Evidence, selectors and verification
Preserve what supports an assertion and describe exactly what was checked
Evidence has two parts: a preserved capture and a precise selection within it. A citation URL identifies a location; it does not guarantee that the bytes seen during the decision remain available or unchanged.
Captures
A capture records acquisition start/finish, requested/final URL where relevant, collector/version, media type, completeness, limitations, integrity and storage identity. Byte captures require a verified object version/hash. Pending, partial and failed captures retain their acquisition state rather than pretending to be complete.
Examples include an original HTTP response, rendered page, screenshot, provider payload, uploaded file or metadata-only capture. Two acquisition attempts remain distinct even when content hashes match. Access and retention policies apply to the evidence itself.
Select exactly what supports the statement
| Selector | Required location/integrity details | Use |
|---|---|---|
| Text span | Start/end offsets, exact quote and quote SHA-256 | One passage in captured text |
| Page region | Page, x/y/width/height and coordinate system | A region in a document or screenshot |
| Time range | Start/end milliseconds; optional transcript | A passage in media |
| Structured path | RFC 6901 JSON Pointer and selected-value SHA-256 | A specific response value |
Links state whether the evidence supports, contradicts or repeats the assertion. Preserve contradictory evidence rather than hiding it to improve a confidence score. Repeated assertions from dependent sources do not establish independent confirmation.
Example: one receipt supports delivered
Suppose the supported collector/upload workflow has preserved a JSON receipt {"order_id":"O-001","delivered":true} in this project. Use its real accessible capture UUID and calculate the selector hash using the collector/runtime's selected-value serialization. Verify the bytes before creating an extracted assertion for subject https://example.com/orders/1, predicate https://example.com/operations#delivered, value boolean true.
Collect the receipt through a supported evidence path, inspect completeness and integrity, select /delivered, and verify that the receipt identifies O-001. Attach that selector to the extracted assertion. This is not performed by entering a citation in the basic human assertion form.
{
"captureId": "<REAL_CAPTURE_UUID>",
"relation": "supports",
"selector": {
"kind": "structured_path",
"jsonPointer": "/delivered",
"valueSha256": "<SHA256_OF_SELECTED_VALUE>"
},
"sourceReliability": 0.9
}This structure is a contract example, not a public HTTP endpoint or an MCP create tool. Replace the capture/hash placeholders with real verified values. Extracted and derived assertions require evidence; inventing a UUID or writing true in a free-form evidence object does not satisfy that requirement.
MCP capture_get inspects the private capture manifest and custody information without returning raw bytes to the prompt. assertion_get inspects actual assertion links. Neither tool creates missing evidence.
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "capture_get",
"arguments": {
"projectId": "<PROJECT_UUID>",
"captureId": "<REAL_CAPTURE_UUID>"
}
}
}Verification tasks and findings
A verification task records a method, instructions and limitations for an assertion. A finding records AI/reviewer origin, supports/contradicts/inconclusive/unverifiable result, rationale, limitations and evidence links. An AI finding cannot review another finding as a human. The contract separates checking from approving the original claim.
For the receipt, the task can ask “Does this receipt identify O-001 and record completed delivery?” A receipt for O-002 contradicts the association even if its delivered value is true. An unreadable receipt is inconclusive/unverifiable, not false.
The task/finding contracts are available in the governed runtime. The currently documented public API/MCP surface does not expose generic task/finding creation calls; do not substitute internal routes as workspace-key APIs. Inspection is available through claim detail.
Derivations and downstream impact
A derivation records input/output resources, execution identity, transform/version, model and prompt provenance where applicable, parameters and recorded usage/cost. It explains how a source capture or assertion produced another assertion/artifact. Corrections and retractions can mark dependent outputs stale or requiring review; they do not automatically recompute every downstream asset.
Inspect the impacted records and rerun/review the relevant consumers. Evidence history explains a past decision; it does not make that decision eternally valid.
Original source provenance
An assertion can reference an existing project-owned source artifact and its
record ID using sourceArtifactId and sourceRecordId; citationUrl is a separate
optional reference. The creation form lists available scoped artifacts and pages
through declared _source_record_id values from preserved JSON/Parquet bytes.
Choose the original field where available. A standalone assertion is still valid.
A manually entered record ID is a reference, not proof that the record exists.
New source links record the artifact revision in metadata. The provenance section checks that revision, content checksums, record uniqueness and the declared field. Unavailable bytes, missing IDs, unsupported formats, conflicting legacy references and missing historical revisions remain visibly unresolved. Source preview is bounded to 32 MiB per file; full record resolution has a 10,000-row preview budget. These preview limits do not truncate forecast evidence or silently confirm a link.
Source provenance describes what a statement concerns. Supporting captures describe why it should be believed. They remain distinct, and capture access policies apply. Assertions do not automatically modify source records or override query bindings.