Compliance mapping — OCP mechanism to regulation

An engineering map from protocol mechanism to the regulatory checkbox it serves. Each row names a mechanism, where it is written down in this repository, and the duty a supervisor or auditor will ask about — so the person answering can point at a SPEC section instead of a slide.

What this document is not: legal advice, a certification, or a claim that using OCP makes a deployment compliant. It cannot be one. The duties below attach to processing, and the deployment's own processing satisfies them or fails to; the protocol contributes mechanisms, evidence, and wire-level disclosure, nothing more. Where a mechanism cannot help, the section after the table says so, in the same register as docs/dpia-template.md §3 — whose first row answers "control:" with "none".

Two scoping facts before the table:

  • The EU AI Act rows for Art. 12, Art. 14 and Art. 26 bind only deployments that are high-risk under Annex III (in application since 2026-08-02, as of 2026-08). Most OCP deployments — a shop's conversational widget — are not. For those, the rows describe hygiene a regulation happens to have words for; for the deployment that is high-risk (credit, essential services), the mechanisms are already on the wire. Art. 50 binds any system that talks to a natural person, risk class or none.
  • Every regulation reference carries as of 2026-08. The Swiss note at the end says why the date is part of the citation.

The mapping

MechanismWritten inDuty served (as of 2026-08)
Per-action audit events. Pair start/claim/approve/deny/revoke, companion link/takeover, every tool invocation, every confirm resolution — "MUST be auditable events", attestation recorded verified/unverifiedSPEC/core.md §14EU AI Act Art. 12 — the system technically allows automatic event logging; Art. 26(6) — the deployer keeps those logs; FINMA Guidance 08/2024 — documentation and inventory of AI use, for which the event stream is the evidence rather than the appendix
Expire-to-deny confirms on writes. Every write MUST be preceded by a confirm; a confirm without a response by expiresAt EXPIRES TO DENY; a pending diff whose Companion dropped MUST NOT auto-apply; "no affirmative action ever results from silence"SPEC/core.md §10, restated §14EU AI Act Art. 14 — human oversight as the default state, not a review screen bolted on: the system's answer to an absent human is refusal, so the ability to intervene is what the machine assumes, not what it grants
answeredBy on session.linked and turn.done. {mode, operator, brainId?} — who runs the model, per settlement and per actually-answered turn, and the two MUST NOT be assumed to agreeSPEC/core.md §5.2EU AI Act Art. 50 — the visitor can be told they are talking to a machine, and whose; FINMA Guidance 08/2024 — explainability starts with naming the model that answered this turn, not the one the contract mentions
The deployment declaration in the discovery document. The deployment field of /.well-known/ocp.json (schema wellknown.ts; discovery SPEC/companion.md §1) — OPTIONAL, additive within wire 1.0; who operates the Hub, in which shape. Absence means only that the Hub predates the fieldSPEC/companion.md §1nFADP Arts. 19–21 duty to inform — the transparency notice's operative facts, machine-readable at a fixed URL instead of prose that decays; self-hosted, it is also how a deployment evidences the privatim posture (the Swiss cantonal data protection commissioners): data and keys under the controller's own control, stated where it can be fetched
The DPIA templatedocs/dpia-template.mdnFADP Art. 22 / GDPR Art. 35 (the template says revFADP — same act) — it exists so the assessment is of the deployment actually running, not of the one the marketing page describes; question zero is which of two opposite privacy shapes you are
Self-hostability. deploy/terraform is a runnable reference under your own credentials — nothing calls home, no account to create; images publish per version tag only, no floating latest, both Hub shapes from the same treedeploy/terraform/README.mdFINMA Risk Monitor 2024 — outsourcing and concentration risk shrink when the trust broker for every session is a box you can inspect, move or destroy yourself; procurement that must exclude US CLOUD Act reach can, because no mandatory third-party operator exists to be reached through

What this does not cover

The rows above are what the protocol can carry. Three things it cannot, stated here because a mapping that omits them stops being an engineering document:

  • The protocol cannot make a model accurate. answeredBy names who to hold accountable for an answer; nothing on the wire makes the answer right. EU AI Act Art. 15 accuracy (as of 2026-08) is the model's and the deployment's to satisfy, and a protocol that implied otherwise would be making one of the claims docs/dpia-template.md §4 forbids.
  • The protocol cannot retain logs the operator deletes. SPEC/core.md §14 makes the events auditable; §11 leaves retention to tenant policy on purpose. Art. 26(6)'s minimum retention (as of 2026-08) is a duty on the operator, and the wire has no view of a deletion that happens after it.
  • The protocol cannot verify a deployment's declaration. The discovery document is self-asserted JSON over TLS: TLS proves who serves it, never that it is true — the same epistemics as the site key in the DPIA template, which identifies and does not authenticate. An auditor gets a claim to check, which is more than no claim and less than a proof.

Swiss note: this table is written against moving law

Switzerland signed the Council of Europe AI Convention in March 2025, and the Federal Council's implementation bill goes to consultation by end-2026 — as of 2026-08 it has not landed. The bill targets transparency, oversight and supervision, which are precisely the three duties the rows above map mechanisms onto. So the citations column of this table has a shelf life the mechanisms column does not: that is why every regulation reference here carries its date. When the bill lands, re-check the citations; the mechanisms stay what they are.