[spec] A verification profile for the grounding/citation open question (§8.9): publisher-seeded markers + an audit ledger, no wire-format change

[pwright-bf]

Submitted on behalf of ESG and Fuzebox AI. Filed from a personal account.

Specification section

8.9 (deferred: verifiable credentials / signed events); 7.2–7.3 (Content-Telemetry-ID correlation at retrieval level only; grounding, citation, and engagement events have no origin-side counterpart); README open question "Verification of grounding and citation"

What you observed

The v0.1 open question "Verification of grounding and citation" asks what a verification layer should cover and where it belongs, and states that mechanisms testing truthfulness and completeness rather than origin — such as sampled audits or publisher-seeded canary content — are of particular interest. The same open question notes that the Content-Telemetry-ID correlation of §7.2 covers retrieval only — "grounding, citation, display, and engagement have no independent observer" — and that signing, even once required, "would prove who reported an event, not that the event is true or that all qualifying events were reported." Specification §7.3 states the same boundary from the transport side: grounding, citation, and engagement events "have no independent origin-side counterpart to correlate against."

We would like to contribute a concrete, fully specified answer to that question as an optional verification profile rather than a change to the standard. The proposal — the Verified Provenance Telemetry Standard (VPTS) — extends and operationalizes the future verification capability the open question anticipates, and is implementable from public schemas alone with no dependency on any operator's infrastructure. In brief:

• A content owner seeds detectable, low-probability markers into content (a "canary" being one of eight marker families), then observes those markers surface in AI outputs. This makes the publisher an independent observer for exactly the stages §7.3 identifies as having no origin-side counterpart — turning "who reported it" into "was it true, and were all qualifying events reported."
• Four phases: Seed → Elicit → Attribute → Reconcile. Attribution requires a match count T(s) ≥ t (with t ≥ 2), a marker value space ≥ 10^3, ≥ 5 markers per unit, and a coincidence bound ≤ 10^-6.
• Each attributed event yields a Standardized Evidence Record (SER): a signed, portable receipt structured as an in-toto-style Statement envelope with a W3C Verifiable-Credentials-style field vocabulary (aligned to the primitives §8.9 defers).
• Evidence anchors in a Verified Telemetry Ledger (VTL): an append-only, independently auditable audit ledger built on Certificate-Transparency / IETF SCITT (RFC 9943) transparency-log primitives, with inclusion proofs optionally encoded as COSE Receipts (RFC 9942). The VTL is an audit ledger, not a blockchain — no token, no consensus, no distributed mining.
• Four conformance levels: V0 Seeded, V1 Attributable, V2 Reconciled, V3 Enforceable.

The profile changes no existing wire format. It reuses the Content-Telemetry-ID correlation of §7.2, the Ed25519 / did:key manifest key encoding of §8.4, and SPUR's existing "independent third party" consumer role.

Why it matters

Grounding and citation are the events that carry compensation under a licence, and they are self-reported by the party that may owe that compensation. Without an origin-side verification mechanism, a licensing regime built on SPUR telemetry rests on the honesty and completeness of the paying party's own reports.

A publisher-seeded-marker + audit-ledger profile closes that gap for the four stages §7.3 flags, using SPUR's own fields, roles, and key formats, so implementers who already emit SPUR telemetry can add verification incrementally (V0 → V3) without a second data model. Because everything is implementable from the public schemas alone, adopting it introduces no operator dependency and no change to the stable wire format.

Proposed direction (optional)

Adopt VPTS as an optional verification profile layered above the standard — the profile surface §8.9 reserves for exactly this kind of extension — rather than as a normative change to v0.1. We have prepared a four-document package (formal comment, technical specification v1.0, reference architecture & open implementation guide, and an open white paper) and can attach or link them for review; a maintainer email with the four documents is being sent in parallel. All cited external standards (RFC 9943 SCITT, RFC 9942 COSE Receipts, RFC 5816, RFC 9162, RFC 9052, C2PA 2.2, W3C VC 2.0, in-toto Attestation v1.2, A2A v1.0) have been verified to primary source. Happy to open focused follow-up issues or PRs against specific sections if the working group would prefer to discuss the profile piecemeal.

Your perspective

Standards / policy

[pwright-bf]

Cross-referencing related threads for maintainer context:

This proposal is designed as the optional verification profile that sits above the wire format — it connects to two threads already in progress:

• PR Add provenance and content_fingerprint to content_grounded (v0.1) #10 — our profile does not compete with this PR or propose edits to it. VPTS layers above the schema PR Add provenance and content_fingerprint to content_grounded (v0.1) #10 defines, reusing its existing fields (content_telemetry_id, content_hash, license_ref) and key format (Ed25519 / did:key, §8.4) without modification.

