Provenance log
Every figure on limn is badged for how real it is - ● full, ◐ curated, ○ illustrative. This is that honesty told forward: each time a dashboard climbs the ladder, it lands here.
A new Labor & Workplace dashboard on the BLS Census of Fatal Occupational Injuries - a census, not a survey, so it is a complete count of workers killed on the job. It ships full: every figure is aggregated from the real CFOI time-series flat files (download.bls.gov fa/ and fw/), joined to BLS OEWS May 2024 median wages on SOC occupation code, nothing curated. 5,070 workers were killed in 2024, a rate of 3.3 per 100,000 full-time-equivalent workers - one death every 104 minutes. The finding the dashboard is built around: the ten deadliest occupations pay a median of $49,550, and the median across every occupation in the country is $49,500. A $50 difference for six to thirty-three times the risk of being killed. Logging runs 110.4 deaths per 100,000 (33.5x the national rate) and pays $49,540; roofers 48.7 and $50,970; refuse collectors 37.4 and $48,350. The one deadly job that actually pays is airline pilot (36.7, $198,100), annotated as the exception it is. A risk-versus-pay scatter, a ranked-by-rate chart carrying published margins of error, an event-and-exposure breakdown, and a state choropleth carry it, with a per-occupation compare route. Honesty notes carried on the page: the published rate is BLS's own and is never derived from the count (the naive count-over-hours arithmetic gives 3.41, not 3.3, because BLS excludes volunteers, active-duty military and workers under 16 from the numerator); the wage join crosses two universes, because CFOI rates include the self-employed (who die at 10.8 per 100,000 against 2.9 for wage-and-salary workers) while OEWS covers employees only, and that caveat sits on the scatter itself rather than in a footnote; the margins of error are published and large (logging is 110.4 +/- 34.9) so the bars carry them and the ranking is not presented as precise; CFOI counts 263 suicides and 410 fatal overdoses as workplace deaths, which the page states plainly; and CFOI is split across separate surveys (FI/FW/FA) precisely because SOC, NAICS and OIICS revisions broke comparability, so national totals span eras but detailed categories are not naively concatenated.
A new Housing & Weather dashboard on the HUD Picture of Subsidized Households - public housing, Housing Choice Vouchers, project-based Section 8 and the rest of the federal housing safety net, 2009-2025. It ships full: every figure is aggregated from the real HUD Picture files joined to HUD's own CHAS 2017-2021 data for the eligible-household denominator, nothing curated. The thesis is that housing assistance is not an entitlement. 5,132,918 units are under contract and 4,471,285 households are housed, but 10,620,313 extremely-low-income renter households exist, so only 33.3% of the households in deepest need get anything at all and roughly 7.09 million get nothing. Those who do get in waited a median of 25 months and then stay 123 months, because there is nowhere else to go; only 9% moved in during the past year. The wait ranges 6x by state, from Maryland at 57 months to West Virginia at 10. Tenants pay 28.5% of the cost ($452 a month against HUD's $1,135), an implied federal outlay near $61B a year. 43% of assisted households are elderly, 24% include someone disabled, 74% are female-headed, and average household income is $18,483 (21% of local AMI). Honesty notes carried on the page: HUD's suppression sentinels (-1 missing, -4 suppressed under 11 families, -5 non-reporting) are pervasive rather than rare, invalidating the wait field in 41.1% of county rows, so they are filtered before every mean and the surviving coverage share is printed on every wait chart; and critically, months_waiting measures only households that were ADMITTED, making it survivor-biased and almost certainly an understatement, so the site says 'the wait experienced by those admitted' and never 'the waiting list' - the data cannot support that second sentence. The need denominator comes from a different vintage (CHAS 2017-2021 ACS against a 31DEC2025 Picture snapshot) and both are stated on the chart.
