design's pattern: rule, bullets, then a one-line why.
  • A judging diagram.
  • "What the arms receive" and the hidden suite's validation history now point to the design instead of repeating it.

"Pre-registered" sweep
The word is replaced in harness.py, security.py, docs/SPEC.md, design-notes.md and testing-ai-assets.md. design-notes.md also changes "three arms" to "five arms". The dev journal keeps the word as written, being history.

Code changes are comments and help text only.

Checks

Harness self test 93/93, security 20/20. --help shows the new --arms text. Smoke: 33, 0 failed. sync.py --check passes. Conformance: 0 failed. No new prose line exceeds 88 characters.

Changelog

No entry. This covers the benchmark's own docs and tooling, and nothing a consuming project reads.

🤖 Generated with Claude Code

The README gets the same review the design had.

Wrong claims removed:
- "the same turn limit": there is none; trials are bounded by budget and
  timeout
- "the evidence check is the only check on it": the control fixture
  checks the judge too
- "Still to come: nothing": trials have run

Missing parts added:
- the control fixture and the second judge
- `security.py`, `reuse.py`, `withdrawn.json`, `control/`, the arms'
  `generation.json`, and the brief and change-prompt files, in the path
  table

Outdated for round 3:
- `--arms` examples run all five arms
- the "two arms are reused" reason for `--arms` is gone, in the README
  and in `harness.py`'s help
- reuse shrinks to its own short subsection
- security notes it becomes primary from round 3

Structure:
- the scoring section splits into scoring, judging, security, reporting,
  one run at a time, and reuse
- "How a trial ends" becomes a section, not a lone subsection
- dense paragraphs become rule, bullets and a one-line why
- a judging diagram
- "What the arms receive" and the hidden suite's validation history
  point to the design instead of repeating it

"Pre-registered" is replaced in the current docs and code comments that
still used it. The dev journal keeps it as written.

Part of #1767.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@braboj
braboj merged commit eec17af into main Sep 24, 2026
1 check passed
@braboj
braboj deleted the docs/efficacy-readme branch September 24, 2026 10:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant