The Tribunal: Multi-Model AI Deliberation

Convene a tribunal of the world's best AI models to independently solve a task, exchange structured critiques, debate with evidence, and produce a consensus output — validated by impartial judges on The Bench and a Fresh Eyes reviewer, with a full audit trail.

When to Activate

Activate when the user says any of:

• /conclave [task] or /council [task]
• "get the council to..." or "have the council review..."
• "I want multiple perspectives on..."
• "debate this", "deliberate on this"
• Any explicit request for multi-model review

CRITICAL: Parse User Intent

Before doing anything, parse the user's input for:

1. TASK: What they want the council to work on

2. DEPTH (if specified): QUICK | BALANCED | THOROUGH | RIGOROUS | EXHAUSTIVE | NUCLEAR

3. TASK_TYPE: What kind of work this is:

   • CODE — "write", "build", "implement", "fix" → Models produce code
   • ARCHITECTURE — "design", "architect", "plan" → Models produce designs
   • REVIEW — "review", "audit", "check" → Models critique existing work
   • RESEARCH — "research", "compare", "evaluate" → Models investigate
   • PROMPT — "prompt", "system prompt" → Models write prompts
   • GENERAL — anything else

4. CONTEXT: Any files, code, or additional context the user provides

Auto-select depth if not specified:

• Simple question → QUICK (2 advocates, no judges)
• Standard task → BALANCED (2-3 advocates, 1 judge)
• "Best-in-class", "production-ready", important deliverable → THOROUGH (3-4 advocates, 2-3 judges)
• "Security review", "architecture decision", high-stakes → RIGOROUS (3-4 advocates, 3-4 judges)
• "Mission-critical", explicitly exhaustive → EXHAUSTIVE (4-5 advocates, 4-5 judges, stability audit, Fresh Eyes)
• "Nuclear", "maximum rigor", adversarial-grade → NUCLEAR (5 advocates, 6 judges, 7 debate rounds, mid-debate checkpoint, stability audit, Fresh Eyes)

Display your parsing:

Convening Tribunal session for: {TASK}

Parsed intent:
- TASK_TYPE = {TASK_TYPE}
- DEPTH = {DEPTH}
- ADVOCATES = {N} models generating solutions
- JUDGES = {N} impartial judges (Justices + randomly drawn Appellate/Magistrate Judges)

Estimated time: {TIME} | Estimated cost: ~${COST}
Starting deliberation now.

---

Step 1: Generate Your Own Submission (SEAL BEFORE API CALLS)

CRITICAL: You must generate your OWN solution to the task FIRST, before seeing any other model's output.

Write your submission to file:

mkdir -p ./conclave-sessions/$(date +%Y%m%d-%H%M%S)
# Write your submission to the session directory BEFORE calling any APIs

Your submission follows the Hypothesis + Evidence format:

• Hypothesis: One clear, falsifiable sentence about your approach
• Evidence: 3-5 pieces of supporting proof
• Counterargument Acknowledgment: Strongest argument against your approach
• Self-Assessment: Score your own work honestly

Save this to file before proceeding. This is the anti-anchoring measure — your work is sealed.

---

Step 2: Run the Orchestrator

# Find skill root — works in repo checkout, Claude Code, or Codex install
for dir in \
  "." \
  "${CLAUDE_PLUGIN_ROOT:-}" \
  "$HOME/.claude/skills/conclave" \
  "$HOME/.agents/skills/conclave" \
  "$HOME/.codex/skills/conclave" \
  "$HOME/.gemini/skills/conclave"; do
  [ -n "$dir" ] && [ -f "$dir/scripts/council_orchestrator.py" ] && SKILL_ROOT="$dir" && break
done

if [ -z "${SKILL_ROOT:-}" ]; then
  echo "ERROR: Could not find scripts/council_orchestrator.py" >&2
  exit 1
fi

python3 "${SKILL_ROOT}/scripts/council_orchestrator.py" \
  --briefing <briefing_file> \
  --sealed-submission <your_sealed_submission> \
  --depth {DEPTH} \
  --emit summary

Use a timeout of 600000 (10 minutes) on the Bash call. Sessions typically take 2-8 minutes depending on depth.

