Three tools with complete JSON schemas and detailed descriptions. All parameters are typed and described. Tool names are action-oriented (detect, verify, sign). Descriptions are comprehensive (200-1000+ chars) and explain WHEN to use each tool and what it returns. detect and verify are read-only; sign is destructive and properly marked. However, descriptions are verbose and could be more LLM-optimized (ideal 50-200 chars per pattern:tool-description baseline). No enum constraints on string parameters (e.g., action, soft_binding, digital_source_type). Error handling guidance is minimal, no recovery hints or actionable error messages visible in tool definitions. Output schemas are not explicitly documented in the tool definitions themselves.
Report what a JPEG, PNG, WebP, GIF, TIFF, HEIC, AVIF, SVG, MP4, MOV, AVI, WAV, MP3 or PDF file CLAIMS about its C2PA / Content Credentials provenance (generator, title, AI flag, claimed signer and signing time). Fast and UNVERIFIED — like reading EXIF; it does not check signatures. Use verify to validate.
Sign a JPEG, PNG, WebP, GIF, TIFF, HEIC, AVIF, SVG, MP4, MOV, AVI, WAV, MP3 or PDF file with THIS SERVER'S configured signing key: embeds a C2PA claim.v2 manifest (title, a created or opened action, the content hash) and writes a signed copy. An asset that already carries Content Credentials keeps them — the prior manifest becomes the new one's parentOf ingredient — which requires `action` opened (the default when a manifest is present); created on such an asset is refused. Input: exactly one of path, url or bytes. Output: `output` names the file to write (required for path and url inputs; an existing file is refused unless `overwrite` is true); a bytes input may omit it to receive the signed asset as `signed_bytes`. The result's `verify` block is the c2pa library's verdict on the output, including the `binding` the manifest just written covers, anchored at the signing certificate chain and without re-validating a prior manifest — call verify on the output for the full picture. Optional `soft_binding: "iscc"` ALSO writes a soft binding: an ISO 24138 ISCC Image-Code (io.iscc.v0) computed from the image content, which survives the re-encoding and metadata stripping that break the hard binding, so a copy can be matched back to this manifest through a provenance store. Still images this build can decode only — JPEG, PNG, GIF, WebP and TIFF — and the hard binding is still written, since a soft binding is never an asset's only content binding. Anyone who can call this tool signs with the content binding. When the server was started with an identity credential, `identity_roles` and `identity_references` also write a cawg.identity assertion: a second signature by a NAMED ACTOR saying who vouches for the content. That actor's key, like the signing key, is the operator's and is never a tool argument. Anyone who can call this tool signs with the operator's key.
Descriptions are verbose (200-1000+ chars) and lack LLM-optimized brevity. Ideal is 50-200 chars per pattern:tool-description baseline. Verbose descriptions waste tokens and bury key decision points.
String parameters lack enum constraints. 'action' (created|opened), 'soft_binding' (iscc|none), and 'digital_source_type' should be enums, not free-form strings. This invites hallucinated values from LLMs.
No error handling guidance in tool definitions. No recovery hints (e.g., 'If file not found, check path or try a URL instead'). Error responses must tell LLMs what to do next per pattern:recovery-guide.
Output schemas not documented in tool definitions. LLMs need to know what fields to expect (e.g., detect returns what structure? verify returns what fields?). Undocumented outputs force LLMs to guess downstream field names.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
Fully validate a JPEG, PNG, WebP, GIF, TIFF, HEIC, AVIF, SVG, MP4, MOV, AVI, WAV, MP3 or PDF file's C2PA / Content Credentials manifest: COSE signature, certificate chain against the trust list, assertion and hard-binding hashes, and the RFC 3161 timestamp. Returns an overall valid flag plus per-step status codes. Read `verified_signer` for who provably signed it — it is empty unless the signature verified AND the chain reached a trust anchor. `signers` is the chain as PRESENTED in the file and is populated even when validation failed, so it is a claim, not proof; the same is true of everything under `detect`. `binding` answers the separate question of whether THESE bytes are the ones that were signed: verified, failed, none (nothing bound the asset), or unevaluated — which is neither a pass nor a failure but "a hard binding exists and this call did not check it", as for a PDF manifest attached to an embedded object, an asset past the scan cap, or a fragmented asset without its fragments. A manifest can be valid with an unevaluated binding. `soft_bindings` lists any perceptual identifiers the manifest carries (a watermark or a fingerprint, such as an ISCC) — REPORTED, never checked: matching one means recomputing it and comparing within a tolerance, which is a policy decision, so nothing here says the content matches. Do not report a soft binding as evidence of anything. `identities` lists any CAWG identity assertions — named actors who signed over the content with their own credential. Read each entry's `name` for who was PROVEN (empty unless `trusted`) and `presented_as` for who was merely claimed — the same split as `verified_signer` and `signers`. An identity means the actor VOUCHED for those assertions: it conveys neither attribution nor ownership, so never say an asset was created by them. For an aggregation credential, `verified_identities` is the AGGREGATOR'S word about signals it checked, proven only as far as that aggregator is trusted. Neither trust list has a default, so unproven is the normal, honest outcome rather than a warning; a cawg.ica.did_unsupported_method failure means this build cannot resolve that aggregator's DID method, which is a limit here and not a defect in the file.
sign tool description does not clearly state it is destructive/irreversible. Per pattern:command-tool, tools that modify state must explicitly say so. Current description is long but lacks a clear 'This writes a file' opening.