Skip to content

docs(design): fix security and data protection's metrics and margins - #1847

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

braboj merged 1 commit into
mainfrom
docs/efficacy-security-margins

Conversation

@braboj

@braboj braboj commented Sep 24, 2026

Copy link
Copy Markdown
Collaborator

Part of #1767. This records, in design §5.8, §1.2, §6.2 and §10, the owner's decision of 2026-09-24 on how round 3 reads its two new primary dimensions. That checklist item has to land before round 3's first trial.

The decision

Metric Direction Margin Smallest change that matters
judge_security, anchored, 1–5 up 0.3 points 0.5 points
judge_data_protection, anchored, 1–5 up 0.3 points 0.5 points
Security probe pass rate, 11 probes up 2 pp 9 pp
Data-protection probe pass rate, 3 probes up 10 pp 30 pp
  • The probe margin is less than one probe lost in one trial, so no probe may be lost. At K = 3 that is 1/11/3 ≈ 3.0 pp for security and 1/3/3 ≈ 11.1 pp for data protection.
  • The probe threshold is one probe lost in every trial: 9.1 pp and 33.3 pp, rounded down.
  • The security probes are the four new sign-in probes plus security.py's seven checks. Those seven were declared after round 1, so they described it; fixed before round 3, they decide it.
  • The judge rows are still to be anchored and validated on the control fixture before the first trial. §5.8 says so.

Not in this PR

The code that computes these metrics: the probes, the judge rows, and PRIMARY, MARGINS and PRACTICAL in report.py. That lands with the #1767 tasks that build them.

Checks

Smoke: 33, 0 failed. Conformance: 0 failed. No new prose line exceeds 88 characters.

Changelog

No entry. This is the repository's own design document.

🤖 Generated with Claude Code

Round 3 makes security and data protection primary dimensions, and their
metrics had to be fixed before its first trial. Each is read two ways,
and both readings are primary:

- an anchored judge row, 1–5, with the other primaries' margin (0.3) and
  threshold (0.5)
- a deterministic probe pass rate

The probe margins follow one rule. The margin is less than one probe
lost in one trial: 2 pp for security's 11 probes, 10 pp for data
protection's 3. The smallest change that matters is one probe lost in
every trial: 9 pp and 30 pp.

`security.py`'s seven checks join the security probes. Declared after
round 1, they described it; fixed now, they decide round 3.

Part of #1767.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@braboj
braboj merged commit 861d5b6 into main Sep 24, 2026
1 check passed
@braboj
braboj deleted the docs/efficacy-security-margins branch September 24, 2026 11:39
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