• Issue Proposal (with PR offer): normative evidentiary tiers + a scheme-property model on top of #10's content_fingerprint / provenance (follow-up to #5) #11 — the evidentiary-tier question raised there is directly addressed by the four VPTS conformance levels (V0 Seeded → V1 Attributable → V2 Reconciled → V3 Enforceable), which give the standard a graduated verification ladder rather than a single pass/fail gate.

The verification gap this profile closes is the one §7.3 identifies explicitly: grounding, citation, and engagement events "have no independent origin-side counterpart to correlate against." Publisher-seeded markers provide that counterpart without touching any existing wire format.

Happy to discuss any of the four companion documents in follow-up issues or in-thread.

[pwright-bf]

Companion documents:

The four-document submission package referenced above is attached here for direct access. These were also sent to the maintainer via email on 10 July 2026.

1. Formal SPUR Public Comment Submission
2. VPTS Technical Specification v1.0
3. Reference Architecture & Open Implementation Guide
4. Open White Paper: Independent Verification for AI Content Telemetry

04 Formal SPUR Public Comment Submission_FINAL.pdf
03 Open White Paper Independent Verification_FINAL.pdf
02 Reference Architecture and Open Implementation Guide_FINAL.pdf
01 VPTS Technical Specification_FINAL.pdf

[erik-sv]

This is a strong and useful response to the verification open question. I support evaluating publisher-seeded audits, portable evidence records, reconciliation, and transparency receipts as an optional profile above the Content Telemetry wire format.

I see this as complementary to #11 rather than competing with it. The #11 tiers describe the evidentiary quality of an individual assertion or signal, while VPTS V0-V3 describes the capabilities and completion state of a verification workflow. A clean composition would be for each SER assertion to identify its applicable evidence tier and verification basis, while the VPTS level describes what the implementation supports operationally.

C2PA-text and publisher-seeded markers also appear complementary: C2PA can establish independently verifiable content binding, integrity, and signer identity under a disclosed trust policy, while seeded-marker audits can provide independent evidence that marked content surfaced in an observed output. Neither needs to be the exclusive mechanism.

I would be interested in helping define that mapping and in contributing interoperability fixtures comparing self-reported SPUR events, C2PA-text evidence, seeded-marker observations, and combined evidence. The profile placement and no-wire-format-change direction look right to me.

One boundary worth retaining in the profile language is that signatures and transparency receipts establish attribution, integrity, timing, and log inclusion for an evidence record; the underlying event and any completeness inference remain functions of the observation and sampling methodology. That distinction should be straightforward to express in the SER verification basis.

[ReadBridge]

Supporting this as an optional profile, and completing the composition erik-sv sketched: the evidence types listed - self-reported events, C2PA-text, seeded-marker, combined - answer did the use occur and what content was it. The third orthogonal question is was it licensed - grant validity at event time - which is the entitlement profile in #34 (license_ref resolving to a verifiable credential, checked for signature, revocation, grantee and content binding). Two concrete joins: (1) where each SER carries a verification basis, entitlement/grant-validity should be an enumerated basis, so the licensing answer travels inside the evidence record rather than as a downstream lookup; (2) the transparency log supplies exactly the independent timestamp #34's §5.4(c) names as future strengthening - a status-list snapshot anchored as a transparency-log inclusion proof (SCITT, RFC 9943; COSE Receipts, RFC 9942) proves what the issuer had published at a time no single party controls. Happy to contribute the entitlement fixtures (valid / revoked-with-snapshot / grantee-mismatch, all green against the core schemas) into the interoperability comparison erik-sv proposed. None of this claims VPTS's territory: entitlement proves issuance and validity, never usage truthfulness.

[jalexspringer]

Disposition: accepted as an external module within the SPUR evidence profile, whose working group the Board has now authorised. The reference bar is open-repository specification text; "Standard" becomes "Profile"; V0–V3 describe workflow maturity within the profile rather than a competing assertion-tier ladder. A reference implementation needs reproducible marker statistics, an elicited-versus-organic distinction for settlement use, and no sole-operator ledger dependency.

