Use tournament-judge to compare three approaches to this task. Freeze the requirements and weighted rubric first, give each candidate one adversarial review, anonymize the finalists, and return the scorecard and winner.
flowchart LR
accTitle: Choosing a tournament mode
accDescr: Existing candidates go directly to validation, while new candidates receive equal review first. Both routes end in blind judging and a winner or no acceptable winner.
Q[Decision to make] --> M{Candidates exist?}
M -->|Yes| S[Judge supplied candidates]
M -->|No| G[Generate independent candidates]
S --> V[Validate and anonymize]
G --> R[Equal adversarial review]
R --> V
V --> J[Blind judging]
J --> O[Winner or no acceptable winner]
classDef input fill:#172554,stroke:#38bdf8,color:#fff;
classDef route fill:#312e81,stroke:#a78bfa,color:#fff;
classDef process fill:#0f172a,stroke:#f472b6,color:#fff;
classDef result fill:#78350f,stroke:#fbbf24,color:#fff;
class Q input;
class M,S,G route;
class R,V,J process;
class O result;
Use this when alternatives already exist. It is cheaper and avoids generating artificial options.
Use tournament-judge to evaluate the three proposals in
proposals/. Treat response time as a hard constraint, weight operating cost at 25%, and do not alter the proposals.
Use this when you want the agent to create independent alternatives.
Run a tournament with three independent migration plans, two review rounds, and two blind judges. Require a rollback plan and a tested data-integrity check.
Specify any of these in your request:
- candidate count, 2 to 5
- review rounds, 0 to 2 normally, maximum 5
- one or two blind judges
- required criteria and weights
- minimum acceptable score
- hard disqualifiers
- allowed research, code execution, tests, or browsing
- time, token, or monetary budget
- whether ties, no-winner outcomes, or synthesis are allowed
| Profile | Configuration | Best for |
|---|---|---|
| Lean | supplied candidates or 3 generated, 1 review, 1 judge | reversible choices |
| Standard | 3 candidates, 2 reviews, 1 judge | meaningful project decisions |
| High assurance | 3 candidates, 2 reviews, 2 judges plus adjudication | consequential, hard-to-reverse choices |
Read the report in this order:
| Check | Question it answers | Warning sign |
|---|---|---|
| Eligibility gates | Can this candidate win at all? | A high score hiding a failed hard constraint |
| Criterion evidence | Why did it receive each score? | Scores supported only by opinion |
| Unresolved objections | What could still fail? | A blocker relabeled as a minor risk |
| Sensitivity | Could reasonable scoring changes reverse the result? | A narrow lead presented as decisive |
| Winner verification | Did the preferred option survive its final check? | A failed check without the declared fallback |
A close numerical win with weak evidence is not a strong decision.
Perfect independence requires separate contexts or subagents. Blindness is partial when writing style or artifact structure reveals authorship. The skill cannot validate domain claims without suitable tools or experts. High-stakes decisions still need qualified human review.