You've already forked mi-apmainas-parvaldiba
Import UAPF package
64ecc3f48d
BPMN admission (AL0 -> AL2), four DMN decision tables and a CMMN case for suspension, withdrawal and appeal. The three rules that answer structural objections structurally: - 17.2 circularity is a precondition, not a warning (avota-neatkariba). A participant cannot be the source of its own admission evidence, including where it commercially supplies company-register data. - 15 adverse decisions fail closed: exhaustive grounds, 30-day notice, cure, 5-working-day emergency review, two distinct officers, conflict-free appeal. - 6.4 a legally significant answer is never T0. Not an AI system (Reg. (EU) 2024/1689 Art. 3(1), Recital 12) - it does not infer. It is the governance of one, and therefore in MIC sandbox scope. resources/mappings.yaml traces every clause to its artefact and names the two gaps the model cannot close alone.
42 lines
3.7 KiB
YAML
42 lines
3.7 KiB
YAML
# Non-normative supplementary file. UAPF does not cornerstone guardrails;
|
|
# they live under resources/ as a host-readable policy snapshot of the
|
|
# normative rules in the AI LV Exchange MCP Node Profile v0.2, Part II.
|
|
authority: lv.pppa.scheme-authority
|
|
version: "0.1.0"
|
|
profile: "AI LV Exchange MCP Node Profile v0.2"
|
|
|
|
separation_of_powers:
|
|
- "PPPA does NOT attest to legal facts. It relies on Uznemumu registrs and records what it relied on, and the date. The output of admission is a reliance statement, not an attestation (14.2)."
|
|
- "PPPA verifies only what is technically reproducible: domain control and key control. Anyone may re-run either challenge and obtain the same answer (14.2)."
|
|
- "A technical service provider acquires no trust authority by supplying software or infrastructure (14)."
|
|
- "The Conformity Assessor function MUST be separated from the Scheme Authority before the Exchange leaves sandbox status, or before it admits its third participant, whichever is first (14.1)."
|
|
|
|
circularity:
|
|
- "A participant MUST NOT be the source of the evidence used to admit it (17.2)."
|
|
- "Legal facts MUST come from Uznemumu registrs, never from the applicant's own data or products - including where the applicant is itself a commercial provider of company-register data (17.2)."
|
|
- "This is enforced in dmn/avota-neatkariba.dmn as a precondition, not as a warning. The default rule rejects."
|
|
|
|
assurance:
|
|
- "AL0 is self-declared and MUST NOT be presented as verified or accredited (6.1, 17)."
|
|
- "AL0 can never be accredited. No credential is issued below AL2."
|
|
- "Assurance level (how thoroughly the participant was verified) and access tier (what the service demands) are different axes and MUST NOT be conflated (6)."
|
|
|
|
access:
|
|
- "The tier is set by the consequence of relying on the answer, not by the secrecy of the datum. A legally significant verification is never T0 (6.4)."
|
|
- "MCPF L2 is where the challenge endpoint appears. A node serving T1 at L1 has an identity claim and no identity proof (6.3)."
|
|
- "Fail-closed: an unrecognised tier resolves to T3 requirements, not T0."
|
|
|
|
adverse_decisions:
|
|
- "The grounds in 15.1 are exhaustive. No other ground is valid."
|
|
- "No single individual may withdraw or suspend accreditation. Two distinct authorised PPPA officers are required (15.4)."
|
|
- "Ordinary withdrawal requires 30 days written notice and a cure period (15.2)."
|
|
- "Emergency suspension is narrow: confirmed key compromise or an active security incident affecting federation trust. It MUST be reasoned in writing and reviewed within 5 working days (15.3)."
|
|
- "Appeal goes to a body excluding any person with a commercial interest in the participant's node or in a competing implementation (15.5)."
|
|
- "Every precondition is checked BEFORE any side effect. Where a side effect is unavoidable, it runs fail-closed: the credential is revoked first, so a later failure leaves the participant untrusted rather than trusted."
|
|
|
|
eu_ai_act:
|
|
classification: "Not an AI system. This package is deterministic governance - rules defined by natural persons, executed without inference. Regulation (EU) 2024/1689 Art. 3(1) and Recital 12."
|
|
scope: "In scope of the MIC regulatory sandbox because it governs admission to the trust and identity plane of the Process-as-Code AI system, which is the declared AI system."
|
|
human_oversight: "Every adverse decision has at least two mandatory human tasks (the two-officer decision, and the appeal). No adverse decision is ever automated."
|
|
logging: "Each decision emits a signed CloudEvent. The onboarding record MUST state which checks PPPA performed, which external sources it used, which claims PPPA makes, and which remain the participant's responsibility (17.1)."
|