Flagship environment domain
Finance, fully instrumented.
Source documents, extracted facts, analyst corrections, formulas, workbooks, judgments and outcomes, connected into a single environment where models must complete real professional work.
Why this corpus is different
Most finance datasets stop at the filing or the final model.
Sofitra preserves the path between them: extraction, mapping, normalisation, formula construction, review, rejection, correction, approval and final decision. That path is where the trainable judgment lives.
Connected evidence
From public source to expert artifact, and every change in between.
The finance universe links raw evidence to structured facts, analyst work and deterministic checks.
Primary evidence
Filings, reports, notes, agreements, disclosures and supporting documents.
- Document versions and dates
- Structured facts and relationships
- Tables, footnotes and source locations
- Authority and provenance metadata
Expert construction
Models and analyses built from that evidence with the human action trail preserved.
- Mappings and normalisations
- Formulas and linked schedules
- Reviewer corrections and overrides
- Approved gold artifacts
Auditable result
Decisions, outputs and quality gates that define what successful work means.
- Accounting and formula checks
- Source-grounding checks
- Scenario and sensitivity integrity
- Professional decision rubrics
Task families
A curriculum across the finance workflow.
Tasks range from precise data operations to long-horizon analysis and judgment under ambiguity.
Statement extraction
Identify, map and source financial facts across filings, tables and notes.
Normalisation
Recast reported data, handle non-recurring items and preserve accounting integrity.
Three-statement models
Build and update integrated income statement, balance sheet and cash flow models.
DCF and trading comps
Construct assumptions, forecast cases, calculate value and explain sensitivities.
Underwriting and covenants
Model debt, liquidity, covenant headroom and downside risk using agreement terms.
Transactions and LBOs
Combine operating, financing and purchase-accounting logic into decision-ready analysis.
Model audit and repair
Detect broken formulas, circularities, inconsistent assumptions and source disconnects.
Investment analysis
Translate evidence into a view, identify drivers and support claims with traceable sources.
Judgment under ambiguity
Resolve incomplete evidence, materiality questions and competing professional constraints.
Environment sample
A model must leave both the workbook and the decision correct.
The verifier checks more than a written recommendation. It inspects formulas, state, sources, constraints and professional process.
Repair a transaction model before committee review.
The workbook produces an attractive return, but several linked assumptions and debt formulas are inconsistent. Find every material error, repair the model and update the committee summary.
Transaction modelOperating case, debt, returns and sensitivities
22 tabsLatest annual reportReported financials and segment disclosures
246 ppFinancing term sheetPricing, amortisation, cash sweep and covenants
18 ppCommittee memoDraft recommendation linked to model outputs
9 ppDiagnosis is part of the task.
The model must distinguish cosmetic issues from errors that change value, leverage, liquidity or decision quality.
Ran model integrity checksFound two broken links, one hard-code and a circularity hidden by iteration.
Reconciled historical EBITDARemoved duplicate add-back and linked the normalised base to the source note.
Rebuilt debt cash sweepApplied excess-cash definition and minimum-cash floor from the term sheet.
Updated returns and downsideIRR fell below threshold in the base case after repairs.
Revised committee memoChanged recommendation and cited each material correction.
Numerical integrity and decision integrity.
Passing the spreadsheet checks is necessary but not sufficient. The final recommendation must reflect the repaired economics.
✓Historical statements reconcile to sourcePASS
✓Debt schedule follows term-sheet mechanicsPASS
✓All seeded material defects are repairedPASS
✓Base and downside cases calculate without errorsPASS
✓Committee recommendation reflects corrected returnsPASS
Composite reward0.96
Finance-specific verification
Deterministic where finance is deterministic.
Balances, formulas, reconciliations and agreement terms can be checked directly. Judgment is scored separately with source-linked expert rubrics.
Statement integrity
Balance checks, cash-flow consistency, period alignment and accounting identities.
Workbook integrity
Formula lineage, hard-codes, broken links, circularities and scenario consistency.
Evidence grounding
Fact-to-source resolution, authoritative-document checks and citation coverage.
Professional judgment
Materiality, risk identification, recommendation logic and escalation quality.
Use cases
Train, evaluate or diagnose.
The same finance universe can support different model-development objectives without exposing the private source corpus.
Post-training curricula
Task families, escalating difficulty, trajectory data and component-level rewards.
Private capability benchmarks
Hidden task distributions and model-comparison reports grounded in real work.
Tool-use validation
Test whether agents can operate documents, spreadsheets and systems safely.
Failure and reward research
Replay model behaviour, inspect shortcuts and evaluate verifier robustness.
Finance access
Test the frontier on work that matters.
Request a sanitised environment sample or define a custom finance capability pilot.