Skip to main content

ADR-0159: PEP tier scores and sanctions defaults are our judgement, recorded as such

Date: 2026-08-01 Status: Accepted Deciders: Adrian (Soft4U BV), Claude Opus 5

Context​

pep_tiers.json and sanctions_defaults.json both carried:

"source": "ADR-008a Default Configuration",
"source_url": null

ADR-008a does not exist. This repository's ADR series is four-digit (ADR-0001 … ADR-0158); grep 008a returns nothing. Neither file cited any public standard.

That matters more for pep_tiers than a config-tidy framing suggests. It is decision-driving — eba_risk_matrix._score_pep_exposure reads it — and it is frozen verbatim into the append-only reference snapshots (scoring/reference_snapshot.py, ADR-0124). So every scoring run persists a content hash of a file whose stated provenance resolves to nothing, and an auditor replaying that run is pointed at a document that has never existed.

Two facts had been conflated under one source field:

  1. The taxonomy — who is a PEP, which sanctions lists exist. This is published law and published authority output. It has a real citation.
  2. The numbers — the 30/25/20/15 tier contributions and the 100/70/60/30 match weights. No public standard assigns these. AMLR defines who a PEP is and what enhanced measures apply; FATF R.12 likewise. Neither assigns a weight. These are our risk judgement.

Attributing our numbers to an external document — any external document — is the defect. Attributing them to a non-existent document is the same defect with the evidence removed.

Decision​

Split the two facts and cite each honestly.

Taxonomy, verified against the ingested corpus rather than from memory. The PEP definition is AMLR (Reg (EU) 2024/1624) Art. 2(1)(34), with family members at Art. 2(1)(35) and known close associates at Art. 2(1)(36); obligations at Art. 44-46; FATF Recommendation 12 alongside. Each of the eight tier keys now records the specific limb it derives from, in a sibling amlr_limbs map.

The issue that scoped this work cited Art. 2(1)(25). Reading the corpus showed the PEP definition is at (34). Writing from the issue text would have replaced a phantom citation with a wrong one — a real article that says something else, which is exactly the failure mode ADR-0121 exists to catch and which no referential check would flag.

Numbers, declared as ours. Both files carry scores_authored_by / weights_authored_by = "Trust Relay", a rationale, and a pointer here. The ordinal reasoning: proximity to sovereign decision-making power (head of state / senior government, 30) ranks above legislative and senior-military office (25), which ranks above regional and judicial office (20), which ranks above derived exposure through a relationship rather than an office (family member / close associate, 15). This is a defensible ordering, not a measured one, and it is recorded so an officer can disagree with it explicitly.

The payload shape is unchanged, deliberately. Provenance went into sibling keys (amlr_limbs, list_authorities), not into data. Two reasons, both verified rather than assumed:

  • eba_risk_matrix._score_pep_exposure reads pep_data[key] and float()s it directly — it does not go through ReferenceDataService.get_risk_score, which would have tolerated a nested {"score": N}. A first attempt at a nested payload broke it, caught by test_risk_config_service.py::test_modified_pep_tiers_changes_computed_pep_exposure.
  • test_reference_snapshot.py:35 pins sanctions_defaults as {"lists": [...ids...]}, and the ADR-0124 reference snapshots hash these payloads.

So every score and every list id is byte-identical to before. This ADR changes what the files SAY about themselves, not what they DO.

Consequences​

Positive​

  • An auditor replaying a scoring run reaches a real citation.
  • The line between published law and our judgement is explicit and per-field, so an officer can contest the numbers without contesting the taxonomy.
  • A standing test (test_reference_data_provenance_930.py) fails on any reference dataset citing a non-existent ADR — the class, not the instance.

Negative​

  • Provenance now lives in sibling keys rather than beside each value, which is less tidy than a nested payload. That is the price of not changing a hashed, directly-read payload.
  • The tier ordering remains a judgement. It is now a recorded, contestable judgement, which is better, but it is not evidence.

Neutral​

  • No scoring behaviour changes. No migration.

Alternatives Considered​

Alternative 1: correct "ADR-008a" to a real ADR number​

  • Rejected: the numbers were never derived from an ADR. Pointing them at a different document would preserve the false attribution and remove the evidence that it was false. The phantom citation is at least visibly broken.

Alternative 2: nest provenance inside data​

  • Rejected on measured grounds: it broke _score_pep_exposure (direct float()), and it would change the ADR-0124 snapshot hash for every historical comparison.

Alternative 3: drop the numbers and score PEP exposure as a flat binary​

  • Rejected: AMLR Art. 44-46 distinguish PEP categories in the measures they require, so a flat treatment would lose real signal. The gradation is defensible; only its attribution was wrong.