Proposal (with PR offer): normative evidentiary tiers + a scheme-property model on top of #10's content_fingerprint / provenance (follow-up to #5)

[erik-sv]

Summary

#5 surfaced the third-party-sourced (resale) coverage gap, and #10 already implements the two fields agreed there: the provenance enum and the scheme-agnostic content_fingerprint object. #10 intentionally leaves two things to a follow-up: the claims-vs-corroborated-facts section its field descriptions are written to slot into, and any structure on the free-string scheme. This issue proposes both, and offers a PR that builds on #10 rather than replacing it.

The intent is not to favour any technique. It is to give the spec a principled, testable way to say which reported signals can be relied on independently of the reporter, so the numbers the standard produces can carry the weight downstream compensation will put on them.

Motivation

Telemetry is a cooperative measurement layer. It reports what willing participants emit, so by construction it cannot represent a non-emitting actor, and it should not try to. That part of the resale gap is an access-layer concern (RSL / CAP / bot-auth), and the right move in the spec is to name the boundary (8.9), as @landomo suggested.

The numbers it does produce are self-reported and become a basis for compensation. #10's descriptions already say detected and preserved_in_output are inference with no origin-side counterpart (the §7.3 gap), and invite "the broader claims-vs-corroborated-facts section when that lands for v0.1." This is that section, plus the scheme semantics that make the distinction operational.

Proposal

1. Normative evidentiary tiers

Define, in a dedicated evidentiary section, three tiers that any reported signal falls into:

• Tier 1, Claim: agent-reported, uncorroborated.
• Tier 2, Origin-corroborated: a matching origin-side event exists.
• Tier 3, Independently verifiable: checkable by any party against a public trust anchor and a trusted timestamp, with no origin-side counterpart required.

Tier 3 is the only tier that closes the §7.3 gap, because it depends on neither the agent's self-report nor an origin-side event. The tiers are scheme-neutral: any technique that qualifies can reach a given tier, and more than one should. This composes directly with #10: its field-level inference labelling becomes "Tier 1 unless the scheme qualifies higher."

2. Scheme-property model for content_fingerprint

Keep scheme a free string, as in #10. Add declared scheme properties that map to the maximum tier the scheme's signals can claim:

• content_borne: the signal lives in the content and survives reprocessing, rather than in detachable metadata.
• identity_bearing: it carries rights-holder identity, not only a match.
• verifiable: it is cryptographically checkable against a public anchor.

A scheme that is content-borne, identity-bearing, and verifiable is eligible for Tier 3. A matching-only fingerprint (for example a similarity hash without a registry) is Tier 1 or 2. An optional evidence pointer on the object can say where the signal is independently checkable.

3. provenance enum

No change proposed; #10's enum and placement are right. Listed here only for completeness.

Worked example: C2PA / C2PA-text as one Tier-3 scheme

Offered as one worked example, not a dependency, and explicitly inviting other Tier-3 schemes. #10 already lists C2PA text provenance as a candidate scheme; this formalises what that entry means.

C2PA text provenance (the A.7 unstructured-text method) binds the manifest into the text itself, so it is content-borne: it survives copy, plain-text save, encoding and serialization round-trips (UTF-8/UTF-16, JSON), and all Unicode normalization forms (NFC/NFD/NFKC/NFKD). A single contiguous in-band block survives whole-text and aggregated reuse; distributed markers extend survival to the fragment case, so a copied span, even mid-sentence, still resolves to the source work. It is identity-bearing (signed publisher certificate) and verifiable (against the public C2PA trust list), with first-publication anchored by an RFC 3161 timestamp, which also answers ISCC's two honest limits: identity rather than matching, and a first-publication anchor.

The "C2PA is attached metadata, stripped on reprocessing" characterisation in #5 is accurate for a manifest embedded in a media file, but not for in-band text provenance, which is a different mechanism.

Survivability: evidence, not assertion

The PR will include a transform-by-transform survivability matrix (Unicode normalization forms, encoding and serialization, copy-paste through common editors, whitespace and markup sanitizers, fragment and aggregation), backed by test output, so robustness is checkable rather than claimed.

Honest boundaries, stated up front:

• No in-band scheme survives genuine paraphrase or summarisation, because the carrier goes with the rewritten text. The resale-path boundary in Provenance gap for third-party-sourced grounding (the scraper-resale supply path) #5 therefore holds regardless of scheme.
• Targeted removal of a specific carrier defeats that one channel; this is mitigated with complementary carriers and error correction, not eliminated.