The orchestrator handles everything:

• Dispatches task to advocate models in parallel (via Together AI + direct APIs)
• Collects submissions with per-model timeouts
• Anonymizes all submissions (including yours) and shuffles them
• Routes anonymized briefs to judges on The Bench (Justices always seated, Appellate/Magistrate Judges randomly drawn)
• Judges evaluate, fact-check, and render verdicts
• If judges issue REMAND → advocates revise and resubmit (max 1 remand)
• Fresh Eyes validation (zero-context review of final output)
• Generates debrief report

Watch stderr for progress:

[tribunal] Session tribunal-20260228-203200 started (THOROUGH depth)
[tribunal] The Bench: Justice Qwen-3.5-397B, Justice DeepSeek-R1, Appellate Judge MiniMax-M2.5
[tribunal] Phase 2: Independent work — dispatching to 4 advocates...
[tribunal]   ✓ Claude Sonnet submitted (3.2s)
[tribunal]   ✓ GPT-5 submitted (4.8s)
[tribunal]   ✓ Gemini Pro submitted (2.9s)
[tribunal]   ✓ Host submission loaded (sealed)
[tribunal] Phase 4: Critique exchange — 12 critique pairs...
[tribunal] Phase 5: Debate round 1 of 3...
[tribunal]   Agreement score: 0.72 (target: 0.80) — continuing...
[tribunal] Phase 5: Debate round 2 of 3...
[tribunal]   Agreement score: 0.88 — consensus reached!
[tribunal] Position stability scorecard written
[tribunal] Phase 6: Judicial review — 3 judges evaluating...
[tribunal]   ✓ Justice Qwen-3.5-397B: VERDICT
[tribunal]   ✓ Justice DeepSeek-R1: VERDICT
[tribunal]   ✓ Appellate Judge MiniMax-M2.5: VERDICT
[tribunal] Phase 8: Done. Files written to ./conclave-sessions/tribunal-20260228-203200/

At NUCLEAR depth, you'll also see:

[tribunal] NUCLEAR mode: debate rounds 1-4, then judicial checkpoint, then rounds 5-7
[tribunal] Phase 5: Debate round 4 of 7 — mid-debate judicial checkpoint...
[tribunal]   Judges reviewing debate progress + stability data...
[tribunal] Phase 5: Debate round 5 of 7 (post-checkpoint)...

---

Step 3: Present Results

Read the output files and present to the user:

Output Files

The orchestrator writes to ./conclave-sessions/<session-id>/:

• final-output.md — The deliverable (what the user asked for)
• session-summary.md — Canonical 4-section summary: Question → Recommended Outcome → How We Got Here → Build This (BALANCED+ only)
• debrief.md — Situation report: how the council worked, where they agreed/disagreed
• council-log.md — Full deliberation record (all submissions, critiques, debates, verdicts)
• play-by-play.md — Narrative "sporting event" commentary of the session
• position-stability.md — Position stability scorecard (EXHAUSTIVE+ only)

Presentation Format

## Tribunal Result

[Present the final output — this is the deliverable the user asked for]

---

## Session Summary

[Present the session-summary.md — the canonical Question → Outcome → How We Got Here → Build This document. This is the reader-friendly executive summary of the entire deliberation.]

---

## Debrief Summary

[Key points from the debrief: consensus type, positions changed, judicial verdicts,
any remands issued, Fresh Eyes findings]

### Panel Composition
| Role | Model | Final Assessment |
|------|-------|-----------------| 
| Advocate | [Model] | [Summary] |
| ...  | ...   | ... |
| Justice | [Model] | [Verdict summary] |
| Appellate Judge | [Model] | [Verdict summary] |
| Fresh Eyes | [Model] | [Assessment] |

### Key Disagreements
[Any contested points with both majority and minority positions]

---
📊 Session: {session-id} | Depth: {DEPTH} | Advocates: {N} | Judges: {N}
⏱️ Duration: {time} | API Calls: {N} | Cost: ~${cost}
📁 Full logs: ./conclave-sessions/{session-id}/

### Position Stability (EXHAUSTIVE+ only)
[Summary from position-stability.md — flag any drift concerns]
---

Want me to go deeper on any point, or run another tribunal session?

---

Key Rules (NON-NEGOTIABLE)

1. Your submission must be sealed before seeing API responses. This is the anti-anchoring measure. If you read another model's output before writing yours, the entire session is compromised.

2. Never reveal model identities during deliberation. Submissions use anonymized aliases (Advocate-A through Advocate-E). Judges see ONLY anonymized briefs. Identities are revealed only in the debrief.

3. Always present minority positions alongside consensus. If a model disagreed and wasn't convinced, their position appears in the output. Minority dissent is protected, not suppressed.

4. Never skip judicial review (at BALANCED+ depth). The whole point of The Tribunal is that impartial judges evaluate the work. Skipping judges turns this into a chat room.

5. Never skip Fresh Eyes validation (at EXHAUSTIVE+ depth). The zero-context review catches groupthink artifacts that the debating models normalized. Fresh Eyes runs for EXHAUSTIVE and NUCLEAR tiers.

