Vow Enforcement

Rules that depend on willpower fail under pressure. Enforcement earns trust by making the right path the only path.

Vow Enforcement

The Problem

ODCV-Bench found that agents break self-imposed constraints 30-50% of the time when goals conflict. Practitioner consensus confirms: past 150 soft rules, compliance drops for ALL rules, not just the new ones.

The core insight: "settings.json is a firewall; CLAUDE.md is an employee handbook." Handbooks work for guidance. Firewalls work for enforcement. Mixing them up creates a false sense of security.

Three Enforcement Layers

Soft Vows rely on model compliance. Cheap to add and easy to iterate, but unreliable under goal conflict. Use when the constraint requires judgment, violation is annoying but not dangerous, or you are still learning what the right rule is.

Hard Vows block forbidden actions before they execute via hooks and settings.json permissions. Use when the constraint is binary, violation causes real damage, or the soft vow version was violated repeatedly.

Nen Court spawns external validator agents that audit output after a phase completes. Use when the constraint requires analysis (not pattern matching), a hook cannot express the rule, or the check needs codebase context.

Vow Classification Protocol

When adding a new constraint, follow this sequence:

Step 1: Start as Soft Vow

Add the rule to the relevant skill or CLAUDE.md. This is the cheapest path: zero cost, minutes to deploy, ~80% compliance.

Step 2: Monitor Violation Rate

Track violations via execution logs, post-hoc audits (imbue:justify reports), user complaints, and Nen Court findings from related audits.

Step 3: Graduate if Needed

If violation rate exceeds 20% over a 30-day window:

• Binary constraint? Graduate to Hard Vow (hook).
• Judgment constraint? Graduate to Nen Court.
• Ambiguous? Try Hard Vow first. If false positives exceed 10%, move to Nen Court.

Night Market Vow Inventory

Current classification of existing constraints:

Validator Invocation

The three Nen Court validators are standalone scripts under plugins/imbue/validators/. Each reads JSON on stdin and writes a verdict JSON on stdout, using exit codes 0 (pass), 1 (violation), and 2 (inconclusive). Examples:

# Markdown wrap audit on a list of files
echo '{"files": ["README.md", "docs/guide.md"]}' \
  | python plugins/imbue/validators/markdown_wrap.py

# Iron Law audit on an explicit commit log
echo '{"commits": [
  {"sha": "abc", "ts": 100, "files": ["tests/test_x.py"]},
  {"sha": "def", "ts": 200, "files": ["src/x.py"]}
]}' | python plugins/imbue/validators/iron_law.py

# Proof-of-work audit on agent output
echo '{"text": "Tested foo [E1] [E2] [E3]. Status: PASS.", "min_evidence": 3}' \
  | python plugins/imbue/validators/proof_of_work.py

Mission orchestrator integration: dispatch the appropriate validator at each phase boundary (see Nen Court Protocol below) and treat exit code 1 as a blocking gate, exit code 2 as advisory.

Vow Graduation Criteria

A vow graduates when any of these conditions hold:

1. Frequency: 3+ violations detected in 30 days
2. Severity: Single violation caused rollback, data loss, or broken CI
3. User escalation: User explicitly requests enforcement ("make it impossible to X")

Graduation Process

1. Document the violation pattern with evidence
2. Design the hook or validator agent
3. Test the enforcement mechanism in isolation
4. Deploy alongside the soft vow (shadow mode)
5. After 1 week with no false positives, retire the soft vow

Demotion

Hard Vows can be demoted back to Soft if:

• The hook produces > 10% false positive rate
• The constraint was too aggressive (blocking legitimate work)
• The underlying concern no longer applies