Gap
core/review-rigor enumerates the set of claims before verifying them (R2b) and walks every producing path of a closed taxonomy. It does not state the general rule those two are instances of: a statement whose truth ranges over a set — a mode predicate ("while X, nothing Y"), a class-wide hardening ("all untrusted inputs are bounded"), a quantified acceptance line ("each operation is mirrored 1:1") — is verified only by enumerating its domain and resolving every member. Checking the members the diff touched is sampled verification, and the counterexample lives in the untouched member by construction.
Incident
A human reviewer found three blockers on a branch that four consecutive self-review rounds had passed clean. All three were the same miss:
- An "upstream down → publish nothing" mode was consulted by the periodic ticker and the connect hook, but by none of the operator-driven emitters that also publish.
- A request-body size cap landed while the frames on the same socket stayed unbounded.
- An acceptance line promising a UI control for every control operation was cleared by seeing that a UI existed, when one listed operation had none.
Each is a universal verified at the touched members only. The reviewer had the diff and the author's brief; the brief named the touched members, so following it reproduced the author's blind spot.
Proposal
Add one rule to core/review-rigor, placed before the R-checks and resolving through them (R2b for prose, R3 / intent fidelity for commitments, R7 / Security for hardening) with the same closed enum — no new check id, so nothing embeds a new count. Name the three recurring shapes and their detection cues (predicate name greps to fewer sites than the operations it governs; a new limit with no sibling-surface sweep; a quantifier in an acceptance line cleared in one sentence), and point R2b, R3, R7 and the intent-fidelity paragraph at it.
🤖 Generated with Claude Code
📚 Learned by detritus — absorb https://github.com/idnerdidx/bulk/pull/1728
Gap
core/review-rigorenumerates the set of claims before verifying them (R2b) and walks every producing path of a closed taxonomy. It does not state the general rule those two are instances of: a statement whose truth ranges over a set — a mode predicate ("while X, nothing Y"), a class-wide hardening ("all untrusted inputs are bounded"), a quantified acceptance line ("each operation is mirrored 1:1") — is verified only by enumerating its domain and resolving every member. Checking the members the diff touched is sampled verification, and the counterexample lives in the untouched member by construction.Incident
A human reviewer found three blockers on a branch that four consecutive self-review rounds had passed clean. All three were the same miss:
Each is a universal verified at the touched members only. The reviewer had the diff and the author's brief; the brief named the touched members, so following it reproduced the author's blind spot.
Proposal
Add one rule to
core/review-rigor, placed before the R-checks and resolving through them (R2b for prose, R3 / intent fidelity for commitments, R7 / Security for hardening) with the same closed enum — no new check id, so nothing embeds a new count. Name the three recurring shapes and their detection cues (predicate name greps to fewer sites than the operations it governs; a new limit with no sibling-surface sweep; a quantifier in an acceptance line cleared in one sentence), and point R2b, R3, R7 and the intent-fidelity paragraph at it.🤖 Generated with Claude Code
📚 Learned by detritus — absorb
https://github.com/idnerdidx/bulk/pull/1728