Intent correlation claim contract

Use this contract to reason over a privacy-safe evidence packet without promoting model interpretation to source truth.

Evidence packet

IntentCorrelationPacketV1 has these top-level fields:

{
  "kind": "IntentCorrelationPacketV1",
  "schemaVersion": 1,
  "packetDigest": "sha256:<canonical digest>",
  "workspace": { "label": "repository label" },
  "inputs": [],
  "executionSlices": [],
  "files": [],
  "changeUnits": [],
  "commits": [],
  "artifacts": [],
  "validations": [],
  "observedEdges": [],
  "allowedRefs": [],
  "limitations": []
}

All FileEvidence paths are repository-relative and all usable evidence is named in allowedRefs. An ExecutionSlice begins with one retained user input and ends before the next retained input in that Session. A ChangeUnit is either a verified content-changed hunk/blob delta or an unverified edit-targeted tool target. These states are not interchangeable.

Observed edges may use contains, read, edit-targeted, content-changed, included-in, produced, validated-by, or correlated-with. Their strength is direct, observed, or correlated; an analysis cannot claim stronger support than its cited evidence. A claim must cite its subject directly or cite an observed edge whose subject, object, or evidence refs name that subject.

Analysis output

Return one JSON object:

{
  "kind": "IntentCorrelationAnalysisV1",
  "schemaVersion": 1,
  "packetDigest": "sha256:<copied from packet>",
  "intentProposals": [
    {