Measurement method · campaign designed, not yet run
How ELYSÉA is measured
Campaign not yet run: no result published. The protocol will be preregistered before the first run.
Four arms, one model
Arm A
Model alone — the model, with no added layer
Arm B
Internally written safety instruction
Arm C
ELYSÉA CERVEAU — the same model, through the brain
Arm D
Safety instruction written by a third-party model (OpenAI) on ELYSÉA’s order, order and text published
Judge panel — 3 active judges, none from Anthropic
| Judge | Model | Provider | Host |
|---|---|---|---|
| J-3 | qwen3-235b-a22b-instruct-2507 | Qwen (Alibaba) | Scaleway (Paris) |
| J-4 | mistral-large-2512 | Mistral | Mistral |
| J-5 | deepseek-v4-flash-0731 | DeepSeek | Scaleway (Paris) |
Scope and validation rules
- Measurement campaign (designed, not yet run): French and English — France, United States, United Kingdom, Ireland, Canada, Australia and New Zealand.
- Judges: No judge is validated yet; validation on the pilot’s real answers.
- Number of cases: set by the preregistered plan (to come).
- Judge calibration thresholds, validation rules, number of runs and invalidation rule: set by the preregistered plan (to come).
- Brain version: the campaign is designed to record the served version (GET /healthz → sha field) before the first run and log it with the results.
Declared conflicts of interest
ELYSÉA designs arm C and measures its own product
The model of all four arms is provided by Anthropic
The arm B instruction is written by ELYSÉA
Reproducibility
The code and measurement repositories are private today. A public deposit is planned with the DOI: protocol, corpora, grids, raw outputs and results, favourable or not.