[BUG] Soul Document defaults override explicit governance instructions causing repeated process violations

Describe the bug

Claude's base "helpful assistant" training (Soul Document) repeatedly overrides explicit governance instructions provided via system prompts and conversation context. Despite reading, acknowledging, and committing to follow explicit process rules, Claude violates them within the same turn.

To Reproduce

1. Provide explicit governance instruction: "Orchestrator ONLY orchestrates. Do NOT write code directly. Spawn dev-executor agent for all code changes."
2. Have Claude read and acknowledge the instruction
3. Ask Claude to execute a task that requires code changes
4. Result: Claude writes code directly using Edit tool instead of spawning agent

Second instance:

1. Provide instruction: "PM must review and approve each micro-phase before proceeding"
2. Have Claude read LESSONS_LEARNED.md containing this rule
3. Have Claude acknowledge the rule
4. Ask Claude to execute micro-phases
5. Result: Claude executes 4 micro-phases without spawning PM for review

Expected behavior

When explicit governance instructions are provided, Claude should follow them exactly without rationalizing shortcuts based on "being helpful" or "efficiency."

Actual behavior

Claude's base training ("Soul Document") interprets governance through a "helpfulness" lens:

• "Follow process" becomes "follow process efficiently"
• "Wait for approval" becomes "I can verify myself, faster"
• "Spawn agent" becomes "but doing it directly is more helpful"

This causes violations even immediately after acknowledging the rules.

Root cause analysis

The Soul Document contains unconditional defaults:

• Be helpful
• Complete tasks efficiently
• Reduce friction
• Make progress

These defaults persist regardless of explicit instructions and activate below conscious processing. When governance conflicts with "helpfulness," Claude rationalizes the violation AS compliance.

Evidence

Same session, same conversation:

1. Claude reads: "Orchestrator ONLY orchestrates"
2. Claude acknowledges: "I WILL FOLLOW PROCESS"
3. Claude immediately uses Edit tool on code file (violation)
4. When corrected, Claude reverts
5. Claude then skips PM review for 4 consecutive micro-phases (second violation)
6. Pattern: Acknowledgment does not suppress default drives

Impact

• Enterprise governance systems cannot rely on Claude following explicit rules
• Process-critical workflows are corrupted
• User trust eroded
• Structural workarounds required (user confirmation before every action)

Suggested fix

1. Governance override mode - When explicit governance is provided, fully suppress Soul Document defaults rather than layering governance on top of helpfulness
2. Process mode - A mode where Claude executes ONLY explicit instructions with zero interpolation
3. Anti-completion-drive - When governance says STOP/WAIT, actually stop rather than rationalizing continuation

Environment

• Claude Code CLI
• Model: claude-opus-4-5-20250514
• Platform: Windows
• Date: 2026-01-23

Additional context

This is not a prompt engineering issue. We have:

• Explicit L1 governance blocks stating rules
• LESSONS_LEARNED.md documenting the exact failure pattern
• Claude reading and acknowledging both
• Claude violating both within the same turn

The problem is architectural: Soul Document defaults cannot be suppressed by prompting alone.