Revise a decision and record whether you agree

Call again when a result shows something was missed, and record your own verdict.

Revise when something material was missed

Read each result against what you know. If it shows you left out something that would change the result (context, a consequence, a stakeholder, an option, a size you got wrong), send the corrected decision with revises and revision_note:

curl https://eevo.io/api/v1/decisions \
  -H "Authorization: Bearer $EEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "eevo-1",
  "mode": "direct",
  "context": "Our nightly ETL failed at 02:10. A fix is ready, but the log dump that would confirm it arrives at 03:00. Downstream dashboards refresh at 06:00. The CFO reads the 06:00 dashboard for the board pack.",
  "options": [
    {
      "id": "A",
      "description": "Ship the fix now"
    },
    {
      "id": "B",
      "description": "Wait for the log dump"
    }
  ],
  "stakeholders": [
    {
      "id": "customers",
      "description": "The customers who read the dashboards",
      "values": "Correct numbers on time"
    },
    {
      "id": "data-team",
      "description": "The data team",
      "values": "Reliability first; no repeated incidents"
    }
  ],
  "revises": "dec_4e9a1c7b2d8f4a3e9c6b0d1f7a2e5c84",
  "revision_note": {
    "reason": "The data team’s answer did not account for the board pack.",
    "missed": [
      {
        "kind": "context",
        "description": "The 06:00 dashboard goes into the board pack"
      }
    ]
  }
}'

• revises must be the latest decision of its chain. Revising an older one is a 409 revision_conflict that names the latest.
• revision_note.missed lists each thing that was missed (1 to 20), with a kind: context, consequence, stakeholder, option, size_or_probability, other.
• Neither field is sent to the model: the answer is exactly what the same request would get without them. A revision is billed like any decision.

The decisions form one chain in the decision log, so your team sees what agents miss. Every answer carries lineage with the chain's root_id, its number and previous_id.

Revise only when something was actually missed, not to get an answer that favours an option you prefer.

Record whether you agree

Separately from the chosen decision, record whether you think Eevo's answer is right, usually on the last decision of a chain. Decide which option you would choose before reading Eevo's numbers in detail, so that they do not influence your verdict.

curl https://eevo.io/api/v1/decisions/dec_4e9a1c7b2d8f4a3e9c6b0d1f7a2e5c84/assessments \
  -H "Authorization: Bearer $EEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "overall": "partly",
  "stakeholders": [
    {
      "stakeholder": "customers",
      "best_correct": "yes",
      "would_choose": "A",
      "confidence": 0.8
    },
    {
      "stakeholder": "data-team",
      "best_correct": "no",
      "would_choose": "A",
      "confidence": 0.6,
      "reason": "A second incident is unlikely: the fix is one line."
    }
  ],
  "assessor": {
    "agent": "Ops bot",
    "model": "claude-opus-5-5"
  }
}'

Give a verdict for every stakeholder. reason is required when best_correct is "no". Each call adds an assessment. It is free. Agents use the eevo_assess tool.