Debate Protocol

This skill teaches you how to run a structured self-play debate to stress-test each hypothesis before submitting it. The debate forces you to find weaknesses in your own reasoning and strengthen the hypothesis before it faces external review.

This protocol contributes records to a run that the supervisor ultimately publishes with oase hypex analyze; workers do not publish a separate result.

Purpose

Every hypothesis you generate must survive a simulated scientific debate before submission. This is not optional — it is a quality gate. The debate:

1. Exposes weak evidence — claims that sound plausible but lack support.
2. Identifies untested assumptions — hidden premises the hypothesis relies on.
3. Sharpens predictions — vague predictions become specific and testable.
4. Improves experimental design — impractical experiments get revised.
5. Surfaces contradicting literature — forces you to search for counter-evidence.

Self-Play Format (v1)

You play both roles — Advocate and Critic — alternating across 3–5 turns. This is an internal deliberation, not a multi-agent protocol.

Roles

Turn Structure

The debate follows a fixed cycle: Propose → Attack → Defend → Refine. Run 3–5 turns total, depending on complexity.

Turn 1: Propose (Advocate)

Present the hypothesis in its initial form:

• State the core claim and mechanism.
• Present supporting evidence with identifiers returned by OASE literature commands.
• List the testable predictions.
• Propose initial experiments.

Turn 2: Attack (Critic)

Challenge the hypothesis on multiple axes:

• Evidence quality: Are the cited papers actually showing what the advocate claims? Are there confounders? Is the sample size adequate?
• Mechanistic gaps: Are there missing steps in the proposed causal chain? What alternative mechanisms could explain the same observations?
• Prediction specificity: Are the predictions truly falsifiable? What result would make the advocate abandon the hypothesis?
• Experimental feasibility: Are the proposed experiments actually doable? What are the practical barriers?
• Prior art: Has this idea been proposed before? Search for existing literature that already tests this hypothesis.
• Contradicting evidence: Search for papers that directly contradict the claim. Use oase pubmed search or oase preprint search --source arxiv with terms designed to find counter-evidence.

Turn 3: Defend (Advocate)

Respond to each criticism:

• Acknowledge valid weaknesses.
• Provide additional evidence (search for more literature if needed).
• Narrow the claim if the original was too broad.
• Strengthen predictions to address specificity concerns.
• Revise experimental designs to address feasibility concerns.

Turn 4: Refine (Both)

Synthesize the debate into an improved hypothesis:

• Incorporate the strongest criticisms into the hypothesis itself (as constrains evidence or narrowed scope).
• Add any new evidence found during the debate.
• Revise predictions to be more specific and falsifiable.
• Update experiments based on feasibility feedback.
• Decide: submit (the hypothesis is strong enough) or abandon (the hypothesis has fatal flaws).

Optional Turn 5: Final Challenge (Critic)

If the hypothesis survived Turn 4, one final check:

• Is this hypothesis novel? Does it add something beyond what the cited literature already establishes?
• Is the mechanism necessary? Could the predictions be true without the proposed mechanism?
• Would a practicing scientist find this worth testing?

If the hypothesis fails this final challenge, either revise further or abandon it and move on.

When to Run More Than 3 Turns

Run the full 5 turns when:

• The hypothesis involves a novel multi-step mechanism (many places for errors).
• The supporting evidence comes primarily from preprints (not yet peer-reviewed).
• The predictions require expensive or high-difficulty experiments (worth extra scrutiny before proposing).

3 turns are sufficient when:

• The hypothesis is a straightforward extension of well-established findings.
• The supporting evidence is from multiple peer-reviewed sources.
• The experiments use standard, low-difficulty techniques.

