Records cite evidence by URI alone, so an appraisal cannot be reproduced against what was appraised

[opento-suggestions]

Spec section affected
§3.1 (Logical schema), §3.2 (Wire format), §3.3 (Verification), and §4.3
(Bindings TRACE adds), as carried by schema/trace-claim.json.

Problem

The observation

Four citation fields in the Trust Record carry a bare URI and no digest, so
what a record cites cannot be checked to be what was appraised.

At 5c69dc1, schema/trace-claim.json carries
four citation fields that are a bare URI with no digest member beside them:
appraisal.policy_ref ("URI to the appraisal policy used."),
transparency (the SCITT receipt URI), runtime.rim_uri, and
model.aibom_uri (appraisal.verifier is also a bare URI but names the
verifier rather than an artifact, and is not counted). Three surfaces pair the URI with a
digest: policy.policy_uri beside policy.bundle_hash,
tool_transcript.transcript_uri beside tool_transcript.hash, and
build_provenance.provenance_uri beside build_provenance.digest; the
references[].digest member is optional, "when the producer holds it at issue
time". The four are top-level fields of the record, not entries in
references; the rules #226 quotes for that mechanism
do not reach them. appraisal.policy_ref appears in spec/trace-v0.2.md once, at line
241, as an example value.

The one bound object shows the spec already knows how: policy.bundle_hash
is required, patterned ^sha(256:[0-9a-f]{64}|384:[0-9a-f]{96})$, described
as "SHA-256 or SHA-384 digest of the policy bundle in force at execution
time", and the spec prose states the binding in its own words (line 590):
"TRACE adds the binding: the policy bundle hash is sealed to the TEE
measurement". #276 (opened 2026-09-03) is doing the same work
for delegation.parent_record_hash, stating a preimage the schema alone left
open.

The consequence

A second verifier resolving the same URI at appraisal-check time verifies
whatever is deployed at that moment, not what was appraised. Where
resolution fails outright, #190 and now
#279 govern what the verifier records; where it
succeeds, nothing at this head makes "the same object" checkable. The tree
states the mechanism for schemas in its own words:

"A validator built from the bundle file alone resolves that by fetching the
published schema over the network, so it checks a nested statement against
whatever is deployed rather than against the file next to it, and with no
network it is unresolvable."
(tests/test_safe_integer_range.py lines 607-610 at
5c69dc1)

Where this has been raised

• [pre-v1.0] Verification depth + action_receipts: two-axis verification model #66, comment of 2026-08-18: appraisal "cites the
  policy that produced its verdict by bare URI: no digest, no resolution step".
• The spec: anchor key revocation to log entry ordering, not iat (#67) #187 review, 2026-08-18: "a verifier that cannot
  resolve a cited object reports a lower depth, or prose, or a third
  classification, or nothing".
• spec: where a verifier records an unresolvable citation: the cross-surface question from #187's review #190 holds the recording-surface half as the deferred
  marker; this question is the binding half, one layer below it.
• spec: disclosed gaps in a receipt chain, and the receipt_gap_disclosed outcome (#117) #260's receipt-outcome vocabulary (disclosed absence
  as a first-class outcome) is the pattern an answer here would join, now
  extracted as a general discipline in Absence of evidence needs a name on every verifier surface: six threads, one primitive #279.
• Absence of evidence needs a name on every verifier surface: six threads, one primitive #279 (2026-09-03) collects the recording half across
  six threads: what a verifier says when it could not establish something.
  This question is one layer below it: what makes "established" reproducible
  against the cited object. An answer here gives its
  spec: where a verifier records an unresolvable citation: the cross-surface question from #187's review #190 row a determinate object to record.
• spec: parent_record_hash's preimage is stated only in a document that binds nothing #245 holds the preimage of parent_record_hash
  (ruled 2026-09-03; normative text in spec: state what delegation.parent_record_hash is a digest of #276); it does
  not hold this question.
• Reproducibility claim: deterministic re-execution as positive evidence for software-only producers #249 holds reproducibility-as-evidence for
  software-only producers; its thread states its own boundary against
  spec: where a verifier records an unresolvable citation: the cross-surface question from #187's review #190's enum; it does not hold this question.

