Documentation/Reference

Reference

Terms and evidence boundaries

SDK 0.7.2a1 · documentation review 2026-11-10.

Term Meaning in this toolkit
Marking Adding a machine-readable indication to an output
Watermark An imperceptible signal embedded in visual samples, carrying a recoverable ID
C2PA credential Signed manifest/assertions bound to an asset; not proof that every assertion is factually true
Provenance Claims about origin and history; separate from factual truth, consent and ownership
Issuer trust Validation against a configured trust policy; mathematical signature validity alone is different
Provider key Signs content credentials; mounted only by the exporter
Report key Signs the verifier's result; distinct from content and TLS identities
TSA Timestamp authority used under the current production signing policy
Registry Operator-owned ID/record/manifest associations; not a global discovery service
Soft binding A recoverable association through a watermark, distinct from cryptographic file binding
Metadata-only recovery Recovery after tested metadata removal without re-encoding media; broader edits need separate evidence
Economy Fresh marking on selected 125 ms windows, with 5 s or 2 s periods
Resilient Fresh per-frame marking; five-position export read-back, not an unremovable mark
Negative Scoped completed detection with no supported mark; does not establish human creation
Incomplete Operational/budget/engine limit; not evidence of absence
Robustness Performance under stated transformations/threats, with parameters and failures reported
Compliance mapping Connection between an applicable norm, technical measure, evidence and operator action
Qualification Exact-version engineering checks within a stated scope; not product-wide legal approval

Ariadne detects supported markings, not every AI-generated file. Unknown provider watermarks require their own detectors. No universal confidence score is calibrated across engines. See verification results.