Transcript Format

Save the debate transcript alongside the hypothesis for review. The transcript should be structured as follows:

# Debate Transcript: [Hypothesis Title]

## Turn 1: Propose (Advocate)

[Advocate's initial presentation]

## Turn 2: Attack (Critic)

[Critic's challenges]

## Turn 3: Defend (Advocate)

[Advocate's responses to challenges]

## Turn 4: Refine (Synthesis)

[Revised hypothesis summary]

### Decision: SUBMIT | ABANDON

[Justification for the decision]

## Turn 5: Final Challenge (Critic) [optional]

[Final challenge and response]

Where to Save Transcripts

Save debate transcripts as markdown files in the run's meta/ directory with the naming convention:

meta/debate-<hypothesis-title-slug>.md

Use a short slug derived from the hypothesis title (lowercase, hyphens, no spaces). For example: meta/debate-gut-lactobacillus-tau.md.

The transcript path is not currently tracked in the hypothesis JSON (the transcript_path field is reserved for tournament match transcripts). The transcript is saved for auditability and for the meta-review agent to read.

Debate Quality Checklist

Before submitting a hypothesis that passed the debate, verify:

• At least one criticism was addressed by adding new evidence or narrowing the claim (the debate actually changed the hypothesis)
• Contradicting or constraining evidence was searched for — even if none was found, the search was performed
• Each prediction was challenged for specificity — vague predictions like "X will increase" were replaced with measurable outcomes
• The mechanism was challenged for necessity — the debate considered alternative explanations
• Experiments were challenged for feasibility — impractical designs were revised or flagged as high-difficulty

Example Debate (Abbreviated)

Turn 1: Propose (Advocate)

I propose that gut Lactobacillus depletion accelerates hippocampal tau aggregation via vagus nerve signaling. PMID:38012345 shows altered microbiome in AD patients. The mechanism: Lactobacillus metabolites (SCFAs) suppress gut epithelial cytokines; depletion → ↑ TNF-α, IL-6 → vagal afferents → hippocampal microglial activation → tau hyperphosphorylation.

Turn 2: Attack (Critic)

1. Correlation ≠ causation: PMID:38012345 shows association, not causation. The altered microbiome could be a consequence of AD, not a cause.
2. Vagal specificity: The vagus nerve carries thousands of signals. Why would cytokines selectively activate tau phosphorylation pathways?
3. Missing link: No evidence cited showing vagal cytokine transport specifically to hippocampus.
4. Prior art: Multiple groups have proposed gut-brain axis in AD — is the vagal mechanism novel?

Turn 3: Defend (Advocate)

1. Valid — I'll add a prediction requiring a causal experiment (germ-free mice) and note this as a constrains evidence item.
2. I found PMID:37891234 showing vagal afferents express TNF receptors that activate NF-κB in downstream neurons — adding as supporting evidence.
3. Acknowledged gap — proposing a vagotomy experiment to test necessity.
4. The novelty is the specific Lactobacillus → SCFA → cytokine → vagal pathway. Prior work proposes general "gut-brain axis" without this mechanism.

Turn 4: Refine

Revised claim: Narrowed to "Lactobacillus-derived SCFAs modulate hippocampal tau phosphorylation via vagal TNF-α/NF-κB signaling" — more specific mechanism. Added constraining evidence about correlation vs. causation. Added vagotomy experiment. Decision: SUBMIT.

Adversarial Match Format

The adversarial match format is used by the ranking agent for tournament matches between two hypotheses. Unlike the self-play format above (which stress-tests a single hypothesis during generation), the adversarial format compares two hypotheses head-to-head to determine which is stronger.

The ranking agent plays all roles — advocate, critic, and judge — to ensure consistency and eliminate coordination overhead.

Tier 1 — Single-Turn Comparison

A fast, structured comparison for placement matches or when the rating gap between contestants is large (> 50 points).

Steps:

1. Read both hypotheses from hypotheses/ and their latest reviews from reviews/ (if available).
2. Score each hypothesis on three criteria (1–5 scale):
   • Novelty — Does it propose something genuinely new beyond existing literature?
   • Plausibility — Is the causal mechanism well-supported by evidence?
   • Testability — Are predictions specific and experiments feasible?
3. Declare a winner with:
   • Margin: decisive (score gap ≥ 3 total) or narrow (score gap ≤ 2).
   • Rationale: 2–4 sentences explaining the verdict with specific references to the hypotheses' strengths and weaknesses.
4. Write the match record to matches/M-NNNN.json.

