VynFi/perfect-audit-crime-data

Perfect Audit Crime Challenge — ledgers

The dataset behind the Perfect Audit Crime Challenge: synthetic general ledgers with unlabeled planted fraud. Flag the fraudulent journal entries and the Space scores you against the held-out ground truth. The challenge is not about crowning a fraud detector — it maps the detectability frontier: where ledger analytics ends and external evidence must begin.

Headline

field	value
Units (multi-entity GLs)	3 — manufacturing, retail, financial-services (small)
Journal entries	25 422
Fraud entries	1 087 (4.3 %) across graded families
Mimetic perfect-crime entries	180 (the joint-invisible family)
Tracks	A (ledger only) · B (ledger + evidence)
Labels	held out — scoring is via the Space
Generation engine	mivertowski/SyntheticData @ v5.34
Determinism	reproducible from the per-unit seeds

The two tracks

Both tracks are the same ledgers; they differ only in what you are given.

• Track A — ledger only. Structured, relational and cross-period memory frauds are catchable. The mimetic perfect crime — fraud fabricated from the ledger's own normal distribution — provably is not (it preserves the joint distribution of every observable, so no ledger-only analytic beats chance on it).
• Track B — ledger + evidence. You additionally receive the ISA-520 expectations (per-account aggregate vs an exogenous expectation) and ISA-505 evidence anchors (external corroboration / dangling accounts). The aggregate-inflation and fabricated-counterparty crimes now become catchable.

The gap between the two tracks on the mimetic family is the result — it shows where ledger analytics ends and external evidence must begin (the companion paper's prop:counter).

Structure

<unit>/                              # manufacturing_small / retail_small / financial_services_small
  track_a/journal_entries.json       # ledger only (engine JEs; fraud columns stripped)
  track_b/journal_entries.json       # same ledger ...
  track_b/external_expectations.json # ... + ISA-520 per-account expectations (deviation, exceeds_band)
  track_b/evidence_anchors.json      # ... + ISA-505 corroboration (corroborated, is_dangling)

• journal_entries.json — a list of { "header": {...}, "lines": [...] }. Each line carries gl_account, debit_amount, credit_amount. The submission key is header.document_id, unit- prefixed (e.g. retail_small:<uuid>) so a combined submission spans all units. The fraud / fraud-type columns are removed — that is the held-out answer.
• external_expectations.json (Track B) — per material account: expected_value, tolerance_pct, lower_bound / upper_bound, actual_value, deviation, deviation_ratio, exceeds_band.
• evidence_anchors.json (Track B) — per material account: total_activity, corroborated, corroboration_method, is_dangling.

Submitting

1. Build a submission — a CSV/JSON listing the document_ids you flag as fraudulent (optional score column for ranking).
2. Upload it at the Space, pick your track, and you are scored + added to the leaderboard.

Metric: PR-AUC on the held-out per-JE labels (plus precision / recall / F1). The leaderboard also reports recall by observability class, so the mimetic family is visibly (near-)unbeatable in Track A and recoverable only with evidence in Track B.

Loading

import json
jes = json.load(open("manufacturing_small/track_a/journal_entries.json"))
for je in jes:
    doc = je["header"]["document_id"]          # unit-prefixed submission key
    total = sum(abs(float(l.get("debit_amount") or 0)) +
                abs(float(l.get("credit_amount") or 0)) for l in je["lines"])
# Track B evidence:
exp = json.load(open("manufacturing_small/track_b/external_expectations.json"))
anc = json.load(open("manufacturing_small/track_b/evidence_anchors.json"))

How it was produced

Each unit is a v5.34 DataSynth general ledger (mixed engine fraud families + relocating campaigns) with a post-injected mimetic perfect-crime family: balanced two-line journal entries whose amounts, contras and (weekday) dates are drawn from the target account's own normal distribution, inflating a credit- dominated aggregate just past the materiality band while leaving every per-entry pattern normal. The Track-B evidence layer (ISA-520 expectations + ISA-505 corroboration) is computed on the full labelled ledger with the ground-truth columns stripped. Reference baselines (a per-account-conditional ledger detector and an evidence detector) seed the leaderboard; on these units the mimetic family scores recall ≈ 0.25 in Track A vs ≈ 0.94 in Track B.

Framing

A research benchmark on purely synthetic data. A high score is not a deployable fraud detector, and the entries no one can catch from the ledger are precisely why audit must combine analytics with external evidence (confirmation, reconciliation, substantive analytics). It is not a recipe for committing fraud — it is a map of detectability and its limits. No real or client data is used anywhere.

License

Apache-2.0 — same as the engine.

Related artefacts

• Challenge Space: VynFi/perfect-audit-crime-challenge
• Engine: mivertowski/SyntheticData
• Companion paper: The Detectability Frontier (in press)