ForensicBench Ledgers

Unlabelled synthetic general ledgers for ForensicBench, a benchmark that evaluates whether agentic LLMs can discover and type coordinated journal-entry fraud schemes over a live ledger (EMNLP 2026 Industry Track).

Five sector ledgers (Energy, Healthcare, Luxury Goods, Manufacturing, Transport), about 300K journal entries each (1.51M in total) over three years, in EUR under the French Plan Comptable General (PCG). Five multi-entry fraud scheme types are injected into each ledger. No fraud label is distributed here: labels live in a private store and are used only by the automatic scorer.

Content

data/<sector>/<table>.parquet, with the tables below (one split per sector in the dataset viewer):

Table	Description
je_header	one row per journal entry (document id, dates, type, business process, creator)
je_line	entry lines (GL account, debit, credit, auxiliary account, lettrage)
chart_of_accounts	PCG account tree
vendors, customers, employees	master data

The HR payroll table used in the paper (hr_employees) is documented in the schema of the reference harness but is access-controlled there; it is not distributed.

Fraud catalogue and agent prompts

The reference harness gives the agent a fraud catalogue describing the five scheme types. It is conceptual: it describes the business process, what normal operations look like and the kinds of breakdown that can indicate manipulation. It does not name expected GL accounts, posting templates or injection parameters, and it contains no labels.

File	Content
catalogue/fraud_catalogue.md	the catalogue exactly as injected into every run
catalogue/fraud_catalogue.json	the same content, one record per scheme type (scheme_type, title, normal_process, warning_signs)
prompts/	the prompt modules of the reference harness: system prompt, orientation, planning, hypothesis worker, schema description, investigation discipline

Use them as they are to reproduce the baseline, or as a starting point for your own harness. Any harness may use a different catalogue and different prompts, as long as it states them in its submission.

Use with PostgreSQL

The reference agent queries a read-only PostgreSQL database. To rebuild one sector:

pip install pandas pyarrow psycopg2-binary
python tools/load_postgres.py --sector energy --dsn "dbname=forensicbench_energy user=postgres"

Task and submission

An agent investigates each ledger and flags fraudulent journal entries with a scheme type from: fictitious_ap_disbursements, revenue_manipulation, vendor_collusion, shadow_payroll, inventory_manipulation.

Each model is run on the 5 sectors with 5 replicates (25 runs). A submission is a CSV with one row per flag (replicate is only needed for the multi run format):

Column	Meaning
sector	one of the five sectors
replicate	replicate index of the run
document_id	je_header.document_id of the flagged entry
scheme_type	one of the five scheme types, or unknown (counts for Entry-F1 only)
rationale	optional free text

Check a file before submitting: python tools/validate_submission.py my_flags.csv.

Metrics

• Entry-F1 (primary): F1 over flagged journal entries.
• Type-F1: an entry counts only if its scheme type is also correct (at most Entry-F1).
• Coverage: share of each fraud family recovered, averaged over families.
• Consistency: stability of Entry-F1 across the 5 replicates (0 to 100).

Entries are ranked by Entry-F1, then Type-F1, Recall, Precision, Coverage and Consistency (no composite score). Per-sector scores are averaged over replicates, then macro-averaged over sectors.

Labels

The ground-truth labels are held out so that the leaderboard stays meaningful. To obtain them for research evaluation (local re-scoring, reproduction), write to guywaffo@gmail.com.

Leaderboard and submissions

Leaderboard, submission form and scoring: https://waguy02.github.io/PHD-Research-Website/forensicbench/leaderboard/

Two formats are accepted: a single run (one run per sector, no standard deviation) or a multi run (same replicates in every sector, reported as mean, standard deviation and Consistency). Each submission must come with the code of its harness (repository URL and commit hash, or an archive); entries are unverified until a maintainer has reviewed it.

Citation

@inproceedings{waffo2026forensicbench,
  title={ForensicBench: Evaluating Agentic LLMs on Journal-Entry Fraud Detection},
  author={Waffo Dzuyo, Guy Stephane and Guibon, Ga{\"e}l and Cerisara, Christophe and Belmar-Letelier, Luis},
  booktitle={EMNLP 2026, Industry Track},
  year={2026}
}