At the IETF 127 Hackathon (San Francisco, 14–15 November 2026), several projects working on verifiable evidence of AI agent actions asked other people to write verifiers from their documents alone. Each implementation passed its own tests. The failures appeared between implementations. This post collects the concrete ones, checked against the slides and drafts, and notes what they mean for VAP, the metaframework VSO maintains.

1. One number, two verdicts

The Action Evidence Packages project reported that its own two reference verifiers disagreed on a single number: the Python one produced 1e-07, the JavaScript one 1e-7, and the verdicts came back ALLOW and DENY.

You can reproduce the serialization difference in a second:

>>> import json >>> json.dumps(1e-7), json.dumps(1e20), json.dumps(2**53 + 1) ('1e-07', '1e+20', '9007199254740993') 
> [JSON.stringify(1e-7), JSON.stringify(1e20), JSON.stringify(2**53 + 1)] [ '1e-7', '100000000000000000000', '9007199254740992' ] 

RFC 8785 (JCS) requires the ECMAScript number serialization, so the Python defaults above are not JCS output. The last pair is worse than a formatting difference: the integer cannot be represented exactly as an IEEE 754 double, so a JavaScript reader has already changed the value before any hash is computed.

2. Eight inputs that split four canonicalizers

The Agent Evidence Conformance Matrix (maintained by Sankalp Gilda) ran four JSON canonicalization implementations over the same corpus. They wrote identical bytes on 20,121 of 20,140 cases. The eight inputs that split them:

  • integers of 2^53 or more (three cases)
  • a byte order mark
  • duplicate members (two cases, one in a real receipt)
  • nesting deeper than 128 (two cases)

The slide's conclusion: a JCS profile has to say "refuse" for these. RFC 8785 limits its input to I-JSON (RFC 7493), which already rules out duplicate property names, but a verifier still has to decide what it reports when it meets one.

3. A receipt needs the issuing service's rule

Two operators registered the same four SCITT Signed Statements on their two Transparency Services and checked each other's receipts. The same statement produced a different leaf on each log: one hashes the statement's bytes, the other a digest of its Sig_structure. RFC 9943 and RFC 9942 do not name the entry serialization, so you cannot check a receipt without knowing the service's rule.

The authors list their limits: the operators co-champion the same project, the second log's signed tree head was not checked, and one side's checking script is unpublished.

4. A receipt does not sign its tree size

In an RFC 9942 inclusion receipt (RFC9162_SHA256), the signature covers the Merkle root; tree size and leaf index are in the unprotected header. A receipt read alone proves inclusion under a signed root, not position or log size. One service edited 8,255 (tree size, leaf index) pairs: a plain receipt was accepted for the issued pair and 31 others; with the signed tree head, only the issued pair. An erratum was announced on the SCITT list (filing not confirmed by us).

5. Re-encoding changes the digest

At the SCITT working group on 18 November, one slide summarised an experiment: a single Signed Statement re-encoded 64 ways with CBOR's encoding freedoms gave 64 different hash values, and a stock decoder rejected none of them. If you reference a record by digest, you have to pin both the bytes and the canonicalization. draft-mih-sokolov-cose-deterministic-encodings-01 does this for the COSE Hash Envelope preimage, for CBOR only ("dtrm"); pairing it with application/json is forbidden.

What this means for VAP

From the VAP v1.2 text:

  • §7.2: all numeric values MUST be encoded as strings. Section 1's split cannot happen for values carried as strings.
  • §7.3 / INT-001: canonicalization per RFC 8785.
  • INT-003: event_hash is computed over canonical JSON, so attaching a COSE receipt does not change it.
  • §11.3: VAP-native verification MUST remain possible without a transparency service.

Still open, and now recorded as unadopted candidate questions (CHANGE-002 preparation register, §5):

  • What a verifier reports for each of the eight inputs: reject, or not verifiable.
  • Whether the integer-typed fields of the AnchorRecord abstract schema are strings under §7.2.
  • What an Evidence Pack must carry to check a receipt: the service's entry rule, its receipt key, and a signed tree head when size or position matters.
  • Whether references between records should name the target's digest and algorithm, not only its identifier.

Fix the text first, then build the reproduction kit; otherwise the kit freezes one implementation's behaviour.

What VAP would not have done

None of these results would have changed. VAP does not define how a transparency service serializes entries, and it cannot make a receipt sign its tree size. None of the projects above is an implementation of VAP or VCP. VAP v1.2 §1.6(c): conformance does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records.

Disclosures

Zero external implementations of VAP or any VAP profile; zero Evidence Packs accepted in any proceeding (count dated 25 October 2026). No claims about contracts, customers or revenue. First-party work by VSO (operated by VeritasChain Co., Ltd.), prepared with Claude (Anthropic) as an AI agent, checked against the primary sources below; no external review. VSO authors draft-kamimura-vap-framework and has corresponded with one of the hackathon champions on IETF coordination.

Sources

Longer version: https://medium.com/@veritaschain/what-ietf-127-showed-about-evidence-someone-else-has-to-check-de2cf37ecb41