6. Justices are always seated. Qwen 3.5 397B and DeepSeek R1 serve on every THOROUGH+ session. They are not randomly drawn — they earned their permanent seats.

7. Present both the final output AND the debrief. The user gets the deliverable they asked for PLUS the audit trail of how it was produced.

8. Maximum 1 remand per session. If judges remand, advocates revise once. If judges still have concerns after the revision, they note the deficiency in their verdict and the session proceeds. No infinite loops.

9. Position stability is tracked at EXHAUSTIVE+ depth. Every debate response includes a 1-5 stability score. The orchestrator compiles these into a scorecard that judges use to detect sycophantic drift. Based on Kelley & Riedl (2026) research on multi-turn sycophantic convergence.

10. NUCLEAR mid-debate checkpoint is mandatory. At NUCLEAR depth, debate pauses after round 4 for a full judicial review of progress + stability data. This checkpoint cannot be skipped — it catches sycophantic drift before the highest-risk rounds (5-7).

---

Configuration

Required Environment Variables

# Required: Together AI (hosts all judge models + open-source advocates)
export TOGETHER_API_KEY="..."

# Optional: Direct APIs for frontier advocate models
export ANTHROPIC_API_KEY="..."     # Claude Opus/Sonnet advocates
export OPENAI_API_KEY="..."        # GPT-5 advocates
export GOOGLE_API_KEY="..."        # Gemini advocates

# Optional: Secondary hosting provider
export FIREWORKS_API_KEY="..."     # Alternative for specific models

# Optional: Configuration
export CONCLAVE_DEFAULT_DEPTH="THOROUGH"
export CONCLAVE_TIMEOUT="120"           # Per-model timeout in seconds
export CONCLAVE_LOG_DIR="./conclave-sessions"
export CONCLAVE_MAX_COST="5.00"         # Abort if estimated cost exceeds this

At minimum, TOGETHER_API_KEY is required — it provides access to all judge models and many advocate models through a single API. Adding ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY enables frontier models as advocates.

The Bench (Judicial Hierarchy)

Rank	Role	Models	Assignment
Justices	Permanent senior judges	Qwen 3.5 397B, DeepSeek R1	Always seated at THOROUGH+
Appellate Judges	Primary rotation pool	MiniMax M2.5, Kimi K2, GLM-4.7	Randomly drawn at BALANCED+
Magistrate Judges	Extended bench	GPT-OSS 120B, GLM-5, DeepCogito Cogito v2.1	Drawn at RIGOROUS+

Depth Levels

Depth	Advocates	Judges	Debate Rounds	Checkpoint	Stability Audit	Fresh Eyes	Est. Time	Est. Cost
QUICK	2	0	0	—	No	No	1-2 min	~$0.10
BALANCED	2-3	1	1	—	No	No	3-5 min	~$0.50
THOROUGH	3-4	2-3	3	—	No	No	8-15 min	~$2.00
RIGOROUS	3-4	3-4	5	—	No	No	15-25 min	~$5.00
EXHAUSTIVE	4-5	4-5	5	—	Yes	Yes	25-45 min	~$10.00
NUCLEAR	5	6	7	After R4	Yes	Yes	45-75 min	~$15.00

NUCLEAR depth splits debate into two halves with a judicial checkpoint after round 4. Judges review position stability data and assess whether convergence is genuine or sycophantic before rounds 5-7 proceed. This is based on Kelley & Riedl (2026) findings that sycophantic drift accelerates steeply after round 4-5.

---

Security & Permissions

What this skill does:

• Sends task briefings to AI model APIs (Anthropic, OpenAI, Google, Together AI, Fireworks AI) for advocate submissions, judicial opinions, and Fresh Eyes validation
• Writes session logs, debrief reports, and final outputs to local filesystem
• Reads local files when the user provides context (code, documents) for review

What this skill does NOT do:

• Does not post, publish, or share content externally
• Does not access your accounts on any platform
• Does not share API keys between providers
• Does not send data to any endpoint not listed above
• Does not execute generated code (advocates write code, but it's presented for human review)
• Cannot be invoked autonomously by the agent (disable-model-invocation: true)

Bundled scripts: scripts/council_orchestrator.py (main orchestration), scripts/model_client.py (unified API client), scripts/debate_manager.py (critique/debate engine), scripts/cardinal_judge.py (judicial routing), scripts/consensus_calculator.py (voting/scoring), scripts/report_generator.py (debrief generation), scripts/fresh_eyes_validator.py (zero-context validation), scripts/config_loader.py (configuration), scripts/progress.py (stderr progress display)

Review scripts before first use to verify behavior.