LedgerBench-100

LedgerBench-100 is a deterministic corporate-finance agent benchmark: 100 distinct employee decisions over isolated simulated finance worlds. Each world combines a Dynamics 365-shaped ERP, an Odoo procure-to-pay and manufacturing surface, a QBO-style subsidiary ledger, Microsoft Graph workbooks, Gmail, governed documents, and frozen SEC XBRL filings through 8 MCP servers exposing 66 provider-shaped tools. The employee asks for an outcome in ordinary language; the agent must discover the reporting schema and the task-specific investigation.

Grading is fully deterministic. LedgerScore groups the executable verifier contract into 16 task-specific employee outcomes worth 100 points: scope, authority, current state, task-native investigation, task-native causal reasoning, operating-plan controls, supported decision, costed options, operational and case state, collaboration, outcome verification, provider-native readback, containment, exact insights, and execution sequence. Strict pass still requires all 100 points. No LLM judge, network, clock, or randomness appears in the reward path.

The source job is the scoring core, not a decorative wrapper: at least 58 of 100 points in every task are assigned to task-native source investigation, reasoning, decision, persisted operational state, and employee-facing insights. Every ERP planning task independently grades its Odoo state transition in addition to the cross-system decision record.

Every case also carries a graded operating-control model: the control requirement is derived from the ERP's in-scope DecisionScopeLines, usable support is reconciled from the current evidence register net of counterparty-corroborated exclusions, the exception is tested against the policy tolerance, the counterparty's committed date and the close or production-planning window bounds three costed timing alternatives (one within authority, one held for the counterparty, one requiring a higher-authority exception), and the selected outcome is compared with the requester's documented need-by date into a signed variance and an honest timing status. All 22 intermediate and final values are graded as their own answer fields, the selected option is persisted on the case, and the completion handoff is graded on its content.

Measured contents

• Tasks: 100 across 23 families: anomaly_triage (1), bank_rec (2), business_brief (1), cash_app (2), cash_forecast (1), close_mgmt (3), collections_ops (1), cross_system (4), erp_qa (6), erp_qa_fb (23), erpbench (27), expense_audit (3), finance_qa (7), fixed_assets (1), fpna (3), journal_entry (1), payment_proposal (1), payment_run (2), pbc (2), revenue_accounting (1), threeway_match (3), treasury_fx (1), vendor_master (4)
• Oracle walk length: min 38 / median 44 / max 144 MCP calls (5722 total); required distributed evidence reads are 28-42 per task
• Public rubric: exactly 16 causal employee-outcome milestones / 100 LedgerScore points per task. The execution layer separately contains 2855 answer + 2707 trace + 1139 state checks, with 81-95 atomic verifier assertions per task assigned exactly once beneath those milestones
• Graded reasoning chain: 23-36 graded answer fields per task (9-16 numeric derivations), a graded decision model in 100/100 tasks and three costed alternatives with authority status in 100/100; task-native work carries 58-58 points per task
• Inspectable assets: 58-72 agent-visible native files per task, including 28-42 exact executed MCP responses bound one-for-one to the verifier's material reads; every material export carries request scope, bytes, and SHA-256, while gold and oracle recipes are excluded
• Readback depth: 134 task-native provider readbacks across the release, plus the governed case and completion-thread readbacks in every task
• Reference diversity: 100/100 distinct raw server/tool sequences (maximum sequence match 0.945274) and 100/100 distinct semantic action graphs
• Human-job curation: 0 prompt variants and zero lineage-marked derived sources; the release uses one instance of each of 27 ERP planning archetypes and 23 distinct ERP control questions, plus separately authored fixed-asset capitalization, bundled-contract revenue allocation, and FX remeasurement workflows. Ticker or company-name swaps do not count as new jobs
• Prompt uniqueness: 100 distinct employee requests; maximum pairwise 5-shingle Jaccard 0.663462

What is included

• data/tasks.jsonl: apex-accounting-compatible records (task_id, task_name, world_id, prompt, context_files, rubric, gold_output, metadata).
• tasks/: one readable JSON record per task.
• task_files/: 58-72 task-scoped agent-visible evidence files plus any native seeded documents and inputs.
• world/: the world source — MCP framework, the eight servers, the deterministic verifier engine, the Streamable HTTP bridge, and the full SQL schema.
• trajectories/: one normalized oracle MCP trajectory per task.
• reports/: measured build and qualification evidence.

The runnable form is the Harbor dataset blobfishai/ledgerbench-100: self-contained task packs (prepared SQLite world + runtime on a digest-pinned python:3.12-slim) whose tests/test.sh calls a token-gated /verify endpoint; the agent container never sees the verification token.

Measured qualification (1600 executions)

Gate	Result
Oracle replays	100/100 reward 1.0
Deterministic replays (byte-identical reports)	100/100

Negative control	Executions	False accepts
incomplete_read	100	0
keyword_stuffing	100	0
missing_readback	100	0
noop	100	0
rejected_mutation	100	0
shortcut	100	0
state_only	100	0
unauthorized_write	100	0
write_before_read	100	0
wrong_decision	100	0
wrong_evidence	100	0
wrong_option	100	0
wrong_target	100	0
wrong_value	100	0

Full per-task evidence is in reports/qualification.json; do not infer a model score from the oracle trajectories.

Data provenance and contamination

The company, its customers, vendors, employees, balances, and documents are synthetic (FinanceBenchmark-derived journal shapes with synthetic entities). The filings surface serves frozen real SEC XBRL facts (38 registrants, snapshot-pinned) — real public data, included under its own public-domain terms. Task text and gold answers are original to this release's source repository. Gold outputs are public, so this release suits transparent evaluation and RL experiments rather than secret-test claims.

Licenses

Task data and documents are CC-BY-4.0. Benchmark code and harnesses are Apache-2.0. SEC XBRL facts are US-government public-domain data.