Grounding the Convergence IO oracle in real cross-border regulation

TL;DR — the grounding is real, but the strength varies sharply

The headline for your thesis: the primitive you emphasized — **NAP, "a denial overrides any capability claim" — is the most solidly grounded of all five, near-literally.** And it has the cleanest live feed. The oracle's strongest leg is exactly the denial leg.

Per-primitive validation

NAP (P2) — "denial overrides capability" — SOLID, near-literal ✅

US sanctions law is civil strict-liability: a prohibition stands regardless of the actor's intent, capability, or justification. OFAC's Economic Sanctions Enforcement Guidelines (31 CFR Part 501, App A) list "Willful or Reckless Violation" (Factor A) and "Awareness of Conduct" (Factor B) as determinants of the enforcement response / penalty, not elements of a violation — OFAC "does not need to prove fault or intent to enter an enforcement action." (Criminal IEEPA prosecution still requires willfulness; the civil strict-liability scoping is exact.)

The "**denial overrides capability unless an authorization/license exists" structure is the literal shape of the blocking prohibition:CFR 594.201(a) (and parallel .201 sections across parts 544/560/590…) — blocked property "may not be transferred, paid, exported, withdrawn or otherwise dealt in" "Except as authorized by … licenses or otherwise.**" Cite the program .201 prohibition, not App A (App A is the penalties framework; the verifier flagged this).

• Cross-border: OFAC (US) is one of several. EU consolidated list, UN Security Council

list, UK OFSI, and US BIS Entity List/EAR are parallel denial regimes (these non-US sources were not independently verified here — treat as well-established but unconfirmed in this evidence set). TheBIS rule making an SDN designation auto-trigger EAR license requirements (no license exceptions) is the same "denial propagates, can't be overridden" pattern — extracted but pending verification.

• Actionable goals: (1) screen every counterparty/destination against the consolidated

deny-list before any action; (2) treat a hit as a hard block that no capability/role can override; (3) record the license/authorization id when an action is permitted.

• Sources: eCFRCFRApp A, Cornell LII. Vote 3-0.

PCSF (P4) — SOLID for concentration/exit; softer for runtime failover ✅⚠️

• EU DORA (Regulation (EU) 2022/2554) — in forceJan 2023, applicableJan 2025.

Establishes an EU-wide oversight framework directly supervising critical ICT third-party providers; explicitly targets systemic/concentration risk from reliance on a few ICT providers. Arts 31–44 give a per-provider Lead Overseer (EBA/ESMA/EIOPA) with inspection powers and penalties up to 1% of worldwide daily turnover. FirstCTPPs designated Nov 2025 (AWS/Azure/GCP, Bloomberg, LSEG) — the concentration concern is operational.

• UK Critical Third Parties (CTP) — PS24/16 (FCA) + PS16/24 (PRA/BoE),Nov 2024,

in forceJan 2025, authority from FSMA(s.312L). Eight Operational Risk & Resilience Requirements (incl. R3 dependency/supply-chain, R8 termination) + six Fundamental Rules. Firm-level bite is designation-gated (HM Treasury designates; 12-month window).

• Honest stretch flag: PCSF's runtime **"circuit breaker / fallback chain / per-tier

quota"** is only analogical to DORA/CTP exit-strategy & substitutability — the regulatory concern is concentration/resilience/exit, not live runtime failover. The concentration mapping is direct; the runtime mapping is softer.

• Actionable goals: (1) maintain a per-provider register of ICT dependencies for

critical functions (DORA Art 28(3)); (2) maintain a documented exit/substitution plan per critical provider; (3) annual severe-but-plausible scenario test (UK CTP).

• Sources: EIOPA DORA, EBA DORA oversight, FCA PS24/16, BoE PS16/24. Votes 3-0.

AAPF (P3) — append-only provenance — SOLID ✅

Four independent, in-force anchors, all 3-0:

• 21 CFR Part§11.10(e) (FDA): "secure, computer-generated, time-stamped audit trails

… Record changes shall not obscure previously recorded information … retained … at least as long as … the subject electronic records … available for agency review." Maps 1:1 to append-only/timestamped/reviewable provenance. Nuance: the text mandates non-obscuring of history, which is slightly weaker than "immutable/no-delete" — the primitive's framing captures intent but is marginally stronger than the letter.

• SEC Rule 17a-4 (2022 amendment) — retains WORM or a new audit-trail alternative

(all modifications/deletions, create/modify/delete timestamps, actor identity, enough to re-create the original). Compliance dateMay 2023, in force.

• GDPR Art 30 — records of processing: a specific, enumerable field set (purposes,

categories of data/recipients, transfers + safeguards, retention, security measures). UK GDPR retains it unchanged.

• EU AI Act Art 12 — purpose-bound traceability (not "lifetime logging" — see refuted),

with a prescribed minimum log field set for Annex III(1)(a) biometric systems.

• Sources: eCFR Part 11, SEC 17a-4, ICO Art 30, AI Act Art 12.

