Methodology
Short, honest notes on provenance, validation, and the limits of the current extract.
What this is
SerffSurf is a shareable snapshot of a workers' compensation rate-filing intelligence database built from public SERFF filings in 11 non-NCCI states (2020–2026).
It is a progress surface for peer review — not a product website, not a filing repository, and not actuarial advice.
How numbers get here
Spine parse (deterministic): the main SERFF “PDF Pipeline” document for each filing is parsed into Filing-at-a-Glance fields, company tables, section text, and correspondence letters.
Entity resolution: company names map to NAIC cocodes and groups where possible; rate rows join through a resolved company mart.
LLM extraction (second pass): LCM and deviation facts are extracted only from supplied filing context. Every accepted fact must include a verbatim evidence quote that appears (alphanumeric-normalized) in that context.
Hard validity gates: LCM values outside 0.4–6.0 and deviation percentages outside −90…+300 are marked invalid. Invalid facts are retained with error metadata; this site only surfaces valid facts.
What to trust — and verify
Evidence quotes are the credibility layer. Prefer them over bare numbers when a figure will drive a decision.
Company-table rate rows are generally more reliable than filing-overall figures repeated across affiliates.
Table-jumble exhibits can yield verbatim-true but semantically mislabeled facts. A hand-labeled gold sample is the planned precision check before treating extracts as benchmark-grade.
Always verify material figures against the underlying SERFF filing.
Architecture of this site
Fully static. Pre-built JSON is exported from DuckDB on the owner’s machine and committed into public/data/. There is no live database, no API keys, and no server-side write path.
Refresh: re-run the export, copy the three JSON files, push — Vercel redeploys.