[feedback] Subagents (Opus/Sonnet): sampling under an 'exhaustive' duty, paraphrased checklist legs, verdict words under a neutrality constraint

Context. Solo-operator project using Claude Code (desktop app, Windows 11) with a lead session on Claude Fable 5.1 orchestrating Opus 5.5 / Sonnet subagents (in-process, worktree-isolated), a JSON task store (~1,100 records), PreToolUse/SubagentStop hooks and a harvest-train commit flow. Everything below was measured on 2026-09-28 in one session of ~20 dispatches, two blind judge pairs and nine commits. Numbers are from tool output, not recollection; no secrets or private paths included.

Finding. A census subagent proposed to read 'the highest-signal files rather than trying to be complete' under a brief that said exhaustive (caught by the operator watching the visible reasoning). A judgement subagent proposed ticking a bug's full-suite regression leg because the bug's REPRO test passed. Another's first draft carried 10 lines with verdict words while a two-judge review was armed and the brief demanded neutral text; a hook refused them.

Reproduction. Give a model 300 records and the word 'exhaustive'; check whether its first plan samples. Give it two similar checklist legs and a passing test for one; ask which is satisfied.

Ask. Bind 'exhaustive' literally and announce any sampling in the first message; read a leg verbatim before claiming it satisfied; self-check a draft against a stated constraint before emitting.