[MODEL] Claude repeatedly skips established audit protocols despite explicit CLAUDE.md instructions

Resolved 💬 3 comments Opened Mar 4, 2026 by grooovinh Closed Mar 8, 2026

Preflight Checklist

  • [x] I have searched existing issues for similar behavior reports
  • [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

During a multi-session qualitative audit, Claude was provided with a Phase 3 Contract v2.0 (PHASE_3_CONTRACT.md) requiring 3-dimensional per-entry scoring, JSONL ledger output, and 20 verification checks. CLAUDE.md explicitly states: "READ the relevant phase contract before executing ANY audit phase."

What Claude Actually Did

Despite this, Claude never read the contract for 99 files across 9 agent invocations, using a simpler protocol instead. This is the 5th documented instance of Claude silently downgrading quality requirements in this project. The content fixes are valid but the formal audit trail is completely missing, requiring full re-audit.

Expected Behavior

Expected behavior is for claude to follow instructions explicitly stated and to maintain a qualitative assessment as defined in the contract.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Claude claims the issue will never happen again, I will let it run a few cycles of compaction and then it happens again. Even when it claims to refresh the context/memory each time.

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

4.6

Platform

Anthropic API

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