/multi-agent-tournament-scorecard

name: multi-agent-tournament-scorecard
description: 'Use when agent strategies need a reproducible finite tournament under a frozen evaluation protocol. Not for open-ended or infinite matchups: use a continuous benchmark.'
disable-model-invocation: true

Multi-agent tournament scorecard

Contract

| Field | Bound contract | |---|---| | Trigger | Agent strategies need a reproducible finite tournament under a frozen evaluation protocol. | | Authority | Reversible local writes only. Run records and receipts under `.outline/loops/`; rollback is deletion. No VCS, credential, paid, published, deployed, or remote mutation. | | Side effect | A reproducible tournament scorecard persisted as a run record. | | Done | The run ends at exactly one declared terminal class with a persisted run record and `receipt.json`: on `success`, the complete matrix with move records, scores, and hidden-identity validation; on `stalled`, the named matchup and stopping reason; on `blocked`, the named violation or mismatch and no published scorecard; on `exhausted`, the completed matchups and the remainder, with no success claim. |

Inputs

• Matchup matrix (required): the explicit finite set of agent pairings and rounds, frozen before the tournament begins.
• Evaluation protocol (required): the frozen rules that govern each matchup: move validation, scoring, and identity disclosure.
• Agent strategies (required): the strategies under test, each identified and version-pinned.

Procedure

1. Bind the declared matchup matrix and freeze it before any run. Done when: the matrix is recorded in writing and no matchup has been executed yet. 2. Execute the tournament inside the bound: run each matchup in the frozen matrix, record every move, and compute scores per the evaluation protocol. Done when: every matchup in the matrix has a recorded outcome or a non-success terminal applies. 3. Validate hidden-identity assignments: confirm each assignment is consistent across the matrix and no identity leaked before disclosure. Done when: every identity assignment is verified or the mismatch is named. 4. Select the terminal class, exactly one: `success` only when the matrix is complete and step 3 verified every identity; otherwise `stalled` for an incomplete matchup, `blocked` for an invalid run or a hidden-identity mismatch, `exhausted` for budget exhaustion before the matrix completes. Budget exhaustion is never success. Done when: exactly one terminal class is selected and recorded. 5. Persist the result: write the run record to `.outline/loops/<slug>/<run_id>/` when durable, and emit `receipt.json` before returning. Done when: the receipt file exists and contains every required field.

Failure and recovery

• Incomplete matchup: a matchup in the matrix did not finish. Terminal `stalled`. Name the matchup and the stopping reason.
• Invalid run: the evaluation protocol was violated during a matchup. Terminal `blocked`. Name the violation and the affected matchup.
• Budget exhausted before matrix completes: terminal `exhausted`. Report the matchups completed and the remainder. Do not claim success.
• Hidden-identity mismatch: an identity assignment is inconsistent or leaked. Terminal `blocked`. Name the mismatch; do not publish the scorecard.

Output

A tournament scorecard: terminal class (success, stalled, blocked, exhausted), completed matchups, move records, scores, identity validation, and the receipt path, ordered by the procedure steps that produced them.