ARarchangelhealth
asclepius

Grade Real Models

"Grade the real models" mode (FEAT-1 / PRD §A): swap the task's A/B pair to two real frontier answers (blinded, source='baseline', truly-random slots) via the assembly ladder (§A3 + §A7):

  • two-frontier (one OpenAI + one Anthropic) is the strong default;
  • on a genuine single-provider failure it reverts to the OLD Anthropic-only method so annotation continues (tagged legacy_fallback, counted);
  • a sustained high fallback rate is treated as an incident → 503 needs_baseline
    • admin alert, never mostly-legacy data shipped silently;
  • V4 real cases stay Anthropic-only unless ASCLEPIUS_TWO_FRONTIER_V4 is on (§A7);
  • NEVER a gold stand-in for a failed model answer.

On submission a per-model failure record is computed.

post/api/asclepius/tasks/{task_id}/grade-real-models

Request

  • Base URL
    https://app.archangelhealth.ai (the document declares no server; this is the origin it was published from)
  • URL
    https://app.archangelhealth.ai/api/asclepius/tasks/{task_id}/grade-real-models
  • Auth
    none declared

Path parameters

task_idstring required

Headers

authorizationstring nullable

Response

Successful Response

unknown required

Changes

No recorded changes to this operation across all 1 revision of this API.