Consequence worth encoding: absence is informative

Once a robust, expected mark is in play, its absence on otherwise-matching content is itself a signal, and a grounding system asserting a reliable source it cannot attribute is carrying that risk. That is the demand-side reason a cooperating agent populates these fields honestly: they are its own due-diligence record, not only publisher telemetry.

Out of scope (explicit)

• Adversary coverage and the resale path itself: that is the access layer, not telemetry. This proposal does not claim to close it; it makes the cooperating channel's signals trustworthy and names the boundary.
• Semantic or paraphrase-robust fingerprinting: pre-standard today and not relied upon here.

Contribution offer

I am happy to author a follow-on PR (on top of #10) that:

1. adds the normative evidentiary-tiers section;
2. adds the scheme-property fields and the property-to-tier mapping to content_fingerprint;
3. adds a scheme-neutral C2PA-text profile as one worked Tier-3 example;
4. attaches the survivability matrix.

Co-authorship with @landomo and @jchomat welcome, and additional Tier-3 scheme registrations more so.

Open questions

• Should the tier be an explicit field, or derived from the scheme's declared properties?
• Is a small registry warranted for scheme strings and their declared properties and tier?
• How should the Tier-2 origin-corroboration path relate to the access layer (for example RSL <reporting>)?

[jalexspringer]

Unattributed consultation feedback for the record - input for triage, not a maintainer decision or resolution.

Direct publisher feedback received outside GitHub during the comment window reinforces the need for the evidentiary-tier work proposed here. The point was that telemetry will only carry weight for attribution, licensing, and compensation if usage reporting can be audited or independently checked.

One useful nuance from that feedback: telemetry should be an evidence layer for a range of licensing approaches, including broader enterprise or relationship-based licences. It should not imply a default move towards granular transactional pay-per-use.

For this issue, that argues for evidentiary tiers as a property of signal quality and verification, not as a counting or pricing model. The tiers can make compensation-grade uses more credible without deciding the commercial formula.

[jchomat]

The evidentiary-tier framing addresses what I see as the core accountability gap in v0.1: self-reported telemetry is a claim, and compensation infrastructure that treats claims as facts will eventually be gamed. The three-tier model is the right structural fix, and I support the PR offer.

On Q1: explicit field vs. derived from properties

Derived-from-properties is cleaner conceptually, but explicit is more useful for consumers — especially smaller publishers without the engineering resources to reimplement tier derivation. I'd suggest both: scheme properties declare what a scheme can support (its ceiling), and an optional tier field on the fingerprint object lets the agent assert which tier it's claiming for this specific detection, constrained by the scheme's declared ceiling. This separates "what this scheme is capable of" from "what I'm asserting about this particular signal."
The risk of explicit-without-derivation: an agent claims Tier 3 with a scheme whose properties only support Tier 1. The risk of derivation-without-explicit: every consumer must reimplement the same logic. The combined approach avoids both.

On Q2: scheme registry

A lightweight informational registry (a well-known URL, not a normative gate) seems right for v0.1. Free strings without any registry create fragmentation — two independent implementers choose c2pa-text and c2pa_text and consumers can't correlate them. A SHOULD-level registry with a canonical list of known schemes and their declared properties resolves the naming problem without creating a governance bottleneck or blocking new schemes.

Rather than enumerating all schemes upfront, I'd suggest anchoring the registry on Tier-3 entries only — these are the ones that need stable names most urgently, since they're the schemes consumers will act on for compensation purposes.

One candidate for v0.1:
c2pa-text

• content_borne: true — C2PA A.7 in-band Unicode markers, not detachable metadata
• identity_bearing: true — signed publisher certificate
• verifiable: true — C2PA trust list + RFC 3161 timestamp
• Honest limit: does not survive paraphrase or summarisation

A second candidate is architecturally sound but not yet deployable:
iscc+registry — ISO 24138 fingerprint + timestamped first-publication hub. 
The ISCC-CODE itself (ISO 24138:2024) is live; 
the ISCC-ID hub that would add the verifiable timestamp is under active development 
and not yet part of the standard. 
Naming it here reserves the scheme string for when that infrastructure exists.

Tier-1 and Tier-2 schemes (IPTC GUID, DOI, standalone ISCC, watermarks…) are better as a community-maintained list outside the normative spec — the space is too large and fast-moving to close in v0.1.
One naming note worth encoding now: c2pa (file-embedded manifest, stripped by reprocessing) and c2pa-text (A.7 in-band) must be distinct entries. They have different content_borne values and therefore different max tiers — collapsing them would be misleading.

