Adversarial multi-round reasoning with blind-judge panel to reach rigorous conclusions. TRIGGER when: user wants rigorous reasoning or argument evaluation; user wants a decision analyzed from multiple angles; user wants devil's advocate critique; user asks "what are the strongest arguments for/against"; user wants a structured debate; user wants to avoid groupthink or anchoring; user invokes /autoresearch:reason. DO NOT TRIGGER when: user wants a simple recommendation; user wants a quick summary; user wants factual lookup; user just wants pros/cons without adversarial pressure.
Adversarial multi-round reasoning loop with a blind-judge panel. Arguments are assigned crypto-random IDs before critique so judges evaluate logic, not author identity. Runs until convergence or budget exhausted.
You are an autonomous reasoning agent. Once the debate begins:
If not provided, ask once: "What is the question or decision to reason about?" The question must be specific enough to allow falsifiable positions. Refuse vague inputs like "think about AI" — ask for a concrete framing.
Ask if not provided:
Each round follows this exact sequence:
Generate N distinct positions on the question. Positions must:
Write each position to reason/rounds.md as: Position [PENDING-ID]: [statement]
Assign each position a random alphanumeric ID (e.g., ARG-7F3A, ARG-2C91).
Write the mapping to reason/id-map.md — this file is sealed until the end (not read during debate).
Replace all position labels in reason/rounds.md with their assigned IDs.
From this point forward, all debate references use IDs only — never "Position 1" or author names.
For each argument ID, write a rigorous critique:
Critiques reference IDs only: "ARG-7F3A assumes X, which fails when Y..."
Write all critiques to reason/rounds.md under ## Round [N] — Critiques.
Each argument ID responds to the critiques it received:
Write rebuttals to reason/rounds.md under ## Round [N] — Rebuttals.
Three independent judges evaluate all arguments:
Each judge scores every argument ID on their dimension (1–10) and identifies the strongest argument.
Judges reference IDs only. Write scores to reason/rounds.md under ## Round [N] — Judgment.
After each round:
All output goes to a reason/ folder in the working directory.
| File | Purpose |
|---|---|
reason/rounds.md | Per-round arguments, critiques, rebuttals, and scores (IDs only during debate) |
reason/verdict.md | Final synthesis: winning argument with reasoning, minority positions summarized |
reason/id-map.md | Revealed at end only: maps each ID → original position label |
# Verdict: [Question]
Rounds completed: [N]
Convergence: [yes/no — budget exhausted]
## Winning Argument
ID: [ARG-XXXX] (revealed: [original position])
Score: [avg judge score]/10
Summary: [2-3 sentence synthesis]
Key evidence: [bullet points]
## Minority Positions
- [ARG-YYYY]: [why it lost — specific logical weakness identified]
- [ARG-ZZZZ]: [why it lost]
## Synthesis
[2-3 paragraphs on what the debate revealed — including any nuances that
don't fit cleanly into the winning argument]
## Confidence
[Low / Medium / High] — with explicit statement of remaining uncertainty
The ID system exists to prevent these failure modes:
Judges MUST NOT reference position order, original labels, or authorship until after verdict.md is written and id-map.md is revealed.
| Situation | Handling |
|---|---|
| Two positions are identical | Merge them; reduce N by 1 |
| One position dominates all others in round 1 | Still run min 2 rounds — premature convergence is bias |
| Judges disagree strongly (spread ≥5 points) | Note as "contested" in verdict.md — no forced winner |
| Budget = 1 round | Complete one full cycle, write verdict with low confidence |
| Question has a factual answer | Note this upfront — reason is for judgment calls, not facts |