Proposed change

Decide one of three, on the record: digest binding of cited evidence is a
norm in v0.2; it is a profile option; or it is explicitly out of scope. Any
of the three closes this issue. Nothing is proposed beyond the decision: no
schema change is attached, and this is not a defect report. It is not
#279's question either: that tracker names how
absence is recorded, per surface; this one asks whether presence can be
bound.

Backward compatibility

• Non-breaking (new optional field, informative addition)
• Breaking (removes or changes required field, changes wire format, changes algorithm set)

Motivation

#271 made appraisal outcomes reproducible from
retained facts, and #279's fifth discipline (no
positive verdict by default) is the same rule one layer up; the citations
those outcomes rest on are the remaining unbound layer.

Measured at 5c69dc1, 2026-09-03T23:02:28Z: 21 open
issues and 6 open PRs, none holding this question.

[lywinged]

Confirmed at 5c69dc1: four citation fields whose target has no digest in the record, three whose target does (policy_uri with bundle_hash, transcript_uri with hash, provenance_uri with digest), references[].digest optional, policy_ref in spec/trace-v0.2.md once, in the example record at line 241, and the two passages quoted match that file at line 590 and tests/test_safe_integer_range.py at lines 607 to 610.

Nothing under src/ in either repository resolves the four. Here the names appear in the packaged schemas, in models.py, and in the two adapters: both write policy_ref and transparency, sandbox.py also writes rim_uri, nothing writes aibom_uri, and sign.verify_record() reads none of them. In agentrust-io/trace-tests at b2a6490, TR-APR-003 checks that policy_ref is an absolute URI, TR-RTE-003 that rim_uri starts with https://, TR-ANC-001 the shape of transparency, and no module reads aibom_uri. Run at that commit, an unreachable https:// value passes both TR-APR-003 and TR-RTE-003, since neither fetches. The one resolver in that suite is TR-POL-003, on policy_uri against bundle_hash, and docs/modules/tr-apr.md says why it is on that field and not this one: the record carries a digest for the enforced bundle and none for the appraisal policy, so there is nothing to compare a fetched policy against; the retarget was agentrust-io/trace-tests#69. The three platform pages describe a verify-hardware --check-rim flow that fetches from rim_uri; no such command exists under src/ at this commit.

The four name different kinds of target. The schema describes rim_uri as the Reference Integrity Manifest for this measurement; two of the three platform pages point it at a certificate-chain service instead, the AMD KDS and Intel Trust Authority / PCCS, and only the NVIDIA page names a RIM and compares the firmware measurement against it. #277 §2, still a proposal, puts accepted_measurements in verifier context. transparency carries a different open question, which entry the URI names: that is the half of agentrust-io/trace-tests#92 that stayed open, §3.1.2 sends the field's resolvability to §7 open question 3, and that question is about which log, and TR-ANC-002 already replays the inclusion proof from a supplied receipt without resolving the URI. Of the four, appraisal.policy_ref and model.aibom_uri are the two with no second handle in the record: transparency has the receipt TR-ANC-002 replays, and rim_uri has runtime.measurement.

references[].digest is the one digest-beside-pointer slot the record has: optional, on rel entries, with §3.1.2 rule 3 keeping a resolved reference from counting as evidence. The four are fields of the record and not entries, as the issue says, and #226 tracks the rel registry.

The sibling repositories read none of the four either: in agent-manifest at 9348097, cmcp at 7178c40, ca2a at cfa4529 and agentrust-telemetry at 29c13cf, policy_ref, rim_uri and aibom_uri occur only in cmcp's schema copy and one test, and no code reads a record's transparency. Not measured here: verifiers outside the organisation, because nothing in these repositories names one.

Tool-assisted: the checks and this write-up.