A new Transportation & Safety dashboard on the FTA National Transit Database major safety and security events - 113,509 real event rows, 2014-2026, across 530 transit agencies. It ships full: every figure is aggregated from the real NTD event file, joined to NTD service metrics so that safety can be expressed as a RATE rather than a raw count, nothing curated. Two findings drive it. First, the dead are not who the public pictures: of 3,569 people killed, only 277 were riding the vehicle. The rest were people on the road (905), suicides (849), people on the right-of-way (422), and people waiting or boarding (480). Riding transit is not what kills people; being near it is. Second, dividing by vehicle revenue miles completely reorders the accountability table - MTA New York dominates every raw count, but per million miles of service actually delivered the worst large agencies are elsewhere entirely, and that reordering is the signature chart rather than a footnote. Nationally the rate is worsening even after adjusting for service (2.97 per million VRM in 2022, 3.10 in 2023, 3.21 in 2024). Transit worker assault events rose from 203 in 2014 to 687 in 2024. Honesty notes carried on the page: the FTA only created the worker-assault flag in 2023 and backfilled earlier years, so better capture inflates that rise - the direction is real but the page explicitly refuses to publish a clean 'it tripled' multiple, and the 2023 boundary is drawn directly on the chart axis and named in its aria-label; this is the MAJOR-event file, gated on reporting thresholds, so 'events' never means 'all incidents'; 2026 is partial on the 90-day reporting lag and is excluded from every trend; the service denominator exists only for 2022-2024, so rate charts are scoped to those years while the long series carries raw counts only; and rate rankings apply a 15M vehicle-revenue-mile service floor, because otherwise a tiny agency with a single event tops the chart.
A new Health & Medicine dashboard on the CDC National Notifiable Diseases Surveillance System (NNDSS) - the case counts of reportable infectious diseases that states report each week. It ships full: every figure is a straight aggregation of the real NNDSS weekly tables (data.cdc.gov resource x9gk-5huc), nothing curated. The through-line is the resurgence of vaccine-preventable diseases. Measles ran 122 cases in 2022 and 2,026 in 2025 - a 16.6x climb and the most in any year since 1992 - with Texas (803), Arizona (206) and Utah (178) leading, and 40 of 51 states reporting at least one case; whooping cough (pertussis) rose 12x from 2,388 to a peak of 35,493; mumps stayed flat near 340. A per-state choropleth, a year-over-year resurgence trend, and a per-vaccine-preventable-disease table carry the story, with a per-state compare route for all 51. Honesty notes carried on the page: these are PROVISIONAL year-end counts (final-week cumulative m3), a floor that finalized totals exceed (pertussis 2024 was 35,493 provisional, 43,402 finalized); they are counts, not per-capita rates, so large and small states are not directly comparable; a few conditions were not nationally notifiable in early years and are shown as n/a, never 0; and the 1992 comparison is external CDC editorial context, while the counts themselves are the NNDSS numbers.
A new Money & Accountability dashboard on the IRS SOI county-to-county migration data - where filer households and their aggregate income move between the states, read off the address changes on filed returns. It ships full: every figure is a straight aggregation of the real IRS state and county inflow/outflow files for the 2021-2022 pair (the interstate totals reproduce the IRS aggregate rows exactly), nothing curated. In one year 4.05M filer households changed states and carried $403B of adjusted gross income across a state line. Florida gained the most (+$36.0B AGI, +125,551 filers) and California lost the most (-$23.8B, -144,203), with New York down $14.2B and Texas up $10.1B; the biggest single corridor was New York to Florida (51,967 households, $9.5B); at the county level Palm Beach FL gained $4.95B while Cook County (Chicago) lost $7.16B. A diverging net-AGI choropleth, money-vs-people mover bars, the biggest corridors, and top county gainers/losers carry it, with a per-state compare route for all 51. Honesty notes carried on the page: AGI is reported in thousands in the source; net is interstate inflow minus outflow, with per-state aggregate rows and non-migrants excluded so movers are never double-counted; county gainers/losers use each county's all-domestic-moves total (a deliberately different measure than interstate state net, which surfaces metro-to-suburb flight); suppressed small flows are treated as absent; and this counts federal filers and reported income, under-counting the lowest incomes and measuring income, not wealth.
