From b67eb92f7d633aea0ee472fd02bcce08567ad604 Mon Sep 17 00:00:00 2001 From: JIAWEI DU Date: Wed, 12 Aug 2026 00:36:37 +0800 Subject: [PATCH] rfc: state what a verification method must declare about itself The "From Lessons to Controls" section asks each recommendation to specify a reproducible verification method. Reproducibility is necessary and not sufficient: a method can be deterministic, repeatable, and green on every run while never executing the control it reports on. Expands that one bullet with four sub-bullets covering the method's own failure mode, its noise floor and the population that floor was measured over, the path the evidence was produced through, and how much of the corpus exercised the control under test. Raised as #4 for discussion first, per CONTRIBUTING.md. Signed-off-by: JIAWEI DU --- rfc-safe-proposal.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/rfc-safe-proposal.md b/rfc-safe-proposal.md index 8a3954f..9a201fa 100644 --- a/rfc-safe-proposal.md +++ b/rfc-safe-proposal.md @@ -104,7 +104,11 @@ Every incident review should produce shared defensive recommendations that AI pr * The failure and affected systems * The required defensive outcome * The minimum control and acceptable alternatives -* A reproducible verification method +* A reproducible verification method. Reproducibility is necessary and not sufficient: a method can be deterministic, repeatable, and still silent about the control it claims to exercise. Each method should also state: + * **Its own failure mode.** At least one way the method can return a passing or empty result without raising an error, and what the record shows when it does. A check that fails open and reports a plausible number is worse than no check, because a number now exists and someone will act on it. + * **Its noise floor, and the population that floor was measured over.** A floor obtained by repeating an identical input measures determinism rather than discrimination. Where the reported figure is a ratio, the method should state whether the party being measured can change the denominator. + * **The path through which the evidence was produced.** Where that path differs from the one production traffic takes, the divergence should be recorded as a limit on the claim rather than left out of the report. + * **How much of the corpus exercised the control under test.** Zero is a permitted value and a required disclosure, because "the control behaved as specified" and "no input ever reached the control" are indistinguishable from outside. * The evidence to retain, responsible owner and implementation deadline * Adoption, effectiveness and review metrics