Challenge the drafts. Every comment gets an answer.
You do not need to join OCC to read, challenge or review its standards. Comment on a specific provision, bring evidence, propose replacement text. Each substantive comment receives a recorded disposition and a reason — the same discipline public standards bodies use for wide review.
What you can comment on
Six rules for comment resolution
Comments attach to a document, a version and a section. A comment that cites the text can be resolved; a general impression usually cannot.
Each substantive comment receives one of six recorded outcomes, with a reason. Silence is not a disposition.
A single well-evidenced objection outweighs many unsupported endorsements. Comments are weighed on evidence, not on who sent them.
Comments are published in the register only with the reviewer’s permission, and attributed only if the reviewer asks to be named.
Reviewers state any commercial or institutional interest. A declared interest does not disqualify a comment; an undeclared one weakens it.
Every change to a published document is recorded in the change log with its date and the comment, if any, that prompted it.
File a provision-level comment
The six dispositions
Issues and responses
The register opens with issues OCC’s editors raised against their own drafts. No external comments have been dispositioned yet; they will appear here, with permission, as they are resolved.
Issue. The upper horizons describe systems that do not exist. Presenting them alongside observable horizons risks implying they are empirically established.
Response. H10–H12 are labelled “Exploratory · speculative” wherever they appear, and the indicative thresholds for them are marked speculative.
Issue. Threshold claims cannot be made until T1–T12 are defined.
Response. T1–T12 published as AIT-12 v0.1, a draft for public comment. The matrix on the AIH-12 page is now a summary of it.
Issue. An earlier proposal used an E0–E5 evidence scale, which would conflict with the planned evaluation profile’s own scale.
Response. Resolved by AIE-12 Autonomy & Impact Evidence v0.1: six evidence classes and twelve profiles, E1–E12, one per threshold. The E0–E5 proposal is withdrawn.
Issue. “OCC” is also the US Office of the Comptroller of the Currency.
Response. The full name, Open Conformance Coalition (OCC), is used on first mention on every page.
Issue. SO-12 (five bands, H0–H4) and AIH-12 (H1–H12) described the same idea in incompatible ways.
Response. SO-12 is superseded by AIH-12 v0.1. The old page and PDF permanently redirect, and the AIH-12 page carries a “formerly published as SO-12” note.
Issue. Some pages and one analysis article still described T-RUE and R-EAL as working drafts, contradicting the status table.
Response. Corrected everywhere: both were published as v1.0 on 5 September 2026.
Issue. Some copy implied August 2026 high-risk deadlines that the Digital Omnibus has moved.
Response. Crosswalk notes Annex III obligations from 2 Dec 2027 and Annex I from 2 Aug 2028; Art. 50 transparency has applied since 2 Aug 2026.
Issue. Two competing versions of the “5 As” were in circulation, and neither was published, so the term could not be cited.
Response. One canonical set published as v0.1: Authority · Access · Activity · Action · Accountability, mapped to GUARD, PACTE and AIT-12. It carries no conformance level and does not replace GUARD.
Issue. The evaluation profile was listed as “name under review”, which left the family incomplete and the E-codes undefined.
Response. Named Autonomy & Impact Evidence, completing the pattern Horizons → Thresholds → Evidence (Hn → Tn → En).
Issue. The family could be read but not applied: there was no way to classify a system, record the result in a portable form, or test the whole chain on a real deployment.
Response. Published an indicative classifier (labelled not certification), the OCC Agent Record with a JSON Schema, machine-readable framework data, and the Agent Assurance Package with terms for a bounded pilot that publishes its failures.
Issue. An Agent Record had no way to show how strong its evidence was, when it went stale, or what happened when the agent failed. The three profiles that would answer this were listed only as “planned”.
Response. Published the Evidence & Attestation profile (A1–A3 attestation levels, SHA-256 digests, mandatory scope of what was not checked), the Agent Lifecycle profile (material-change triggers, reassessment clock, suspension and retirement; no status is final) and the Incident & Disclosure protocol (S1–S4, S1/S2 always public, lessons before close). All v0.1 drafts for public comment.
Issue. Incidents could not be coded against the OWASP Agentic Top 10 or shared incident-registry fields. Agent identity and multi-hop delegation were not aligned with NIST NCCoE work. Any self-declared record went straight onto the public register with no check. New visitors had no guided path.
