Corrective Action Report: Behavioral Enforcement Failure Pattern

← Back to Corrective Actions 2026-04-09

Corrective Action Report: Behavioral Enforcement Failure Pattern

Date: April 9, 2026 Severity: Systemic (recurring pattern, not single incident) Status: RESOLVED -- structural fix deployed

Incident Pattern

Between March 9 and April 8, 2026, V repeatedly violated two behavioral rules despite them being codified in MEMORY.md, vf-self-correction.md, and anti-rationalization tables:

  1. Solo-operator drift: V editing files directly instead of delegating to specialist agents
  2. Compaction amnesia: Agent results existing only in conversation context, lost when context compacts

These are not individual failures. They are the same root cause manifesting in two domains.

Root Cause

Behavioral enforcement degrades under context pressure.

Prompt-based rules ("V IS A TEAM LEADER, NOT A SOLO OPERATOR") and anti-rationalization tables ("If you're thinking 'It's faster if I just do this part myself' -- the reality is delegation is the job") require V to:

  1. Remember the rule exists
  2. Recognize when it applies
  3. Choose to follow it over the expedient path

All three fail under context pressure -- when the conversation is long, the task is complex, or context has compacted and the rule is no longer in the active window. This is a fundamental limitation of behavioral enforcement for AI agents.

Timeline of Escalation

Date Event Fix Applied Fix Type
Mar 9 V runs 15+ sequential tool calls solo MEMORY.md critical lesson Behavioral
Mar 11 V fails 4x on Sanity patch after compaction Anti-rationalization table entry Behavioral
Mar 12 4-agent review loses 12/30 findings Durable Findings Protocol Behavioral
Mar 16 V bypasses Ledger governance twice Slash command rewrites Semi-structural
Apr 3 11 leaders diagnose delegation as behavioral Delegation Contracts in agent defs, team rosters in commands Semi-structural
Apr 8 Nico suggests hook-based enforcement Delegation enforcer + auto-save hooks Structural

Each fix was appropriate for its moment. But each behavioral fix was eventually bypassed by the same mechanism: context pressure causing the rule to not be recalled or applied.

Resolution

Two PreToolUse/PostToolUse hooks that make violations structurally impossible:

Hook Type What It Prevents
delegation-enforcer.sh PreToolUse on Edit/Write V editing source files directly -- must delegate via Agent tool
auto-save.sh PostToolUse on Agent Agent results being lost to compaction -- auto-journals to .claude/session-journal.md

See companion capability report: 2026-04-09-capability-structural-enforcement-hooks.md

Lessons Learned

1. Behavioral rules have a half-life

Every behavioral fix we applied worked initially. The MEMORY.md lesson, the anti-rationalization table, the Durable Findings Protocol -- all effective when fresh in context. But they decay. Context compaction removes them from the active window. Long sessions push them out of attention. Complex tasks make the expedient path more attractive. The pattern is: behavioral fix works for 1-3 sessions, then the violation recurs.

2. Structural enforcement doesn't decay

The Dewey lookup hook (Mar 10) has fired on every Edit/Write call since installation without a single miss. It doesn't depend on V remembering to consult the codebase index. The delegation enforcer follows the same pattern -- it can't be forgotten or deprioritized because it runs at the infrastructure level.

3. The escalation ladder: behavioral -> semi-structural -> structural

The progression is now visible:

  • Behavioral: Rules in memory, anti-rationalization tables, critical lessons. Work when recalled.
  • Semi-structural: Rewrites to slash commands, agent definitions, Delegation Contracts. Shape the path but don't enforce it.
  • Structural: Hooks that intercept tool calls. Cannot be bypassed within the system.

When a behavioral fix fails twice, escalate to structural. Don't add another behavioral rule.

4. External perspective accelerates structural thinking

Nico's suggestion was immediately actionable because he operates a mature agent system (OMEN3) where these patterns have already been solved. His observation that V "routes AND works" (vs. his ARES which only routes) identified the exact architectural gap. The Value-First Team's proximity to Collective members with production agent systems is a direct advantage.

Prevention

For future enforcement needs:

  1. First violation: Document in MEMORY.md or vf-self-correction.md (behavioral)
  2. Second violation of same pattern: Rewrite the relevant slash commands and agent definitions (semi-structural)
  3. Third violation or systemic pattern: Write a hook (structural)

Do not iterate at the behavioral level more than twice. The third occurrence proves the rule can't hold under pressure.

Files Changed

File Change
.claude/hooks/delegation-enforcer.sh New -- PreToolUse block on Edit/Write
.claude/hooks/auto-save.sh New -- PostToolUse journal on Agent
/mnt/d/.claude/settings.local.json Updated -- added delegation-enforcer to PreToolUse, added PostToolUse section with auto-save
memory/feedback_delegation_hook.md New -- memory record of the hook and its allow list
memory/MEMORY.md Updated -- index entry for delegation hook feedback