Test type · implemented
Model
A single model, one prompt, a structured output contract. No browsing, no retrieval, no tools — the model alone against the frozen evidence.
Adapter capabilities
OpenRouter catalogue discovery with fixture fallback for offline CI
One eligible flagship per provider — minis, aliases, and specialists excluded with recorded reasons
Roster freeze with exact canonical IDs, pricing, and context windows
Execution with JSON-schema response format and temperature 0
One schema-repair retry on malformed output, billed separately
Usage extraction and cost computation with explicit cost-basis labels
Currently running: The Flagship Fight
All test types