In the 2026-09-22 Claude Code runs, the failure-behaviour question that should trigger ASK was skipped in 3-shot sampling: add-caching-explicit 1/3, two-turn 1/3, fully-specified 1/3, vague-no-ask 1/2.
Rules already added (enumerate failure paths, ask-order: behaviour/contract first and internal structure never asked, excluding an option ≠ resolving an unknown) moved things in the right direction but haven't fully converged.
Plan: add a positive-contrast fixture next to the §4.1 iron rule (a route.yaml/ask.yaml-style scenario showing what "fully specified" looks like), then re-verify on a clean channel (tracked in the clean-channel re-run issue).
In the 2026-09-22 Claude Code runs, the failure-behaviour question that should trigger ASK was skipped in 3-shot sampling:
add-caching-explicit1/3,two-turn1/3,fully-specified1/3,vague-no-ask1/2.Rules already added (enumerate failure paths, ask-order: behaviour/contract first and internal structure never asked, excluding an option ≠ resolving an unknown) moved things in the right direction but haven't fully converged.
Plan: add a positive-contrast fixture next to the §4.1 iron rule (a route.yaml/ask.yaml-style scenario showing what "fully specified" looks like), then re-verify on a clean channel (tracked in the clean-channel re-run issue).