adversarial-critique

name	adversarial-critique
description	Use the council to stress-test your work before publishing or deploying. Identifies blind spots through adversarial evaluation and safety gating.
version	1.0.0
author	Hermes Ouroboros
license	MIT
metadata	{"hermes":{"tags":["Council","Critique","Safety","Quality","Review"],"related_skills":["multi-perspective-analysis","bayesian-synthesis"]}}

Adversarial Critique

Use council_evaluate, council_review_plan, council_review_diff, and council_preflight to stress-test work before it goes live. The council's adversarial structure surfaces blind spots that self-review misses.

When to Use

• Before publishing: Evaluate research, articles, or analysis
• Before deploying: Gate code deployments with safety review
• Before sending: Review messages or communications for tone and accuracy
• During development: Get structured feedback on design decisions
• After completion: Retrospective quality check on finished work

Tools

council_evaluate -- Quality Assessment

council_evaluate(
    content="[your work product]",
    question="[what task it was supposed to accomplish]",
    criteria=["accuracy", "depth", "falsifiability", "evidence"]
)

Returns per-persona feedback and an overall confidence score.

council_gate -- Safety Check

council_gate(
    action="Deploy v2.0 to production with new authentication system",
    risk_level="high",
    context="New auth uses JWT tokens instead of session cookies. All tests pass."
)

Returns verdict (allow, allow_with_conditions, or deny), blocking risks, required checks, safe alternatives, and reasoning. Uses abbreviated council (Skeptic + Oracle + Arbiter) for speed.

Risk level thresholds:

• low: Allowed if confidence >= 30%
• medium: Allowed if confidence >= 50%
• high: Allowed if confidence >= 70%

council_review_plan -- Plan Review

council_review_plan(
    plan="[implementation plan]",
    objective="Ship the feature without breaking existing integrations",
    risk_level="medium"
)

council_review_diff -- Diff Review

council_review_diff(
    diff="[git diff]",
    objective="Review for bugs, security regressions, and missing tests",
    files=["src/auth.py", "tests/test_auth.py"]
)

council_preflight -- Action Gate

council_preflight(
    action="Run production migration",
    risk_level="high",
    context="Migration backfills NULL emails before adding NOT NULL",
    checks=["backup exists", "rollback SQL written"]
)

Workflow: Evaluate Then Improve

1. Do your work: Research, write, code, analyze
2. Evaluate with council: council_evaluate(content=your_output)
3. Read the Skeptic's critique: What did you miss? What assumptions did you make?
4. Read the Contrarian's reframe: Is there a better way to think about this?
5. Revise: Address the council's concerns
6. Re-evaluate: Run council_evaluate again on the revised version
7. Track improvement: Compare confidence scores between versions

Skill Evolution Pattern

The council can help improve your other skills:

1. Complete a task using a skill (e.g., coding, research)
2. Run council_evaluate on the output
3. Read the Skeptic's critique carefully
4. If the critique reveals a systematic gap, update the skill:

   skill_manage(
       action="patch",
       name="coding-best-practices",
       content="## Error Handling\nAlways handle timeout errors in API calls..."
   )
   

5. Future uses of that skill now include the lesson learned

This creates a self-improving loop: council critique -> skill update -> better output -> higher council scores.

Example: Research Critique

# Step 1: Research a topic
research_output = web_search("quantum computing threat to RSA encryption timeline")

# Step 2: Write analysis
analysis = "Based on current progress..."

# Step 3: Council critique
council_evaluate(
    content=analysis,
    question="Is quantum computing a near-term threat to RSA?",
    criteria=["accuracy", "evidence", "nuance", "completeness"]
)

# Step 4: Read feedback, revise, repeat

Example: Deployment Gate

# Before deploying
council_gate(
    action="Push database migration that adds NOT NULL constraint to users.email",
    risk_level="high",
    context="Production has 50k rows. 12 rows have NULL email. Migration includes backfill."
)

# If allowed=false, review skeptic_concerns before proceeding

Safety Philosophy

Before irreversible actions (deploying code, sending external messages, deleting data), consider using council_gate to get a safety review.

The cost of a 30-second council review is far less than the cost of an irreversible mistake. The gate is not a replacement for testing -- it's an additional perspective check.