Section 3
Corpus
Public documents, chosen to be long, messy, cross-referencing, and naturally partitioned into tenants.
The manifest is what makes the corpus reproducible without redistributing gigabytes: replay the URL, verify the checksum. Raw documents are cached locally and not committed.
3.1 Sources
SEC EDGAR 10-K filings
Long, structurally messy, heavy with cross-references and tables. They partition naturally by filing company, which is what makes tenant isolation meaningful to test rather than theatrical. Only the primary document of each filing is fetched.
CUAD — Contract Understanding Atticus Dataset
Commercial contracts annotated with clause spans, CC BY 4.0, which supplies ground truth for a subset of queries. These contracts are themselves drawn from EDGAR, so they are a different document shape from the same publisher rather than an independent source, and they are deduplicated against the filing set by containment.
RFCs
Documents from the HTTP and families. Freely redistributable, densely cross-referencing, and a clean hard-mode subset — the only genuinely independent third source.
3.2 Attribution
: Hendrycks et al., NeurIPS 2021; The Atticus Project. Licensed 4.0. Chunking and re-indexing constitute modification and are indicated as such. filings are published by the US Securities and Exchange Commission. s are published by the and RFC Editor under 78.
Corpus edgar-cuad-rfc-v1