A collection of self-contained document environments spanning finance, housing and census data, baseball analytics, and historical statistical atlases. Each environment pairs heterogeneous source files with expert-authored, multi-step questions, canonical answers, evidence locations, reasoning paths, and records of prior model attempts.
Available Now · Sample Available
Document Environments
Traceable document-reasoning tasks grounded in self-contained professional corpora.
Volume
0
Volume
Scalable Volume
Coverage
Legal, Finance, Enterprise
Supplied Trajectory Reward
0.0
%
Opus 4.6
Targets
MMLongBench-Doc, DocVQA, LegalBench-RAG
Modality & Delivery
DocumentImageText→Text
Formats
PDFJSONImages
What You Receive
Coverage
Questions cover document retrieval, PDF and table parsing, cross-document synthesis, chart and map interpretation, numerical reasoning, statistical modeling, document version selection, and long-context grounding. Every reference answer is linked to its supporting documents, pages or regions, and a canonical reasoning chain.
Package Structure
- sample01.json
- sample02.json
- Environment01/
- EnvFiles01/
- File01
- File02
- EnvFiles02
- EnvFiles03
- EnvFiles01/
Differentiators
- Self-contained professional corpora with traceable ground-truth answers.
- Exact document and page-level evidence annotations.
- Canonical multi-step reasoning chains and answer tolerances.
- Historical model attempts with expert diagnoses of failure modes.
- Mixed-format environments containing PDFs, spreadsheets, CSVs, JSON, HTML, images, and specialized data files.
Record Schema
Results
| Subject | Type | Metric | Result |
|---|---|---|---|
| Opus 4.6 | Trajectory reward | Pass rate | |
| Opus 4.7 | Result | — |
Methodology & Quality
Methodology
SME involvement
Domain-expert curated with up to L2 review
Stressors
Large and deeply nested document collections.
Dense PDFs with varying extractability and OCR quality.
Mixed PDF, XLS/XLSX, CSV, JSON, HTML, image, and specialized data formats.
Cross-file reconciliation and multi-document synthesis.
Chart, map, and table interpretation.
Long-context grounding and document-version selection.
Multi-step numerical and statistical reasoning.
Strict answer formats, units, rounding rules, and tolerances.
Design criteria
Build assumptions
Source-document paths are relative to each corpus root and use forward slashes. PDF page references use one-indexed PDF positions rather than printed page labels. Solvers are expected to work primarily from the supplied corpus, except where a question explicitly carries the external_knowledge skill tag. Numerical answers are graded using the stated units and tolerances; text and categorical answers use exact canonical values. The reasoning chain documents the canonical solution but is not the only acceptable path.
Quality approach
Known limitations
Customization & Scalability
Custom sets available. Additional volume in progress