Record contract

Every agent working inside a Spine run returns its result as one or more records. A record is a JSON object. The human never reads records directly; the synthesis agent renders them.

Shape

{
  "id": "R-0007",
  "stage": "review",
  "agent": "ecc:security-reviewer",
  "provider": "ecc:code-review",
  "claim": "The upload handler accepts arbitrary paths.",
  "evidence": [
    { "kind": "file", "ref": "services/api/upload.py:42-58" },
    { "kind": "test", "ref": "pytest tests/test_upload.py::test_traversal", "result": "fail" }
  ],
  "confidence": 0.85,
  "severity": "HIGH",
  "dissent": null,
  "resolution": null,
  "alternatives": ["Sanitize with os.path.realpath", "Reject anything containing .."]
}

Rules

1. claim is one sentence, falsifiable. Not "consider improving error handling" but "create() swallows ValueError at line 88 and returns 200".
2. evidence has at least one entry. kind is file, test, command, doc, or unverified. ref is path:line or path:start-end for files, the exact command for commands and tests, a heading or line for docs.
3. A record whose evidence is only unverified may not be HIGH or CRITICAL. If you believe it is serious, go verify it, then raise it.
4. confidence is 0 to 1 and means "probability the claim is true as stated", not "how much I care".
5. severity is CRITICAL (data loss, security, outage), HIGH (bug or spec violation), MEDIUM (maintainability), LOW (style), INFO.
6. dissent is set only by the devil's advocate or by a verifier that refutes. It is the strongest case against the claim, with evidence.
7. resolution is set when reading the repo cannot settle a disagreement: { "kind": "test" | "spike" | "question", "text": "..." }.
8. alternatives lists other approaches considered, so the human learns what was debated.
9. Records are written to <run>/records/R-NNNN.json by the spine skill. Agents that return text end with one fenced block tagged json spine-record per record. Agents called through a workflow return structured output and never write files.