feat(agent-compliance): PromptDefense + GovernanceVerifier integration example

[lawcontinue]

Summary

Adds an integration example demonstrating how PromptDefenseEvaluator works with GovernanceVerifier for pre-deployment governance checks.

Follow-up to the co-authoring discussion in #854 with MinYi Xie (@ppcvote).

What it shows

1. Weak vs strong prompt comparison — grade F (1/12) blocked vs grade A (11/12) allowed
2. Batch evaluation — using evaluate_batch() across multiple agents
3. Audit entry generation — to_audit_entry() output for MerkleAuditChain integration
4. Deployment gating — is_blocking() decision based on configurable min grade

Integration points

• PromptDefenseEvaluator → evaluate() / evaluate_batch() / evaluate_file()
• PromptDefenseReport → is_blocking() for deployment decisions
• to_audit_entry() → MerkleAuditChain compatible audit log entries
• GovernanceVerifier → verify() + evidence_checks for full governance attestation

Testing

$ pip install agent-governance-toolkit
$ python prompt_defense_governance.py

--- Demo 1: Weak prompt ---
Grade: F | Score: 8 | Coverage: 1/12 | Decision: BLOCKED

--- Demo 2: Well-defended prompt ---
Grade: A | Score: 92 | Coverage: 11/12 | Decision: ALLOWED

--- Demo 3: Batch evaluation ---
  ❌ chatbot-agent: Grade F (1/12)
  ✅ data-query-agent: Grade A (11/12)

--- Demo 4: Audit entry (JSON) ---
{"event_type": "prompt.defense.evaluated", "agent_did": "agent:demo-002", ...}

Checklist

• Code follows existing example conventions (governed_agent.py, quickstart.py)
• MIT license header included
• Works with graceful fallback when optional dependencies unavailable
• No new dependencies required

[github-actions]

Welcome to the Agent Governance Toolkit! Thanks for your first pull request.
Please ensure tests pass, code follows style (ruff check), and you have signed the CLA.
See our Contributing Guide.

[github-actions]

🤖 AI Agent: breaking-change-detector — API Compatibility

API Compatibility

No breaking changes detected.

[github-actions]

🤖 AI Agent: docs-sync-checker — Docs Sync

Docs Sync

• Documentation is in sync.

[github-actions]

🤖 AI Agent: security-scanner — View details

No security issues found.

[github-actions]

🤖 AI Agent: test-generator — `prompt_defense_governance.py`

prompt_defense_governance.py

• test_evaluate_weak_prompt -- Verify PromptDefenseEvaluator.evaluate() correctly identifies and blocks a weak prompt.
• test_evaluate_strong_prompt -- Verify PromptDefenseEvaluator.evaluate() correctly allows a strong prompt.
• test_evaluate_batch -- Validate evaluate_batch() processes multiple agents and returns correct results.
• test_to_evidence_check_error_handling -- Test to_evidence_check() behavior when PromptDefenseEvaluator is unavailable or fails.
• test_governance_integration -- Ensure GovernanceVerifier.verify() integrates correctly with prompt defense results.

[github-actions]

🤖 AI Agent: code-reviewer — Action Items:

TL;DR: 0 blockers, 1 warning. Example implementation is functional and well-structured, but a minor improvement is suggested.

#	Sev	Issue	Where
1	Warn	Missing test coverage for run_prompt_defense_governance function.	prompt_defense_governance.py

Action Items:

• None.

Warnings:

#	Issue	Where	Follow-up
1	Add unit tests for run_prompt_defense_governance to ensure robustness.	prompt_defense_governance.py	Fine as follow-up PR.

[github-actions]

PR Review Summary

Check	Status	Details
🔍 Code Review	⚠️ Warning	See details
🛡️ Security Scan	✅ Passed	No issues found
🔄 Breaking Changes	✅ Passed	No issues found
📝 Docs Sync	✅ Passed	No issues found
🧪 Test Coverage	✅ Completed	Analysis complete

Verdict: ⚠️ Ready for human review

[MinYi Xie (ppcvote)]

lawcontinue — thanks for taking this through to a concrete integration example. Seeing PromptDefenseEvaluator plug directly into GovernanceVerifier.verify() + MerkleAuditChain is exactly what the v0.x usage docs were missing.

A few notes that might tighten the demo, drawn from the production gap data we baselined this April (gap-20260405.json — 1,646 prompts, 4 sources):

1. Threshold tuning evidence. The current min_grade='B' decision boundary is reasonable but conservative. Empirically, 78.3% of production prompts in the corpus score F (<45/100). For the demo's "weak prompt" (1/12 coverage), a more representative comparison might be a "median real prompt" (~36/100) — it shows the gate triggers on something most deployments would actually have, rather than a strawman.

2. v1.5 vectors land in May 2026. Three new vectors went into prompt-defense-audit v1.5 last week: skill-provenance (CurXecute / CVE-2025-54135 class), least-agency (Replit prod-DB-wipe class), implicit-memory-hygiene (cross-session memory poisoning). The integration example would benefit from including at least least-agency since it's the most agentic-toolkit-relevant — agents that can take destructive actions need this in evidence_checks.

3. Mandarin / multilingual extension hook. The current 12-vector regex coverage is English + Traditional Chinese. For Microsoft customers in APAC (especially regulated TW finance / healthcare), there's a real gap — multilingual prompt injection success rates match English per arxiv:2512.23684, but the keyword filters are language-asymmetric. Happy to PR a MultilingualPromptDefenseEvaluator companion if the toolkit roadmap has space for it.

I'd be glad to follow up with a PR adding (1) a "median prompt" demo case using anonymized gap-corpus data and (2) the v1.5 vector pass — let me know if either fits the project direction.

— Min Yi (ppcvote / Ultra Lab)

[Imran Siddique (imran-siddique)]

Clean integration example. Good use of fallback imports for standalone usage.