For legal & compliance teams
Evidence of licensed acquisition, prepared for legal review.
Kelp Market supplies training datasets accompanied by a documented, independently verifiable record of provenance and license. The record comprises three elements: verification of the capture device; a signed license grant; and a fixed verification procedure. On an infringement inquiry, counsel produces this record as documentary evidence of licensed acquisition. Verification executes against public records and requires no reliance on Kelp Market.
Artifacts delivered with each bundle
Each delivered dataset includes the following artifacts. Each is signed, and each resolves at a permanent public URL.
| Artifact | What it evidences |
|---|---|
| LICENSE.md | The complete license instrument, delivered as a file within the bundle. The contractual text is fixed at delivery; it is not a reference to terms subject to later modification. |
| manifest.json | The transaction record: asset count, aggregate price, currency, and the license terms as fixed at quotation. Signed; any subsequent alteration is detectable. |
| Per-asset signed manifests | One certificate per asset stating the asset identifier, its location within the bundle, the applicable license, the rights holder, and the price. Each certificate bears a digital signature; alteration of a single byte is detectable. |
| SEAL records | Tamper evidence for each certificate: a content fingerprint (any alteration to the media changes the fingerprint) and a signature over the certificate bytes, bound to the verification key published by Kelp Market. |
| Manifest URLs | Each grant resolves at a permanent public URL (/api/v1/manifests/:id) and may be cited in an audit, a negotiation, or a court filing. |
Technical annex: signature and hash specifics
Certificates are signed with Ed25519; content is fingerprinted with BLAKE3; the public verification key is published as a DNS TXT record at _seal.kelp.digital.
Scope of the license
The bundle's LICENSE.md constitutes the operative contract. The full license
text is delivered within the bundle rather than incorporated by reference to terms that may be modified
after purchase. The standard licenses grant, in substance:
- Training rights. Use of the licensed assets to train, fine-tune, or evaluate artificial intelligence and machine-learning models.
- Limitations of scope. Training-only licenses exclude display, distribution, or public presentation of the assets — or of outputs reproducing them — absent a separate grant.
- Deployment. Where the license so provides, models trained on the assets may be deployed and operated, including in commercial products, subject to the remaining terms.
The applicable clause set is fixed per license at purchase and recorded in the signed manifest.
License catalog
- kelp-training — training-only scope; deployment of trained models permitted; no resale or marketing use.
- kelp-platform-v1 — broad platform rights: reproduction, distribution, derivatives, agency sublicensing, attribution.
- royalty-free — training plus reproduction, distribution, and derivatives.
- commercial — the foregoing plus express commercial use.
Effect of this evidence in a dispute
- Authorship is evidenced by hardware, not by assertion. Assets enter a set only after verification tied to the physical camera: metadata linkage, a challenge pass, or a sensor-fingerprint match. The authorship claim rests on evidence concerning the device.
- The license is a signed instrument, not a receipt. The manifest records the pricing and license terms as fixed at quotation, under a signature that detects alteration. Terms cannot vary silently after purchase.
- Verification is independent. Verification executes in the browser against public records: retrieval of the manifest, check of the fingerprint, check of the signature, resolution of the published key. An auditor, opposing counsel, or court-appointed expert requires no account, no API key, and no reliance on Kelp Market.
- The record is permanent and citable. Grants reside in a tamper-evident public record at stable URLs, suitable for citation in an audit, a negotiation, or a filing.
This evidence addresses the licensee's acquisition. It does not of itself preclude a challenge to the licensor's upstream title; that residual risk is addressed by the capture-verification tiers below, which bind each asset to a verified physical camera.
Strength of capture evidence
Three verification tiers. The tier of each asset is reflected in its pricing; stronger evidence costs more.
| Tier | Method |
|---|---|
| Self-reported metadata | The camera serial number as recorded in the file's embedded metadata at upload. Weakest tier; priced accordingly. |
| Challenge verification | The uploader photographs a single-use challenge code with the claimed body and lens; the serial numbers visible in that image are cross-checked against the record. |
| Sensor fingerprint | Each sensor exhibits a stable, unique noise pattern (photo response non-uniformity). The image is matched against the fingerprint of the physical sensor, confirmed by a blind three-sample comparison. Strongest tier. |
Technical annex: sensor fingerprinting
The method is photo response non-uniformity (PRNU): manufacturing imperfections in a sensor leave a stable noise pattern in every image it captures. The three-sample triangle test requires a blind evaluator to match the image to the correct sensor among three candidates. A pass establishes that the match is not attributable to chance.
Comparison with media-embedded provenance
Standards that embed provenance in the media (C2PA / Content Credentials) exhibit documented structural weaknesses. The present design does not inherit that architecture.
| Documented gap | Embedded-manifest approach | Kelp approach |
|---|---|---|
| Timestamp ambiguity | Formal analysis (UMBC / Hacker Factor, 2026) showed that validators and generators may disagree on trusted timestamps — competing timestamps for a single claim. | One signature over the manifest; one hash; no timestamp-authority chain to dispute. |
| Validator divergence | Public validators disagree on identical files; malformed fields may be silently ignored. | A fixed verification procedure: if any check fails, the proof is shown as unverified. There is no partial pass. |
| Defective revocation | Certificate revocation management was found inadequate; trust lists add certificate-authority complexity. | One trust anchor: the verification key is published, so any party may check signatures itself. |
| Provenance embedded in the media | Embedded chains can be stripped or altered in transit; tooling faults have allowed altered manifests to pass. | Certificates reside in a tamper-evident public record at permanent URLs. Nothing is embedded in the image to strip. |
Gap analysis informed by: Golaszewski, Krawetz et al., "Verifying Provenance of Digital Media: Security Analysis of C2PA and its Implementation" (UMBC Cyber Defense Lab, 2026).
Limitations of the evidence
- Capture evidence attests that the image originated from the verified camera. It does not attest that the scene is unedited after capture; no provenance system makes that attestation.
- A license grant evidences the transaction and its terms. It is not legal advice and does not substitute for independent review of the license scope.
- Verification depends on the manifest URL and the DNS-published key remaining reachable. Both are public infrastructure under the control of Kelp Market; the key material is disclosed so that grants remain independently checkable.
Verification demonstration
Open a specimen proof to execute the verification procedure — hash, signature, SEAL record, key resolution — in the browser.