On Q3: Tier-2 and RSL

This connection is worth making explicit in the spec. If an agent retrieves content under an RSL licence that includes , a corroborated content_retrieved event is precisely what RSL mandates — it's the same signal, expressed in telemetry format. Tier-2 for content_retrieved could therefore be defined as: a matching origin-side event with the same content_telemetry_id exists or an RSL/CAP token exchange is on record for the same retrieval.
This grounds Tier-2 in existing access-layer infrastructure rather than requiring entirely new corroboration tooling.

Additional angle: compensation weight
One thing worth encoding — as informative guidance, not normative text (too early) — is that different tiers should carry different downstream weight:

Tier 1 signals: useful for trend analysis and dispute investigation
Tier 2: the reasonable floor for contractual compensation claims
Tier 3: supports fully automated compensation

From the perspective of European press publishers operating under the neighbouring rights framework (EU Copyright Directive 2019, transposed in France as droits voisins): the tier model maps directly onto the evidentiary requirements for asserting those rights.
You need to demonstrate that content was used, not merely fetched (gets you to Tier 2), and ideally that it's attributable to a specific rights-holder (Tier 3). Having shared vocabulary here would meaningfully accelerate licensing negotiations.

On the C2PA A.7 characterization vs. #5

@erik-sv's correction on in-band text provenance is important and worth tracking empirically.
The key question from a real publisher workflow: does the mark survive (a) CMS export to plain text, (b) copy-paste into a browser AI assistant prompt, (c) ingestion through an intermediary like Firecrawl?
If it does, the picture changes substantially from #5's assessment.

Happy to contribute test cases from typical French press CMS outputs (SPIP, WordPress, Arc XP) if that helps populate the survivability matrix.
Co-authorship on the PR: yes.

[jalexspringer]

Thanks @jchomat - adopting all three.

Ceiling plus assertion for the tier field. The scheme's declared properties set the ceiling, an optional tier on the fingerprint object carries the agent's assertion for that detection, and an assertion above the ceiling is invalid.

Registry: informational, SHOULD-level, a well-known URL, Tier-3 entries only. c2pa and c2pa-text are separate entries. iscc+registry is reserved and marked pre-standard until the ISCC-ID hub is live.

Tier 2: a matching origin-side event, or an RSL / CAP record of the same retrieval referenced through license_ref. This goes in the spec text.

Compensation weight stays out, including as informative guidance. A tier-to-weight mapping starts deciding the commercial formula, which the consultation feedback above warned against. Tiers let parties set the evidentiary bar for their use case; which tier a deal requires lives in the governing terms (terms_ref, #3). The droits voisins mapping belongs in a compensation-oriented profile, and would make a good first one.

@erik-sv - green light on the follow-on PR. The wire format on main is frozen for the comment window, so there is now a v0.2-draft branch: I'll review #10 and land it there this week (it's retargeted already) - point the follow-on at the same branch. Everything merges to main together after the window closes on 24 July. Co-authorship as offered, and yes to the French CMS test cases for the survivability matrix.

[CKBrennan]

Posting from issue #9 as suggested — we are the signals intermediary referenced there, and the provenance enum as drafted does not carry our case. Wanted to explain why before #11 is finalised.

The third_party_sourced value is designed for the scraper-resale flow in #5: a buyer agent retrieves content from a non-emitting intermediary, the content enters the buyer’s generation context, and the buyer emits content_grounded with provenance: third_party_sourced to flag that the content came via a reseller rather than directly from the publisher.

Our flow is structurally different in a way that makes any value on the existing enum insufficient:

Publisher article
↓
Overtone indexes it, derives signals (tone, brand safety, article type)
↓
Buyer agent calls get_signals → receives signal IDs and metadata only
↓
Content never enters the buyer agent's generation context
↓
No content_grounded event exists on the buyer side to attach provenance to
third_party_sourced is a value on a content_grounded event. In our flow there is no content_grounded event on the buyer side — because nothing was grounded there. You cannot add a provenance value to an event that does not exist.

The right model, as suggested in #9, is Overtone as a cooperating index emitter: we emit content_grounded ourselves at the point get_signals returns, with source_role: index. That part the current fields can carry.

What the fields cannot carry is the forward link from the grounding event to the derived signal that was served. Without that, the publisher sees “your article was accessed by Overtone’s index” but not “it contributed to signal X which was used for advertising decisioning by buyer Y.” The attribution chain is broken at exactly the point that has commercial value to publishers.

