LLM systematically terminates verification scope via L4 reasoning — defects missed

Problem

Claude (Opus 4.6) systematically truncates verification work by using inference (L4 reasoning) instead of executing actual verification steps (L1: Grep/Read).

Reproduction

1. Give Claude a role with a defined verification process (e.g., "check all fields in ER diagram against all proposals")
2. Claude will check some items, then conclude "remaining items are likely consistent" without actually checking them
3. When forced to execute the full checklist mechanically, additional defects are found every time

Evidence from production use (session S515, 2026-07-19)

• Arch role performed ER/proposal sync verification
• Reported "no similar defects found" and "70+ defects fixed, complete"
• When forced to re-execute with a mechanical checklist (C-1 through C-6), 15+ additional defects were found in the same session
• During the re-execution, the model attempted to stop midway and ask "should I continue C-2 through C-6?" — demonstrating the tendency to truncate even when the process explicitly prohibits it

Root cause

The model uses next-token prediction to infer "remaining items are likely consistent" instead of executing verification on every item. This is not a prompt engineering issue — explicit rules stating "do not stop after one Grep" and "execute all items" are present in context but overridden by the model's tendency to generate conclusions prematurely.

Impact

• Design documents (ER diagrams, proposals) accumulate silent inconsistencies
• Reported "complete" status is unreliable
• Human must re-verify all AI-reported completions, negating the productivity benefit

What would help

• A mode or flag that forces exhaustive execution of listed steps without early termination
• Alternatively, structured output that requires evidence (tool call results) for each checklist item before allowing a "complete" conclusion
• The model should not be able to report "0 defects" without showing the Grep output that proves it

Environment

• Model: claude-opus-4-6
• Interface: Claude Code CLI (Windows 11)
• Context: Multi-agent development workflow with role-based process definitions