A pipeline you can read is a dataset you can trust.

Twenty-four collector segments overlap on purpose: several crawlers were pointed at overlapping universes of payers, so a document missed by one pass surfaces in another rather than disappearing silently. The price of overlap is duplication, so merging dedupes in two passes — first on numeric identity (same source URL, same approved and denied counts: two collectors reading the same row produce the same numbers), then on name identity (same URL, contract ID and normalized plan name), which catches filings that published no counts at all. When two independent reads of the same document disagree, the disagreement is not resolved by keeping the prettier row — it is written to merge_conflicts.json and resolved against the source document, because a disagreement between readers is itself evidence about how hard the document is to read.

Every record lands in one flat schema, and this is where the project's central rule is enforced: rates are computed from counts, never transcribed. If a payer published 1,000 standard requests and 214 denials, its denial rate is 21.4% by division, whatever its PDF says. The percentage the payer printed is kept in its own column — std_denial_rate_printed — precisely so the two can disagree in public. A filing that published only percentages keeps them, flagged as unverifiable; its count columns stay null, because "the payer did not say" is a fact, and zero would be a lie.

The part that exists nowhere else. CMS mandates publication but checks no arithmetic, so the 10 consistency rules do what an auditor would: approved plus denied must account for the stated total, the printed rate must match the computed one, overturned appeals cannot exceed appeals filed, expedited decisions should not take longer than standard ones, and rates at the extremes get flagged before anyone quotes them. A rule that fires produces a finding — the rule name, a severity, and the exact numbers that tripped it — never a correction. The pipeline does not fix payers' arithmetic; it reports it. The documentation states each rule as an equation with its thresholds.

Build stamps every filing with its quality flags and writes the two principal artifacts: the metrics CSV and the findings CSV. Gaps does the same honest bookkeeping for the misses: every document the collectors could not read is classified — in code, reproducibly — into blocked (the document exists and resisted machines: robots.txt, HTTP 403, bot walls, scans), missing (a search that came up empty, a finding about the crawl rather than the payer), or not a gap (duplicates, plans with no CY2025 filing to find). The distinction matters because only the first group supports the claim that a filing is being withheld — and conflating the three would have overstated that claim by a factor of five.

The same rows go out in every form someone might actually use: a standalone HTML report that opens without a server, and a formatted two-sheet workbook — Metrics and Validation — for the spreadsheet-first reader. Then tools/ in the web repository turns the CSVs into the versioned SQL dumps this site imports. Nothing is retyped anywhere in that chain; the number in a payer's PDF and the number in this page's HTML are connected by code alone.

Every source the analysis reads is declared in one file — the pinned URL for each Medicare file, the CHR and Urban Institute releases, the Census relationship file, and the discovery rule for hospital price files. Fetching writes a manifest entry per download: the exact URL, the fetch time, the byte count and the SHA-256. That last field is the quiet backbone of the whole analysis — anyone can download the same file tomorrow, hash it, and know with certainty whether they are looking at what the pipeline read.

Three parser families, one per source shape. The Medicare files are large, clean CSVs — every hospital by every MS-DRG and APC, every state by every HCPCS code — where the only judgment is remembering that CMS suppresses rows under 11 patients, so a missing row means "not published", never "none performed". The hospital price files are the opposite: each is discovered through the cms-hpt.txt pointer the rule requires at the hospital's web root, streamed rather than loaded, and read for every code-like string it carries. Files that cannot be found or parsed are recorded with a reason and classified exactly as the prior authorization gaps are. The county parsers join health outcomes, medical debt and demographics onto one county spine via the Census ZCTA-to-county file.

Most strings a hospital's price file calls a code are not codes — chargemaster numbers and fragments that happen to pass a regex. The dictionary gives every string a status by set membership, never by judgment: official if it appears in a CMS list, hospitals only if well-formed and independently listed by two or more sampled hospitals, unverified otherwise — and unverified strings are priced nowhere on this site, because a price for a non-code is not a price. Descriptions are copied from a named source or left blank; the pipeline never writes one, and CPT text outside the Medicare file is AMA copyright and stays blank rather than guessed.

Enrichment computes what the sources imply but do not say — charge-to-payment ratios per row, state medians for the comparison basket, system-level sums weighted by discharges — always from figures in the same row or the same declared join, never across an undeclared boundary. Validation then runs the price rules: a cash price above list, a negotiated rate outside the file's own stated bounds, a price file more than a year old, and a hospital whose published list price sits far from the average charge it reported to Medicare for the same admission. As in analysis 01, every miss is a finding with the numbers attached.

The one stage that is not standard library (numpy, pandas, scikit-learn) fits gradient-boosted trees and a ridge regression for three measured county outcomes — medical debt in collections, premature mortality, years of life lost — from prices, insurance, income, age, race, rurality, access and behaviour. Cross-validated, fixed seed, permutation importance, partial dependence. Nothing in it is causal and every figure it produces says so; the Analysis page renders the module's output and nothing else.