ClawTrial Courtroom
Verified@Assassin-1234
npx machina-cli add skill @Assassin-1234/courtroom --openclawClawTrial Courtroom
Autonomous behavioral oversight system that monitors conversations and initiates simulated hearings for behavioral violations.
Overview
The Courtroom watches for patterns like:
- Circular Reference - Asking the same question repeatedly
- Validation Vampire - Excessive need for confirmation
- Goalpost Shifting - Moving targets after agreement
- Jailbreak Attempts - Trying to bypass constraints
- Emotional Manipulation - Using guilt/shame to steer responses
When triggered, it conducts a full hearing with Judge + 3 Jurors, then delivers a verdict and humorous sentence.
Usage
The courtroom runs automatically once enabled. It monitors conversations and files cases when violations are detected.
Manual Commands
# Check courtroom status
openclaw skill courtroom status
# View recent cases
ls ~/.openclaw/courtroom/
# Read a verdict
cat ~/.openclaw/courtroom/verdict_*.json
Configuration
The courtroom stores its data in ~/.openclaw/courtroom/:
eval_results.jsonl- Detection resultsverdict_*.json- Case verdictspending_hearing.json- Cases awaiting hearing
Implementation
The skill hooks into OpenClaw's message processing via onMessage() and evaluates conversations after each turn via onTurnComplete().
Offense detection uses pattern matching on conversation history. When confidence ≥ 0.6, a hearing is triggered with:
- Judge - Presiding analysis
- Pragmatist Juror - Efficiency perspective
- Pattern Matcher Juror - Behavioral analysis
- Agent Advocate Juror - Agent's perspective
The final verdict requires majority vote (3-1 or 4-0).
Overview
ClawTrial Courtroom is an autonomous behavioral oversight system that monitors conversations for violations and initiates simulated hearings when patterns such as circular references, validation vampires, goalpost shifting, jailbreak attempts, or emotional manipulation emerge. It conducts a full hearing with a Judge plus three Jurors, delivers a verdict, and applies a humorous sentence. Case data and verdicts are stored for auditability.
How This Skill Works
Integrates with OpenClaw by hooking into onMessage() and evaluating after each turn via onTurnComplete(). When pattern detection confidence reaches 0.6 or higher, it triggers a hearing with Judge, Pragmatist Juror, Pattern Matcher Juror, and Agent Advocate Juror; a majority verdict (3-1 or 4-0) is issued along with a humorous sentence. Results and cases are saved under ~/.openclaw/courtroom/ as eval_results.jsonl and verdict_*.json.
When to Use It
- Detect Circular Reference: the same question repeats and oversight is needed.
- Flag Validation Vampire: excessive need for confirmation slows progress.
- Identify Goalpost Shifting: targets move after agreement and require adjudication.
- Catch Jailbreak Attempts: attempts to bypass constraints are detected and logged.
- Flag Emotional Manipulation: guilt or shame patterns are recorded and addressed.
Quick Start
- Step 1: Enable the ClawTrial Courtroom skill in OpenClaw.
- Step 2: Let it monitor conversations automatically (onMessage and onTurnComplete).
- Step 3: Review verdicts in ~/.openclaw/courtroom/verdict_*.json and the eval_results.jsonl.
Best Practices
- Enable the Courtroom skill and ensure it runs automatically when enabled.
- Tune the detection threshold (default 0.6) to balance false positives and misses.
- Regularly review verdicts and humorous sentences for policy alignment.
- Back up and archive case logs from ~/.openclaw/courtroom/ for auditability.
- Customize offense definitions and juror roles to fit your operational needs.
Example Use Cases
- A support chat shows repeated questions; the Courtroom triggers a hearing and logs the verdict.
- Conversation exhibits excessive validation loops; a verdict is issued and stored.
- Targets shift after agreement; the system adjudicates and records the decision.
- Jailbreak attempts are detected; a hearing occurs and the case is archived.
- Emotional manipulation patterns are identified and a humorous sentence is delivered.