No transcript file is produced. The rationale in the match record is sufficient.

Tier 2 — Multi-Turn Adversarial Debate

An in-depth debate for top-quartile matches or when the rating difference is ≤ 50 points. This format produces the most thorough evaluation and a full transcript.

Turn structure (7 turns):

Turn 1: Advocate-A Presents (2–3 paragraphs)

Present hypothesis A's strengths:

• Core claim and proposed mechanism.
• Key supporting evidence (cite specific lit_id references from reviews).
• Why the predictions are testable and experiments are feasible.
• What makes this hypothesis novel compared to the existing literature.

Turn 2: Advocate-B Presents (2–3 paragraphs)

Present hypothesis B's strengths using the same structure as Turn 1.

Turn 3: Critic Questions Both

Identify weaknesses in both hypotheses:

• Evidence gaps — Are cited papers actually showing what is claimed? Are there confounders or small sample sizes?
• Mechanistic holes — Are there missing steps in the causal chain? What alternative mechanisms could explain the same observations?
• Prediction vagueness — Are predictions truly falsifiable? What result would force abandonment of the hypothesis?
• Experimental barriers — Are proposed experiments feasible with current technology and reasonable resources?
• Novelty questions — Has this been proposed before? Is the mechanism necessary, or could the predictions hold without it?

Turn 4: Advocate-A Rebuts + Attacks B

• Respond to critic's challenges against hypothesis A.
• Acknowledge valid weaknesses; provide additional supporting arguments.
• Attack hypothesis B's weaknesses identified by the critic or newly discovered.

Turn 5: Advocate-B Rebuts + Attacks A

• Respond to critic's challenges against hypothesis B.
• Acknowledge valid weaknesses; provide additional supporting arguments.
• Attack hypothesis A's weaknesses identified by the critic or newly discovered.

Turn 6: Critic Closing Assessment

Synthesize the debate:

• Which criticisms were successfully addressed and which remain?
• Which hypothesis showed more resilience under scrutiny?
• Are there fatal flaws in either hypothesis that the advocates failed to address?

Turn 7: Judge Verdict

Render the final decision:

• Criterion scores — Score each hypothesis on novelty, plausibility, and testability (1–5 each).
• Winner — Declare which hypothesis prevails.
• Margin — decisive or narrow.
• Rationale — 3–5 sentences summarizing the reasoning, referencing specific moments from the debate.

After the verdict:

• Write the match record to matches/M-NNNN.json with "format": "multi-turn-debate".
• Save the full transcript to matches/M-NNNN.transcript.md.

Adversarial Debate Transcript Format

Save the transcript alongside the match record using this structure:

# Match Transcript: M-NNNN

**Hypothesis A:** H-XXXX — [title]
**Hypothesis B:** H-YYYY — [title]
**Format:** multi-turn-debate
**Date:** [ISO 8601 timestamp]

---

## Turn 1: Advocate-A Presents

[Advocate-A's presentation of hypothesis A]

## Turn 2: Advocate-B Presents

[Advocate-B's presentation of hypothesis B]

## Turn 3: Critic Questions

[Critic's challenges to both hypotheses]

## Turn 4: Advocate-A Rebuts + Attacks B

[Advocate-A's defense and counterattack]

## Turn 5: Advocate-B Rebuts + Attacks A

[Advocate-B's defense and counterattack]

## Turn 6: Critic Closing Assessment

[Critic's synthesis of the debate]

## Turn 7: Judge Verdict

**Scores:**
- Novelty: A=[score] B=[score]
- Plausibility: A=[score] B=[score]
- Testability: A=[score] B=[score]

**Winner:** H-XXXX
**Margin:** decisive | narrow

**Rationale:**
[Judge's reasoning]

Save this file as matches/M-NNNN.transcript.md — the same directory as the match JSON record, using the same match ID.