A running list of data-quality issues across the atlases: datasets withheld, errors fixed, named limits, and pending fixes. Cross-referenced to the dataset and route each affects. All 25 published gold datasets have been run through a reconciliation audit (part-vs-whole, bounds, source-id, anomaly scans, plus targeted invariant checks); it found 2 silent corruptions (the Justice caseload and Care physician-payments entries in Fixed, below; the other Fixed entries are earlier corrections) — both now fixed and reconciliation-verified — and 0 outstanding reconciliation breaks.
Built but not published — the data is wrong.
Nothing is currently withheld. The federal civil-docket dataset that was previously withheld here (a misaligned case-category column) has been fixed and published — see Fixed, below.
Was wrong or overclaimed; corrected.
The federal civil-docket case-mix was corrupt: the parser assumed the AOUSC Table C-3 workbook had 27 columns, but it has 36 (merged headers, a spacer column, prisoner-petitions split into sub-columns), so a positional misread put the private all-civil total ($221,592) into the "intellectual property" row (real federal IP ≈ 12,500/yr). Fixed: decoded the real 36-column layout, verified by reconciliation (every district: party total = sum of its categories), summed the prisoner sub-columns, and added a reconciliation guard so the class can't recur silently. IP now reads 12,500. The page is published at /justice/jurisdictions and reconciles.
Physician-payment totals silently understated/mixed payment types: the headline total came from the CMS "by nature of payment" report (general payments only) while the company breakdown summed the "by company" report across all types — so research grants and ownership-stake values inflated companies above the total (a top company could exceed the doctor's "total"). Fixed: the three relationships are now kept distinct — general (the influence signal, with top companies, reconciling), research, and ownership/investment as separate labeled figures (a $91M physician-owned-distributorship stake no longer reads as a "payment"). 979,136/979,136 rows reconcile.
The structured data for state hazard pages advertised a "FIO/NAIC insurance non-renewal rate" in its own description and boundary — a figure that does not exist in our data (the source gold is empty for every state). We struck the reference rather than fabricate a number. A data source must never describe a figure it doesn't carry.
The cross-state storm-death comparison was a raw total, so large states ranked highest largely for being large (Texas read #1 partly because it's the second-biggest state). We added a population-normalized rate (deaths per million) and its own rank, and labeled the raw count as such. Texas is #1 raw but #5 per-capita; the per-capita number is the honest "is that a lot where I live" answer.
Shipped, with the limit named on the page — read these before citing the figure.
An initial 50,000-filing sample of the ~800K-filing DOL Form 5500 series — labeled as a sample wherever it appears and not treated as a census. An employer's absence means "not in the sample," not "no filing." The full ingest is queued.
Premiums are sticker prices before income-based tax credits. Most marketplace enrollees pay far less; a credit tied to income can bring a plan near $0. We show how plans compare and the unsubsidized cost — not what you'll pay — and do not compute net cost (the 2026 enhanced-subsidy rules are unsettled). Scope is the 30 federally-facilitated-marketplace states; state-based-exchange states (CA, NY, …) are labeled as such, never shown as $0.
The recorded storm history is a single year (2025), not a multi-year trend — one severe event can swing a state's rank. Per-capita ranks are unnormalized single-year totals; read them as direction, not a precise standing.
The full GLEIF register (~3.35M legal entities), one row per LEI. It carries registration status and structure only (e.g. ISSUED / LAPSED, jurisdiction, name) — a point-in-time public record, never a verdict on a counterparty's conduct or solvency. An entity's absence means no LEI is on file, not that the entity doesn't exist.
BLS OEWS is an establishment survey that lags ~12 months and suppresses estimates for small-population occupations and areas — a missing wage means suppressed, not zero. The Anthropic Economic Index used alongside it is non-government and labeled where used; the federal BLS/O*NET figures are the authoritative spine.
The cross-registry corporation match is a name heuristic (labeled DERIVED), superseded by the hard-code (FDIC/GLEIF) joins; treat a cross-registry match as a lead, not proof.
Named gaps with a known fix, not yet shipped.
Pages for closed/merged FDIC institutions don't yet carry the closure date or successor institution (the gold lacks the FDIC CHANGEC/successor fields). The page says so and routes you to FDIC BankFind for it; a successor re-ingest is the queued fix.
Premiums are summarized across a state's rating areas; per-rating-area / per-county localization (needs the CMS rating-area→county file) and the subsidy-adjusted net cost (SLCSP, gated on 2026 subsidy law) are both queued.
For agents. This ledger is machine-readable at
/corrections.json — each issue carries its atlas, affected dataset/route, status (withheld / fixed /
known-limit / pending), and detail. Check it before treating any Atlas Commons figure as authoritative.