{
  "schema": "vedokrok.public-item.v1",
  "release_id": "MHC-RPUB-20260920-75ad787a",
  "url": "/knowledge/grade-the-final-state-and-the-agent-trajectory-separately",
  "id": "MHC-D-RESEARCH-0721",
  "version": "0.1.0",
  "title": "Grade the final state and the agent trajectory separately",
  "summary": "The destination can be correct while the route is unacceptable.",
  "kind": "protocol",
  "body": "Use outcome graders for the resulting files, data or decision, and separate trajectory checks for properties that matter during execution: unauthorized tool calls, destructive retries, excessive loops, fabricated sources or other process constraints. Do not grade every stylistic path when only the end state matters; trajectory rules should protect real risks or costs.",
  "limits": [
    "Overly prescriptive trajectory graders can punish creative valid solutions; inspect the path only where the path itself matters."
  ],
  "topics": [
    "union-ai-eval-learning-workflow"
  ],
  "intents": [],
  "source_ids": [
    "RS-ED3CA2F877BE88CA"
  ],
  "evidence": [
    {
      "claim": "Agent evals may grade both end-state outcomes and transcript or trajectory properties because a correct-looking final artifact can hide undesirable process behavior.",
      "source_id": "RS-ED3CA2F877BE88CA",
      "role": "supports",
      "note": "Do not overconstrain legitimate alternative paths when only the end state matters.",
      "locator": "Graders can evaluate outcome or transcript"
    }
  ],
  "use_when": [
    "An agent can reach the right answer through risky, wasteful or policy-breaking intermediate actions."
  ],
  "avoid_when": [
    "Overly prescriptive trajectory graders can punish creative valid solutions; inspect the path only where the path itself matters."
  ],
  "example": "A deployment agent must produce the correct configuration and must not bypass the approval gate on the way there.",
  "check": "A run can pass the result while still failing a meaningful process constraint, and the reports distinguish those failures.",
  "steps": [
    "A run can pass the result while still failing a meaningful process constraint, and the reports distinguish those failures."
  ],
  "sources": [
    {
      "id": "RS-ED3CA2F877BE88CA",
      "title": "Demystifying evals for AI agents",
      "url": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents"
    }
  ],
  "relations": [
    {
      "from": "MHC-D-RESEARCH-0721",
      "to": "MHC-D-RESEARCH-0959",
      "type": "useful_with",
      "url": "/knowledge/monitor-the-whole-agent-trajectory-not-only-individual-allowed-actions"
    }
  ],
  "collections": [
    {
      "id": "RC-ABC17A288CB87C85",
      "title": "Make AI work improve from failures instead of accumulating rituals",
      "url": "/collections/make-ai-work-improve-from-failures-instead-of-accumulating-rituals"
    }
  ]
}