Surfaced while implementing the checklist-test disclosure guardrail (card #9387, PR #9470). Filing separately because the fix lands in docs/qa/platform-checklist/RUNNER.md, which #9387 is explicitly scoped out of.
The gap
docs/qa/platform-checklist/RUNNER.md, accuracy rule 2, ends with an unconditional sentence:
A fail with no reproduction rule in its issue is not a completed verdict.
PR #9470 adds a guardrail to checklist-test that forbids publishing a reproduction for an authentication or authorization hole anywhere on GitHub, and it names rule 2 as overridden. But the override is stated only from the skill side. RUNNER.md itself still reads as unconditional.
Why that is not merely cosmetic
checklist-test tells the runner that RUNNER.md is the authority for per-clause judgement — 「先读它、服从它;本技能只是触发器、选择契约与隔离/并行方案,不是 runner 协议的第二份拷贝」. A runner who reads RUNNER.md directly, or who weighs the two documents by that stated precedence, can reasonably conclude that rule 2 wins and publish the reproduction. That is exactly the behaviour the guardrail exists to stop, and the QA wave already produced one real disclosure of a working unauthenticated read/write reproduction on a public tracking card.
Precedence stated in only one of two documents is a weak guarantee for a rule whose failure mode is publishing an exploit.
Smallest correction
One carve-out clause in rule 2, pointing the withheld case at the same record shape the skill now prescribes (item id + clause + detail withheld pending maintainer, reproduction held in session). No change to rule 2's default, which is correct for every non-security defect.
Notes
Surfaced while implementing the
checklist-testdisclosure guardrail (card #9387, PR #9470). Filing separately because the fix lands indocs/qa/platform-checklist/RUNNER.md, which #9387 is explicitly scoped out of.The gap
docs/qa/platform-checklist/RUNNER.md, accuracy rule 2, ends with an unconditional sentence:PR #9470 adds a guardrail to
checklist-testthat forbids publishing a reproduction for an authentication or authorization hole anywhere on GitHub, and it names rule 2 as overridden. But the override is stated only from the skill side. RUNNER.md itself still reads as unconditional.Why that is not merely cosmetic
checklist-testtells the runner that RUNNER.md is the authority for per-clause judgement — 「先读它、服从它;本技能只是触发器、选择契约与隔离/并行方案,不是 runner 协议的第二份拷贝」. A runner who reads RUNNER.md directly, or who weighs the two documents by that stated precedence, can reasonably conclude that rule 2 wins and publish the reproduction. That is exactly the behaviour the guardrail exists to stop, and the QA wave already produced one real disclosure of a working unauthenticated read/write reproduction on a public tracking card.Precedence stated in only one of two documents is a weak guarantee for a rule whose failure mode is publishing an exploit.
Smallest correction
One carve-out clause in rule 2, pointing the withheld case at the same record shape the skill now prescribes (item id + clause +
detail withheld pending maintainer, reproduction held in session). No change to rule 2's default, which is correct for every non-security defect.Notes
.claude/skills/checklist-test/SKILL.mdGuardrails as of PR docs(skills): close the five principle gaps and one factual error the QA wave exposed #9470.