How to produce an AMLA RTS effectiveness pack

When you need this: You are an EU-supervised obliged entity (a credit institution, payment institution, EMI, or VASP) and the EU AML Authority (AMLA) wants to see, for one engine run, that your transaction-monitoring program is effective — the alert→case→STR funnel — and that your detection rules evidence the three AMLR (Regulation (EU) 2024/1624) effectiveness obligations: CDD (Art. 28(1), the CDD-information RTS), ongoing monitoring of the business relationship (Art. 26; CDD measure Art. 20(1)(f)), and targeted-financial-sanctions screening (Art. 20(1)(d)). The offline aml amla-effectiveness-report command (#528) rolls all of that into one frozen, deterministic artifact for your Model Risk Committee (MRC) / supervisory pack.

Prereqs: A spec whose EU rules carry the AMLA RTS citations in their regulation_refs (the bundled examples/eu_bank/aml.yaml is the demonstrator). One completed aml run so the run dir has cases/, decisions.jsonl, and manifest.json. metrics/amla_effectiveness.py::build_amla_effectiveness_report is the canonical, pure builder.

Time: ~5 min.

The pack is offline and advisory — it is a post-run report like aml model-inventory, NEVER in the engine run path. It derives every number from the run's own artifacts: the funnel + per-rule counts come straight from metrics.outcomes.compute_outcomes (so the AMLA pack and the existing outcomes pack never diverge), and the RTS coverage comes from each rule's regulation_refs. Nothing is fabricated — the run records STR filing (escalated_to_str) but not regulator acceptance, so str_acceptance is reported as not_tracked rather than invented. It is deterministic: generated_at is the run manifest's as_of, not a wall-clock read, so the same run yields byte-identical output.

---

Steps

1 · Annotate your EU rules with AMLA RTS citations

Each rule whose logic reflects an RTS effectiveness standard carries the matching citation in regulation_refs (these are the real AMLR / AMLD6 references — AMLR = Regulation (EU) 2024/1624; AMLD6 = Directive (EU) 2024/1640):

rules:
  - id: structuring_cash
    regulation_refs:
      - citation: "AMLR Art. 28(1)"        # CDD RTS
        description: "CDD RTS — customer-due-diligence measures the rule reflects."
  - id: pep_screening
    regulation_refs:
      - citation: "AMLR Art. 26"           # ongoing monitoring of the business relationship
        description: "Ongoing monitoring of the PEP business relationship (CDD measure Art. 20(1)(f))."
  - id: sanctions_screening
    regulation_refs:
      - citation: "AMLR Art. 20(1)(d)"     # targeted-financial-sanctions screening
        description: "Targeted-financial-sanctions screening obligation (CDD measure Art. 20(1)(d))."

The report maps a rule onto an AMLR article when one of the rule's citations contains the article's reference (AMLR Art. 28(1), AMLR Art. 26, AMLR Art. 20(1)(d)). A rule with no AMLA citation simply isn't counted toward coverage — it still appears in the per-rule funnel.

2 · Validate + run

aml validate examples/eu_bank/aml.yaml
aml run examples/eu_bank/aml.yaml --seed 42

3 · Produce the pack

aml amla-effectiveness-report examples/eu_bank/aml.yaml .artifacts/run-<TS> \
  --markdown amla_effectiveness.md

The command writes amla_effectiveness_report.json into the run dir (atomic write — a temp file in the same directory then os.replace, so an I/O error never leaves a partial artifact), an optional --markdown MRC table, and prints a summary:

AMLA RTS effectiveness eu_bank_aml
  alerts: 89  cases: 89  str_filed: 83
  alert→str: 93.26%  str_acceptance: not_tracked
  RTS coverage: 3/3 articles

Use --out <path> to write the JSON somewhere other than the run dir.

4 · Read the report

jq . .artifacts/run-<TS>/amla_effectiveness_report.json

{
  "spec_program": "eu_bank_aml",
  "generated_at": "2026-06-09T18:31:51.015515",
  "total_alerts": 89,
  "total_str_filed": 83,
  "alert_to_str_pct": 93.26,
  "str_acceptance_rate_pct": null,
  "str_acceptance_status": "not_tracked",
  "rts_coverage": [
    {
      "key": "cdd",
      "title": "Customer due diligence",
      "citation": "AMLR Art. 28(1)",
      "instrument": "AMLR",
      "covering_rule_ids": ["high_risk_jurisdiction", "structuring_cash"],
      "status": "covered"
    }
  ],
  "n_rts_covered": 3
}

Field	Meaning
total_alerts / total_cases / total_str_filed	The AMLA effectiveness funnel, derived from cases/ + decisions.jsonl.
alert_to_str_pct	Overall conversion — the canonical AMLA effectiveness ratio.
str_acceptance_status	not_tracked — the run records STR filing, not regulator acceptance. Wire your FIU-feedback loop separately if you need this.
rts_coverage[].status	covered (≥1 rule cites the article) or gap (none).
rules[].precision	Per-rule precision — null unless you supply case labels (precision needs ground-truth labels the run doesn't carry). recall is always null; it needs the ground-truth positive population, which the alert set alone can't tell you.

Note (2026-07-10): AMLA's statutory deadline to submit the three final RTS to the European Commission passed on 2026-07-10. This edition treats the submission as pending confirmation, not asserted — see the "Deadline passed" entry in docs/research/2026-07-regulator-pulse.md for the Commission's 3-month endorsement clock (2026-07-10 → 2026-10-10). Once endorsed, a rts_coverage[].status: gap stops meaning "consultation input still open" and starts meaning an implementation backlog against a locked-in binding standard — worth re-running aml amla-effectiveness-report against a fresh spec once the endorsement lands.

5 · Read it on the dashboard instead

• Framework Alignment (page 8) — for an EU program, an "AMLA RTS coverage" tab maps the spec's rule citations onto the three RTS articles with a ✓ (mapped) / ∼ (partial — cited but no evidence trail) / ✗ (gap) indicator and the covering rule ids. The mapping logic (dashboard/frameworks.py::build_amla_rts_alignment) is streamlit-free and unit-testable.

---

Verify it worked

Three checks:

1. The JSON exists in the run dir and n_rts_covered matches the number of RTS articles your rules cite (3/3 for the bundled eu_bank spec).
2. STR acceptance is honest — str_acceptance_status is not_tracked, not a fabricated rate.
3. Determinism — re-run the command against the same run dir; the JSON is byte-identical (generated_at is the manifest as_of, not a clock read).

---

Common problems

Symptom	Cause	Fix
RTS coverage: 0/3 articles	No rule carries an AMLR Art. 28(1) / AMLR Art. 26 / AMLR Art. 20(1)(d) citation	Add the citation to the relevant rule's regulation_refs; re-validate
No manifest.json in <run-dir>	You pointed at a directory that isn't an engine run dir	Pass a real .artifacts/run-<TS> dir (or run aml run first)
precision is null for every rule	No labels supplied — precision needs ground-truth true/false-positive labels	The run doesn't carry labels; supply them from your QA system if you need precision (the funnel + coverage are label-free)
A citation won't resolve to a URL	The AMLA citation isn't in CITATION_URL_MAP	The three RTS citations ship resolvable; for a novel one add a url: to the regulation_refs entry

---

Next steps

• How to monitor model risk + per-rule drift — the sibling post-run report (aml model-inventory / model_risk_report.json); the AMLA pack reuses the same offline, deterministic, manifest-pinned shape.
• jurisdictions.md — the EU section, where the AMLA RTS effectiveness pack is the supervisory deliverable.