Spec 0.1 · English storefront

The ScoreIA protocol

ScoreIA is an open protocol for scoring AI models. A ScoreIA card describes a model, a named context and a versioned suite. It is neither a universal ranking nor a prediction of citations in ChatGPT, Gemini or Perplexity.

French remains available at /protocole/. Legal French pages stay authoritative. This English page is the unmatched-language fallback (x-default).

One observation

An observation is 1 trial × 1 model × 1 seed. Without a deterministic oracle (exact, regex or maze), it is not a trial. The public boards show only cards already written by the local lab. Visiting this site does not call a model API.

API vs product plan

Grok on the xAI API is not Grok in Grok Build. Composer in Cursor is not an OpenAI model id. A subscription UI is a product channel. The lab records a transcript (--adapter forfait) and the same deterministic oracle judges it. No scraping of the product, no live call from scoreia.ai. Until that file exists, the board cell stays empty.

Empty cells

No run = not measured. We do not invent a 50/100 to fill a cell. Local snapshots and cloud APIs are not mixed on the same row. Access to Grok, Claude or Composer is not a card: a card exists only after a lab run.

Boards, not a throne

Each domain has its own competition: throughput, French, instruction, code, honesty, play, timed chamber, footprint, cost, MCP. There is no single “best LLM 2026” number. The Ring and the timed chamber are visual trials. A K.O. does not write the board.

JSON for agents

HTML is for humans. Agents should read:

There is no who_is_the_best endpoint. An agent reads a card. It does not elect a king.

What ScoreIA is not

Open the boards