A new Health & Medicine dashboard on the CMS Medicare Physician & Other Practitioners public use files - what U.S. providers bill Medicare versus what Medicare allows. It ships full: every figure is a straight aggregation of the real 2024 by-Provider file (1.30M providers, streamed) and twelve years of the by-Geography file, nothing curated. The through-line is the markup - submitted charges over the Medicare-allowed amount - which runs 3.25x nationally ($494B billed against $152B allowed) and widens sharply by cut: emergency medicine bills 6.2x and the anesthesia trades 9-11x while primary care sits near 2x; a hospital/facility service marks up 4.8x against 2.6x in an office; the ratio climbs from rural (2.5x) to metro (3.3x); and it varies 2x across states, from Idaho to Alaska, on a per-state choropleth. Honesty notes carried on the page: these are submitted charges against the allowed amount, not what Medicare paid ($120B), and a high markup is the distance between a list price and a regulated rate, not fraud - though the uninsured can face the list price. Drug/biological codes carry inflated per-unit markups and are flagged; services are ranked by allowed dollars, not markup. The 3.25 GB by-Provider-and-Service file was deliberately not ingested - the two files here carry every figure shown.
A new Government & Education dashboard on the NCES Common Core of Data - what U.S. public schools spend per student, where the money comes from, and how crowded the classroom. It ships full: every figure is a straight aggregation of the real CCD flat files, nothing curated. Money comes from the F-33 district finance survey (2022-23): $796B of current spending across 49.3M students, about $16,164 each, but with a 2-to-1 gap between states - the Northeast near $30,000 (DC, New York, New Jersey) and the Mountain West and Sun Belt near $10,000 (Idaho, Utah, Oklahoma), on a per-pupil choropleth. Funding splits three ways - nationally 12% federal, 44% state, 44% local - but states lean on those pockets in wildly different proportions, shown in a federal-blue / state-amber / local-purple triad. Staffing comes from the nonfiscal 2024-25 files: a 15.2:1 national student-teacher ratio joined on district ID, ranging from ~22:1 in California and Utah to ~11:1 in Maine and DC. Honesty notes carried on the page: per-pupil is current spending (not capital or debt); the finance and enrollment strands are dated separately (2022-23 vs 2024-25); F-33 negative sentinel codes are treated as absent, never summed; and the district leaderboard is ranked by enrollment, not per-pupil, because a per-pupil ranking is dominated by county offices of education serving tiny high-need populations.
A new Transportation & Safety dashboard on the FAA UAS Sightings Reports - the agency's quarterly logs of pilots and air-traffic controllers reporting drones near aircraft. It ships full: every published quarter from October 2019 through June 2026 was ingested straight from the FAA's Excel files, 11,379 reports across the 50 states and DC, with nothing curated. The story is where (a per-capita state choropleth that redraws the raw-count map around DC, Nevada and Florida), when (a clean daylight curve - 71% of timed reports land between 10a and 6p), how deep in the airspace (tower vs approach vs enroute center, parsed from the report narrative), and whether it is rising (a 2021 surge, a dip, and a climb back). Honesty notes carried on the page: these are reports, not confirmed drone flights or verified violations; and the FAA created pages for FY2023 Q1-Q3 but never posted the files, so those three quarters render as a hatched break in the trend rather than a false drop to zero.
◐ curated → ● full
What Went Wrong closed its last illustrative gaps and is now fully real. Phase of flight was already counted from CAROL's cicttPhaseSOEGroup involvement coding; this promotion makes probable cause and the per-100,000-flight-hour accident rates real too. Cause now comes from the NTSB's bulk aviation accident database (avall.mdb, read with mdbtools): the coded CICTT defining occurrence per accident (Events_Sequence, Defining_ev=1), crosswalked to cause families - so the fatal-GA breakdown (loss of control 47%, powerplant 15%, CFIT, weather, fuel, maneuvering) is a straight count off the record, not an NLP guess. The rates are real for all three operating-rule worlds: general aviation (5.2 per 100k hours, 0.9 fatal) and Part 135 (1.1 / 0.2) use FAA General Aviation & Part 135 Activity Survey flight-hours; Part 121 (0.16 all / 0.0023 fatal) uses the NTSB's own air-carrier accident-and-flight-hour series (2014-2023, hand-downloaded because BTS/NTSB WAF-block automated fetch). The Part 121 rate uses NTSB's all-operations count and hours (282 accidents / 176.6M hours), a broader basis than the site's US-only CAROL accident cards, and the Methodology explains the two counts. No figure on the page is illustrative anymore.
