Business-health risk read-out (FR + EN, closed grid, official data only)

Risk read-out from official data: `synthese` (resume FR + resume_en, strengths, warning signs, missing data, verdict, activity trend, confidence level), `grille` (the reconciled evaluation grid, machine-readable codes) and `divergences_modele` (where deterministic guards overruled the model). A model only picks values from closed lists: no free text, no name and no figure of its own. WHAT TO EXPECT ON THIN FILES: entities that file no annual accounts — non-profits, sole traders, companies under a year old, ceased companies — return `verdict: "non_concluant"` with most grid fields set to `non_evaluable`; the prose then states explicitly that no accounts are available and whether that is normal at this age, and a ceased company is flagged with the `entreprise_administrativement_fermee` warning. Measured on a random sample of 47 SIREN: half of them are such files. Includes model, prompt version and cache status. Cached 7 days. Preserve coverage and provenance. Missing data is not zero; a name match is not proof of identity.

price-ascending: candidates are ordered by the listed price of the operation, with no managed-first preference applied. Unlike /v1/apis, this surface does not rank Apiosk-settled listings above federated ones — a comparison that reorders on commercial grounds is not a comparison. Pass sort=managed_first for the catalog's ordering.

Compared: listed price, settlement rail and input compatibility. Not measured yet: latency, reliability, result quality, and provider terms such as rate limits, jurisdiction and licensing.