Blind arbitration

Blind Arbitration

Preserve independent first passes, expose only the evidence needed for one controlled challenge, and make a traceable decision against declared criteria.

Rules

• Keep one lead as the sole information router and final adjudicator.
• Define decision criteria before collecting submissions. Do not change criteria merely to favor a proposal after seeing it.
• Give agents the same neutral decision brief unless compartmentalization is an explicit constraint. Record any information withheld from an agent.
• Do not disclose another submission, identity, preferred answer, or emerging consensus during the blind phase.
• Define mutation authority before delegation. Keep submissions read-only unless the user authorized changes; never convert selection work into external writes.
• Require claims to include exact evidence, confidence, assumptions, risks, open questions, and validation run or not run.
• Use votes only for subjective preferences. Resolve factual conflicts through evidence or a discriminating test.

Workflow

1. Define The Decision Contract

State the decision, candidate scope, hard constraints, weighted criteria, evidence standard, mutation boundary, and stop condition. Choose two or more agents whose roles add distinct expertise without assigning them a preferred outcome.

2. Collect Blind Submissions

Send the decision contract to every agent independently. Ask each to return:

• proposal or conclusion
• criterion-by-criterion case
• source-backed evidence and confidence
• assumptions, tradeoffs, failure modes, and reversibility
• validation performed, missing proof, and one decisive follow-up question

Collect every first pass before relaying any submission. If agents exceed the available slots, run them in waves or sequentially while withholding all earlier answers. State the degraded capacity, but preserve blindness.

3. Build The Evidence Ledger

Normalize submissions without erasing material differences. Record each claim, evidence, confidence, assumptions, criterion score, and dissent. Remove duplicates and separate factual conflicts from different value judgments.

Reject unsupported certainty. Do not let eloquence, agent count, or proposal order stand in for evidence.

4. Cross-Examine Once

If a material conflict remains, relay only the minimum opposing claim, evidence, or question needed to test it. Do not reveal unrelated submissions or a running ranking. Give each affected agent one response and require new evidence, a revised claim, or a precise reason the evidence is unavailable.

After two dead ends on one proof path, classify the blocker and pivot to a higher-signal artifact or reversible experiment.

5. Adjudicate

Apply the declared criteria to the evidence ledger. Select, synthesize, or defer. Preserve minority findings and explain rejected proposals against the criteria. When evidence cannot distinguish finalists, prefer the cheapest reversible test instead of manufacturing certainty.

6. Close With Evidence

Report the decision contract, submissions considered, information withheld during the blind phase, decisive evidence, criterion-by-criterion result, dissent, validation gaps, and next narrow proof. If authorized implementation follows, assign it separately and keep broad validation batched.

Stop Conditions

Stop after one blind submission per agent and at most one controlled cross-examination, or earlier when decisive evidence resolves the choice. Preserve unresolved dissent. Do not use this skill for open peer conversation, causal diagnosis through rival experiments, or joint editing of one artifact.

Example

“Use $blind-arbitration to collect three independent architecture proposals, cross-examine the finalists once, and choose against these constraints.”