Loading ToolBench
Read-only AI provenance auditing. Locate C2PA evidence structurally, verify it cryptographically with c2patool, measure whether a publishing pipeline preserves it, audit metadata for privacy exposure, check disclosure records, and run text-provenance detectors. Nothing here removes marks, and no tool asserts human or AI authorship.
Seven read-only tools with strong, domain-specific descriptions (avg 180 chars). All tools have complete input schemas with required fields and type definitions. Naming is clear and verb-forward (audit_, inspect_, verify_, map_, check_, detect_). However, parameter descriptions are sparse, most lack guidance on format, constraints, or valid ranges. Output schemas are undocumented; LLMs cannot predict response structure. Error handling is minimal; no recovery guidance. No tool annotations (readOnlyHint present in metadata but not in schema). Composition is sound: each tool has one responsibility, and the front-door audit_provenance orchestrates others.
Audit a file for privacy-sensitive metadata: location (GPS, geotags), identity (author, creator, email), device (camera make/model), software (tools used), timeline (creation/modification dates), and comments. Reports findings by category and severity, never values.
Front door. Answers whether provenance was located, cryptographically verified, and trusted under a named policy, whether the scan was complete, and what remains unknown. Composes the other tools; performs no authorship classification.
Check whether a file carries AI transparency disclosures: C2PA manifest claims, machine-readable metadata, or human-readable statements. Reports what was found, not whether it is sufficient.
Run text-provenance detectors (Watermark, Copyrighted, Synthetic) on text to measure statistical evidence of AI generation. Reports detector outputs and confidence; never asserts authorship.
Inventory a file or string for structurally located C2PA evidence, hashes, and hidden Unicode channels. Never asserts authorship; a literal mention of C2PA in text is recorded separately and is not evidence.
Output schemas undocumented. LLMs cannot predict response structure, field names, or types. Downstream tool chaining and result extraction require guesswork.
Parameter descriptions lack format/constraint guidance. 'path' appears in 6 tools with no hint on absolute vs relative, symlink handling, or file type restrictions. 'c2patool' and 'trust_anchors' lack format examples.
No error handling guidance. Tools return raw errors with no recovery hints. E.g., if c2patool is not found or trust_anchors is invalid, LLM has no actionable next step.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Compare an original with derivatives to measure whether a publishing pipeline preserves Content Credentials. Reports observed outcomes, never intent.
Verify C2PA Content Credentials with the official c2patool, reporting manifest presence, cryptographic integrity, and signer trust as independent results. Requires c2patool 0.20.0+.
Tool annotations missing from schemas. Risk metadata (READ_ONLY) is in comments but not in inputSchema as readOnlyHint. Clients cannot auto-detect safety properties.
map_provenance_survival 'derivatives' param uses 'LABEL=PATH' string format with no validation or parsing guidance. Ambiguous for LLMs; no schema constraint (enum, pattern, or format).