Response. Added the OWASP ASI01–ASI10 crosswalk, optional ASI and AIR mechanism fields on incidents (schema, portal and public record), NIST NCCoE identity alignment and a multi-hop delegation note. Added a separate reviewer role and a listing check: a secretariat-appointed reviewer confirms each record is complete and not misleading before it appears on the public register. The listing check is not a review of the agent. Added a role-based Start here page and a portal onboarding checklist.
Issue. Signatures were proposed but nothing was signed, so a copied record could be altered without detection. Digests had to be computed outside the portal. Valid-until and withdrawal were rules on paper with no field to record them.
Response. The register now signs every public Agent Record and attestation (Ed25519 JWS over canonical JSON), publishes its key set at /occ/jwks.json and offers a verify page. A register signature proves the register published the content unaltered; it is not endorsement. Issuer signatures stay optional pending comment on issuer key discovery. The portal computes SHA-256 digests in the browser without uploading files, records method and valid-until, and withdraws evidence with a reason that stays on the public record.
Issue. The framework said nothing about agents with bodies: robots and humanoids driven by the same models. Physical actions are harder to reverse, and functional-safety standards do not cover authority, evidence or revocation.
Response. An Embodied & physical agents profile is added as planned. It applies the existing family (AIH-12, AIT-12, F-ACT GUARD, T-RUE, Incident & Disclosure) to physical agents, adds no new framework or levels, and states that it complements functional-safety standards (ISO 10218:2025, ISO 13482, ISO 25785-1 in draft) and regulation (EU Machinery Regulation, EU AI Act) rather than replacing them. Robotics safety reviewers are invited before drafting.
Issue. Register signatures used Ed25519 only. Agent records may need to be checked well into the 2030s, after NIST (draft IR 8547) and the Australian Signals Directorate expect traditional public-key signatures to be phased out. There was also no published key rotation policy. Separately, a "quantum AI" profile was suggested.
Response. The register now adds an ML-DSA-65 signature (NIST FIPS 204, August 2024) over the same canonical bytes as the Ed25519 one; either verifies the content. Both public keys are published at /occ/jwks.json and the verify page reports which method it checked. A key policy is published: rotation triggers, retired keys kept so old records stay checkable, and handling of suspected compromise. Long-term key custody, including a possible move to an independent foundation, is an open decision. No separate quantum AI profile is added: OCC classifies agents by capability, authority and impact, not by the hardware they run on.
Issue. An accuracy pass against current sources found: (1) the Embodied profile said an AI safety component of machinery is high-risk under the EU AI Act, but the adopted Digital Omnibus took machinery out of that direct route; (2) the crosswalk did not mention Colorado’s replacement law or the 2 December 2026 grace date for marking AI-generated content; (3) three OWASP Agentic Top 10 names did not match the published list; (4) the NIST note named the Model Context Protocol among the concept paper’s sources. A linked whitepaper also said the original Colorado AI Act took effect in June 2026.
Response. Machinery wording corrected to the adopted Regulation (EU) 2026/1744. Colorado SB 26-189 (signed 14 May 2026, effective 1 January 2027) added to the crosswalk, and the whitepaper corrected. The 2 December 2026 date for systems already on the market added. ASI03, ASI04 and ASI08 now use the published OWASP names. The NIST note now cites the concept paper (5 February 2026, comments closed 2 April 2026) and the standards it names.
Issue. The reported Medicare data-matching disclosure showed how a report can stall at the first hop: an AI developer emailed a general Services Australia inbox on 10 September 2026, it was read on 11 September, the Australian Signals Directorate was told on 15 September and the Minister on 17 September. Agent records had no field for where a report about an agent should go. Separately, incident fields did not line up with the OECD incident/hazard definitions or with legal reporting clocks such as EU AI Act Article 73, so operators had to re-code incidents before reporting them.
Response. Organisations can now declare an incident and vulnerability disclosure contact (email or web form, optional phone and policy). It is shown on each public register entry, carried in the Agent Record as disclosure_contact and listed at /occ/disclosure-contacts.json; a reviewer can mark it as checked, and changing it clears the mark. Incidents gain optional event_type (incident or hazard), OECD-aligned harm type, affected stakeholders and AI-system role, jurisdictions, and a regulatory block recording the regime, Art. 73 class, awareness date, report dates and regulator reference. OCC records whether a legal duty applies and when the operator reported; it does not calculate deadlines, produce filings or file reports. Where the law sets a deadline, the law applies.
Issue. From 10 December 2026 the Australian Privacy Act (APP 1.7–1.9) requires privacy policies to describe the kinds of personal information used and the kinds of decisions made solely, or substantially and directly, by a computer program where they could significantly affect people. EU AI Act Article 50 transparency duties have applied since 2 August 2026. Agent records had no place to point to these disclosures, agent by agent. Separately, there was no way to follow changes to the public register without re-reading it.