CCF (P4) — capability claim / attestation — REASONABLE, but time-sensitive ⚠️

Maps to the US software supply-chain attestation regime: CISA Secure Software Development Attestation Form (releasedMar 2024), grounded in EO→ OMB M-22-18 / M-23-16 → NIST SSDF (SP 800-218). Producers of federal-government software must attest (prove a claimed secure-development capability) or file a POA&M — the "prove capability at action time, honesty floor" shape, near-literally.

CURRENCY RISK (critical, flagged by the verifier): OMB rescinded M-22-18 and M-23-16 onJan 2026 in favor of a risk-based approach. The historical/structural grounding is accurate, but the underlying mandate is no longer in force as of the 2026-06-25 research date. An oracle MUST track this — it's the live example of why regulatory grounding has to be polled, not memorized.

• Sources: CISA attestation form, NIST EOSSDF. Vote 3-0 (with the currency caveat).

DCF (P1) — data classification — STRETCH / aspirational ❌ (unverified)

No surviving verified claim grounds DCF. Specifically, the "label propagates through transformations" concept has no regulatory anchor in this evidence set. Regimes that require classifying data at rest exist (GDPR Artspecial-category, ISO/IEC 27002:2022 control 5.12, NIST SP 800-60 / FIPS 199), but a mandate for a propagating, machine-readable label that survives transformations was not substantiated. This is the weakest mapping — honestly aspirational, not yet grounded. (Open question below.)

The live oracle feeds — what keeps it current

The meta-question — can a cross-border oracle stay live? — is YES for the denial and EU-law layers, because authoritative machine-readable feeds exist (verified 3-0):

• OFAC Sanctions List Service (SLS) — launchedMay 2024, sanctionslist.ofac.treas.gov.

SDN + Consolidated lists in XML/CSV/fixed-field, the UN/Wolfsberg Advanced Sanctions Data Model (XML), plus a delta file for incremental updates. → grounds NAP.

• OpenSanctions (data.opensanctions.org) — I verified this live separately: a

consolidated feed (OFAC + EU + UN + UK + more), ~100k entities, updated daily, keyless. The single best cross-border NAP oracle.

• EUR-Lex — a SOAP web service (XML metadata/search) + the Cellar RESTful API

(full-text retrieval by UUID). → grounds DORA, GDPR, AI Act text. (Note: a 10,000-results-per-search cap takes effectJan— a bulk-ingestion tweak.)

Feasibility verdict: keeping the oracle current is feasible for NAP + EU-law (real feeds). Decomposing statute into machine-actionable goals is still partly human-in-the-loop rule-authoring — no regulator yet ships a structured "obligations" dataset rich enough to auto-decompose (open question).

What was refuted, and what's still open

Refuted (1, vote 1-2): "AI Act Artmandates automatic lifetime logging." — Overreach. Artrequires purpose-bound traceability + a prescribed minimum field set, not unbounded lifetime logging. The primitive holds; the maximalist framing does not. This is the discipline working — the verifier killed the one claim that overstated.

Open questions (carry-forward):

1. Can DCF be grounded at all? Does any in-force regime mandate a propagating label

(vs. classify-at-rest), and is it exposed as a schema an oracle could ingest (ISO catalog, NIST OSCAL)?

1. NAP cross-border completeness: do the EU FSF / UN / UK OFSI / BIS feeds reconcile with

OFAC's ASDM identifier scheme cleanly enough for a unified deny-list without false matches?

1. Given the OMB rescission, what's the new in-force basis for CCF attestation?
2. Is the decompose-to-machine-actionable-goals step automatable, or fundamentally manual?

Wiring it — the RegulationChannel oracle

This converts directly into a grounding channel for the Question Loop, exactly like the flourishing WebChannel:

• SanctionsChannel (NAP) — polls OpenSanctions / OFAC SLS; resolves "is this entity

denied?" with a hard block + source. The highest-confidence oracle leg.

• EurLexChannel (DORA/GDPR/AI Act) — Cellar API by CELEX id; grounds AAPF/PCSF/CCF

obligations in current statute text + its last-modified date (catches rescissions like OMB's).

• CAP/NAP gating still applies: the regulation oracle informs the deny-list; the existing

AuthorityGate enforces it. Denial overrides capability — now grounded in the realCFR .201 structure it was always modeling.

Methodology + honesty note

Run by the deep-research harness:angles →sources fetched →claims extracted → topverified by 3 independent adversarial voters (2-of-3 refutes to kill) →confirmed,killed →synthesized findings. First run reported "allrefuted" — but that was a rate-limit artifact (the verifier hit the account session limit and every vote came back 0-0/abstain, which defaulted to "refuted"). That false "all refuted" is itself a textbook ungrounded-collapse failure: a system reporting confident verdicts while actually blind. The resume (after the limit reset) re-ran only the verify stage and produced the real 24/25. The lesson is the certificate's: trust the verdict only when the verifier actually ran.

Source of record: deep-research run wf_70e7797a-28e, 2026-06-25. All findings 3-0 unanimous on primary sources; cross-border completeness and DCF remain the honest gaps.