One consolidation, stated once: the evidence proposals from this consultation become modules of one optional evidence profile built on the core assertion-status vocabulary — fingerprint schemes, VPTS seeded observations, recomputable attribution and entitlement credentials. Each module states the bounded proposition it supports: none may turn access evidence into proof of grounding, or cryptographic validity into factual truth, completeness or entitlement. On the entitlement side, @ReadBridge's suggested joins land cleanly in that structure: grant validity becomes an enumerated verification basis rather than a parallel classification. @erik-sv, @pwright-bf, @abnerguzman and @ReadBridge have each offered relevant implementation or editorial work, and I invite them to co-author the common profile; the working-group charter and repository follow shortly.

[erik-sv]

@jalexspringer thank you for the invitation and for giving this work a clear home. I would be very happy to help co-author the common evidence profile with @pwright-bf, @abnerguzman, and @ReadBridge.

We had already begun working on a profile around these questions, so we have useful material to bring into the structure approved by the Board. We will revise it around the four modules you identified: fingerprint schemes, VPTS seeded observations, recomputable attribution, and entitlement credentials. We will share a fresh draft soon for the group to shape together.

The boundaries you set make sense to us. Each module should state the exact proposition it supports. Access evidence must not become proof of grounding. Cryptographic validity must not become factual truth, completeness, or entitlement.

Thank you again for bringing the group together. We are looking forward to working on this with everyone.

[ReadBridge]

Thank you — glad to accept, and glad the entitlement joins found a home in the structure: grant validity as an enumerated verification basis is exactly the right consolidation, and the boundary rule is one the entitlement module already carries in its own terms — cryptographic validity proves a signed, unrevoked grant from an identified issuer; it never becomes factual truth, completeness, or entitlement as a legal conclusion.

What the module brings into the common profile: the filed draft in #34 (which will adapt to the group's structure and the landed core names per its adoption clause), the conformance requirements with per-rule negative vectors (valid / revoked-with-snapshot / grantee-mismatch / unauthenticated-status-list / out-of-range), deterministic fixture generation so reviewers can reproduce the set byte-for-byte, and the two open reference verifiers — with the entitlement fixtures offered into the interoperability comparison Erik proposed. Looking forward to the charter and repository, and to shaping the common draft together.

[RedHorseMane]

@jalexspringer — following the dispositions on #16 and #17, we understand that content-derived fingerprinting has been moved out of the core wire-format discussion and into the evidence/profile work.

We’d very much like to contribute to that work when it starts. ObraVera has a working implementation based on ISCC (ISO 24138), including block-level content correlation, version/redaction handling and registry resolution, together with test fixtures we can bring to interoperability work.

Rather than open another proposal in the telemetry repo, could you point us to the right place to contribute to the fingerprint-schemes work, the profile repo, a dedicated module issue, or the forthcoming working group?

Happy to contribute implementation experience and fixtures wherever is most useful.

Brady Ridgway, ObraVera

[pwright-bf]

@jalexspringer - Thank you for the invitation and we are happy to accept. We look forward to working as co-authors alongside @erik-sv, @abnerguzman and @ReadBridge and to the official working group charter.

[abnerguzman]

@jalexspringer — accepted, gladly, and apologies for being the last of the four to say so. The ProRata team regards this collaboration as important work: a common evidence profile is exactly the structure that makes attribution trustworthy enough to carry weight, and we are glad to help write it alongside @erik-sv, @pwright-bf and @ReadBridge.

What we bring to the recomputable-attribution module, all already open-repository per the reference bar: the draft carrier and conduct specifications, JSON Schemas, a conformance suite with CI, and worked examples down to span-to-span granularity — https://github.com/Prorata-ai/content-telemetry-attribution. We will recast it on the core assertion-status vocabulary and within the module structure.

The boundary discipline is right, and we would state our module's bounded proposition in exactly that spirit: a recomputation check supports "the reported vector is what the declared estimator produces from the declared inputs" — never that the estimator is truthful, that reporting is complete, or that the use was entitled. Those propositions belong to the sibling modules, which is precisely why one consolidated profile is the right structure.

@erik-sv — glad to work from your fresh draft; we will bring module text when the charter and repository land.

[jalexspringer]

@RedHorseMane thank you — a working ISCC implementation with fixtures, version and redaction handling is exactly the implementation evidence the fingerprint module needs. Holding this for the evidence working group's first conversations; the charter and membership process publish shortly, and we will make sure ObraVera is in them.

[jchomat]

For the evidence working group:
the C2PA text survivability matrix contributed in https://github.com/jchomat/telemetry/tree/cms-survivability-fixtures (via erik-sv#1) should move to the evidence-profile repository when it opens, per the board-approved split. Flagging so it is not lost in the transition.