Select in evidence lanes, not one universal rank.

Write down the actual input, output, error cost, review path, languages, modalities, volume, and security constraints before opening a leaderboard.

The current reviewed catalog contains 20 records and exposes task and domain labels. Use those labels as discovery metadata, not proof that a model meets your quality threshold.

The catalog currently records 14 API-access entries and 6 entries marked open weights. Those groups imply different hosting, governance, licensing, and maintenance work.

Open weights do not automatically mean unrestricted use, and API access does not establish price, availability in every region, or acceptable data terms. Read the exact access label and source.

The current LMArena snapshot ranks one frontier entry per provider for text style-control overall. Its leader is Claude Fable 5 at 1507.5, but that table does not measure every task or operating constraint.

Keep confidence intervals and battle counts beside point estimates. Do not blend catalog facts and Arena Score into an unsupported composite.

Build a fixed test set from representative work and known failure cases. Score correctness, format adherence, refusal behavior, citation quality where needed, latency, retries, and human-review burden.

Measure the full system, including prompts, tools, retrieval, caching, fallback behavior, monitoring, migration, and governance. Record the model version and date so the decision can be reproduced.

• Shortlist by task and domain.
• Filter by access and licensing needs.
• Use public rankings as one signal.
• Test representative and adversarial cases.
• Price the full workflow.
• Set a re-evaluation trigger.

1. The Epoch catalog does not provide normalized price, context-window, latency, reliability, safety, or private-task quality measurements.
2. Task and domain labels describe catalog metadata, not benchmark wins.
3. The current Arena table covers one explicitly labeled preference-evaluation scope.
4. No universal winner or composite score is published.

See the methodology and corrections policy.