Response. The Agent Record gains optional adm (solely automated decision types, substantially related decision types, personal information kinds, privacy policy link) and eu_art50 (AI-interaction disclosure, synthetic content marking) blocks. They describe kinds, never personal information; they point to the operator’s privacy policy and do not replace it; OCC does not review whether a disclosure is adequate. The public register now publishes a change feed of status changes and public incidents at /occ/records/feed.xml (Atom) and /occ/records/feed.json (JSON Feed).
Issue. Signed records prove what the register published for a record at the moment of signing, but not that the register has not since changed, back-dated or quietly removed an entry. A register that wants to be trusted by people who do not trust its operator needs a way for outsiders to detect rewriting.
Response. The register now keeps an append-only transparency log at /occ/log. Each listing, record version (by SHA-256 of the canonical record JSON), status change and public incident is written once as a canonical JSON statement and hashed into a Merkle tree using RFC 9162 hashing. Each new tree head is signed with the Ed25519 and ML-DSA-65 register keys. Inclusion and consistency proofs are public at /occ/log/proof.json, with a browser checker. Entries are never edited or removed; withdrawals and unlistings are new entries. Inclusion proves publication, not truth, safety or compliance. The log borrows from Certificate Transparency and IETF SCITT but is not a SCITT transparency service.
Issue. Many agents already publish an A2A Agent Card at /.well-known/agent-card.json describing their name, endpoints and skills. Operators would have to retype those facts into an OCC record, and someone holding only the Agent Card had no way to find the agent's OCC record.
Response. The portal can prefill a draft record from a pasted Agent Card (name, endpoint identifier, version, provider, description and skills), filling only empty fields and never inventing the rest; skills are copied as advertised capabilities, not authority. Each listed public record offers an A2A extension block at /occ/records/<ref>/a2a-extension.json, under the URI https://openconformance.org/occ/ext/agent-record/v1, carrying the record URL, its SHA-256 digest and signing key IDs. It is never required and adds no methods. A discovery convention at /.well-known/occ-record.json is proposed, labelled as a proposal and not registered with IANA. The bridge does not verify Agent Card signatures and does not grant, scope, delegate or revoke authority.
Issue. The schemas fixed the shape of a record but nothing showed implementers what a correct checker should accept or reject, and some rules (matching horizon, threshold and profile numbers; dates that run forwards) cannot be written in JSON Schema. Running a checker over our own fictional example record found ten schema errors.
Response. Published 20 fictional test vectors with expected verdicts at /occ/test-kit/vectors.json, ten profile rules with stable codes (OCC-R1 to OCC-R10: four errors, six warnings), and an in-browser checker that can fetch a listed public record by reference. The example record was corrected. The kit checks that a record is well formed and internally consistent; it does not check that the claims are true and is not certification.
What changed, and when
| Date | Document | Change |
|---|---|---|
| 2 Oct 2026 | OCC Agent Record v0.1; Incident & Disclosure v0.1 | Conformance test kit: 20 open test vectors, ten profile rules (OCC-R1 to OCC-R10) and an in-browser checker at /occ/test-kit. The fictional example record was corrected to match the schema (OCC-E-021). |
| 2 Oct 2026 | OCC Agent Record v0.1; Interoperability crosswalk | A2A Agent Card bridge: prefill a record from an Agent Card, an optional OCC extension block for Agent Cards, and a proposed /.well-known/occ-record.json discovery convention (OCC-E-020). |
| 2 Oct 2026 | Evidence & Attestation profile v0.1; public register | Register transparency log: append-only Merkle log (RFC 9162 hashing) of listings, record versions, status changes and public incidents, with dual-signed tree heads and public inclusion and consistency proofs (OCC-E-019). |
| 2 Oct 2026 | OCC Agent Record v0.1; public register | Optional automated-decision (adm) and EU Art. 50 (eu_art50) transparency fields, mirroring the Australian Privacy Act duty from 10 December 2026; register change feeds in Atom and JSON Feed (OCC-E-018). |