The field needed is something like derived_output_id — an opaque reference to the downstream artifact (signal ID, model output, etc.) the grounded content contributed to. Flagging here since #11 is under active drafting and this gap may affect the scheme-property model scope.

We can provide fixture examples for the signals flow if useful for the test suite.

[jchomat]

#9 already documents this gap thoroughly. From a publisher perspective, option A is the right answer architecturally: advertising usage and editorial usage are distinct boundary crossings, and the spec's design principle is to mark boundary crossings with dedicated event types, not to label existing ones.

One naming note: content_used_for_decision breaks the existing pattern (content_retrieved, content_grounded, content_cited, content_displayed, content_engaged — all content_ + single past participle). content_activated would be consistent with that pattern and maps naturally to the advertising concept of activation.

In my view, the evidentiary tier model and scheme-property work there apply to content_grounded events and are unaffected by whatever shape this new event takes. The two are independent — #11 does not need to wait on this, and this does not need to wait on #11.

[erik-sv]

Thanks @jalexspringer. I will proceed with the follow-on PR against v0.2-draft, building on #10, and implement the suggestions you added here.

@jchomat, thank you for the co-authorship offer. The SPIP, WordPress, and Arc XP cases would be useful inputs to the survivability matrix, particularly the CMS export, browser copy-paste, and intermediary-ingestion paths you identified.

On @CKBrennan's derived-signals flow, I agree with @jchomat that this is independent of the evidentiary-tier and scheme-property work. It describes a separate advertising or decisioning boundary rather than a change to the evidentiary properties of content_fingerprint.

I also agree that Option A from #9, a dedicated event type, is cleaner than overloading content_grounded, and that content_activated fits the existing naming pattern better than content_used_for_decision. I commented separately on #9 from the perspective of another signals provider with a similar content-to-signal-to-decision flow.

I will keep the #11 PR scoped accordingly and proceed with the approved follow-on work.

[pwright-bf]

We've submitted a related proposal in #18 — an optional verification profile with four conformance levels (V0 Seeded → V1 Attributable → V2 Reconciled → V3 Enforceable) that addresses the evidentiary-tier question raised here. The profile is additive and does not modify the wire format. Flagging here so the connection is visible.

[CKBrennan]

Thanks to everyone who weighed in — the convergence on content_activated and Option A is clear. Agreed that this is independent of #11 and should be its own PR.

@erik-sv happy to coordinate on the PR. We can contribute Overtone’s fixtures from our AdCP signals flow alongside Encypher’s Prebid/OpenRTB cases, and i imagine these converge more and more as Prebid manages the seller agents for AdCP. We can follow up directly to align on scope before drafting.

Will also look at #18 — the verification profile looks relevant to how signals-derived events would be treated once content_activated lands.

[jalexspringer]

The proposed v1 direction is to narrow this. Core would keep the assertion-scoped status names, emitter-requested/consumer-derived distinction and no-overclaim rule. Tier-2/RSL linkage and all evidence mechanics would move to the evidence profile.

I am also walking back my earlier proposal for a normative scheme registry. Scheme discovery should be a community-maintained informational list in the evidence profile, not a core registry. Origin or access evidence corroborates what the origin observed; for grounding or citation it may be one part of composite supporting evidence, not proof that the sensitive event occurred. Compensation weighting remains outside the standard. The scope boundary is in #26, and the split proposed for PR #21 preserves the larger body of work.

This issue should remain open while that split is tested. Erik, Jean-Christophe and the other contributors: please challenge the proposed core/profile seam, especially if a fixture demonstrates that the small core vocabulary or extension hook cannot support an independently implemented evidence profile.

[jalexspringer]

Disposition: superseded in part, adopted in part. Core v1 keeps the assertion-scoped status names, the emitter-requested/consumer-derived distinction and the no-overclaim rule. Tier-2/RSL linkage and all evidence mechanics move to the evidence profile, whose working group the Board has now authorised. The normative scheme registry is walked back: scheme discovery becomes a community-maintained informational list in the profile. Origin or access evidence corroborates what the origin observed; for grounding or citation it can be part of composite supporting evidence, not proof the sensitive event occurred. Compensation weighting stays outside the standard. Open until the PR #21 split is reviewed.

[jalexspringer]

Status correction so this thread reads true against the published text: core v1.0 carries content_fingerprint.scheme/detected/value with the non-escalation rule (section 6.4) and a generic evidence slot admitting detached references (section 6.8). The assertion-status vocabulary did not enter core - read both my earlier "adopting all three" comment and the status-names clause of the 12 August disposition as superseded by this note and the closing note on #8. The #21 split remains the path for the rest.