| 2 Oct 2026 | Incident & Disclosure protocol v0.1 | OECD-aligned coding: optional incident/hazard type, OECD-aligned harm and stakeholder fields, and a record of any legal reporting duty (such as EU AI Act Art. 73) and when the operator reported. Organisations can publish a disclosure contact on every register entry, in the Agent Record and in a public directory (OCC-E-017). |
| 2 Oct 2026 | Framework crosswalk; OCC-ASF v0.1 | Accuracy pass: EU AI Act machinery wording corrected to the adopted Digital Omnibus; Colorado SB 26-189 and the 2 December 2026 marking date added; OWASP ASI names aligned with the published list; NIST NCCoE note made precise (OCC-E-016). |
| 2 Oct 2026 | Evidence & Attestation profile v0.1 | Post-quantum readiness: every public Agent Record and attestation now carries an ML-DSA-65 (NIST FIPS 204) signature alongside Ed25519; both keys are in /occ/jwks.json and the verify page checks both. Key rotation policy published; long-term key custody remains open (OCC-E-015). |
| 2 Oct 2026 | OCC-ASF v0.1 | Framework map published (linked diagram and downloadable image). Embodied & physical agents profile added as planned, with a mapping to the existing family and the safety standards it does not replace (OCC-E-014). |
| 2 Oct 2026 | Phase 6 v0.1 | Register signing live: public Agent Records and attestations are signed (Ed25519 JWS), with a public key set and a verify page. Portal adds in-browser SHA-256 hashing, attestation export, valid-until and recorded withdrawal of evidence (OCC-E-013). |
| 2 Oct 2026 | Phase 5 v0.1 | OWASP Agentic Top 10 crosswalk, AIR/ASI incident fields, NIST NCCoE identity alignment and multi-hop delegation note. Assurance Portal adds a reviewer role, a listing check before public listing, an onboarding checklist and a role-based Start here page (OCC-E-012). |
| 2 Oct 2026 | Phase 4 v0.1 | Published for public comment: Evidence & Attestation profile, Agent Lifecycle profile and Incident & Disclosure protocol, with JSON Schemas. The Assurance Portal now records evidence digests and attestation levels, shows a reassessment clock, and logs incidents to the public register (OCC-E-011). |
| 2 Oct 2026 | Phase 3 v0.1 | Published for public comment: indicative AIH-12 classifier (not certification), OCC Agent Record with JSON Schema, Agent Assurance Package with bounded-pilot terms, and machine-readable framework data at /occ/framework.json (OCC-E-010). |
| 2 Oct 2026 | AIT-12 v0.1 | Thresholds T1–T12 published for public comment: controls, authority ceiling, GUARD level and reassessment interval per horizon (OCC-E-002). |
| 2 Oct 2026 | AIE-12 v0.1 | Named Autonomy & Impact Evidence and published for public comment: six evidence classes, profiles E1–E12. E0–E5 withdrawn (OCC-E-003, OCC-E-009). |
| 2 Oct 2026 | 5 As v0.1 | Canonical set Authority · Access · Activity · Action · Accountability published for public comment (OCC-E-008). Founding Review Cohort confirmed at 24 reviewers. |
| 2 Oct 2026 | OCC review | Public review register, comment form, change log and Founding Review Cohort opened. Assurance Observatory opened with entry OBS-001. AIH-12 Horizon Card v0.1 template published. |
| 2 Oct 2026 | T-RUE · R-EAL | Status statements corrected to “published v1.0, 5 September 2026” across all pages (OCC-E-006). |
| 2 Oct 2026 | Crosswalk | Added NIST NCCoE agent identity work, the reported FTC position on developer liability, and US state laws in Connecticut and Illinois. |
| 2 Oct 2026 | OCC-ASF v0.1 | Framework overview, status table, decision tree, glossary and “What OCC does not claim” published as v0.1 for public comment. |
| 2 Oct 2026 | AIH-12 v0.1 | Published for public comment. Supersedes the SO-12 working label (OCC-E-005). |
| 2 Oct 2026 | PACTE v0.1 | Published for public comment, with proposed evidence classes. |
| 5 Sep 2026 | T-RUE 1.0 · R-EAL 1.0 | Published as v1.0 under the royalty-free pledge. |
| 19 Aug 2026 | F-ACT 1.0 | Published as v1.0 under the royalty-free pledge. |
AIH-12 and PACTE Founding Review Cohort
24 reviewers in the first round — expressions of interest open. Drawn from:
- — Read one draft (AIH-12, AIT-12, AIE-12, PACTE, the 5 As or the Horizon Card) and file provision-level comments.
- — Optionally, apply the Horizon Card or PACTE to one system you know well and report where it broke.
- — Declare any interest. Take part as an individual expert unless your institution agrees otherwise.
- — A written disposition for every substantive comment.
- — Early sight of revised drafts before public release.
- — Named acknowledgement in the register, only if you ask for it.