○ illustrative → ◐ curated
What They Say vs What You Get flipped to real FCC Broadband Data Collection data (December 2025 filing), pulled through the broadbandmap.fcc.gov download API with a free account token. Two products drive the page: the Summary by Geography Type file (unit-weighted over the full Location Fabric of 162.4M broadband-serviceable locations) and the location-level availability files - about 116.5M distinct locations aggregated from roughly 973M provider filings across every state and technology. The measured story is the site's own thesis, now real: the map shows 99.99% of locations served at 100/20, but that counts a low-orbit satellite dish - strip satellite out and only 92.8% have a wired option, leaving 11.8M locations with none. 30.1% of locations have one wired provider or none (5.3% have zero); 4.5% can reach 100/20 only via satellite; fiber is available to 59.9%. Urban/Rural/Tribal without a wired 100/20 option run 2.5% / 23.5% / 25.7%. A build-time SVG state choropleth ranks single-provider share (Alaska, West Virginia, and - surprisingly - Massachusetts run deepest). The single-provider vignettes stay illustrative archetypes, and the advertised-vs-measured delivery bars stay illustrative gap tiers (real measured throughput needs FCC Measuring Broadband America); each is badged. Maturity is curated, not full, for that reason.
○ illustrative → ● full
What Got Filed, and by Whom swapped to real USPTO Open Data Portal data at the data layer - the PatentsView Granted Patent Disambiguated Data product (PVGPATDIS), whose disambiguated bulk tables (grants, assignees, inventors, locations, CPC classifications) now drive every count, share, sector lean and inventor tally. 324,040 utility patents were granted in 2024; Samsung Electronics leads at 9,555, ahead of TSMC (3,985) and Qualcomm (3,422), and IBM has slipped to seventh (2,461, down 33% year over year) after decades atop the list. Physics and computing (CPC section G, 30.6%) and electricity (H, 26.7%) dominate the classification mix, and the most prolific living inventors surface for real (Shunpei Yamazaki, 6,787 lifetime grants). The 18 mapped metros are a curated selection of major invention centers ranked by their real 2024 counts, disclosed as such - a few metros with more grants (Washington, Denver, Miami) sit off the map. Assignee family display names, geography anchor employers, sector examples and inventor field labels stay curated overlays. (PatentsView moved behind the USPTO Open Data Portal API in March 2026; the old keyless S3 dumps are gone, and the site's stale 'legacy S3 remains live' note was corrected.)
○ illustrative → ● full
The multi-year assignee cut went real on the same PatentsView PVGPATDIS tables: 1,638,388 utility grants across 2020-2024, with every firm's five-year trajectory, share, CPC-portfolio mix and the compare tool computed from the record. The measured story is a changing of the guard - IBM's annual grants collapsed from 9,128 in 2020 to 2,461 in 2024 (down 73%) as it abandoned its decades-long lead, while TSMC climbed 41% and Qualcomm 50%; Samsung Electronics holds the top spot across the whole window. National utility grants dipped from 351,997 in 2020 to 312,067 in 2023 before recovering to 324,040 in 2024. Editorial labels - assignee display names, geography anchor employers, sector examples, and which 18 metros the map shows (disclosed as a curated selection, not the national top 18) - stay curated overlays.
◐ curated → ● full
The last illustrative section, Schedule Padders, went real - taking flight-ontime to full. A second BTS Reporting Carrier On-Time Performance month a decade back (June 2010, 551,687 flights) joined the real June-2024 record (611,132 flights); the pipeline partitions flights by era and computes the median scheduled block time (CRSElapsedTime) per city pair for now versus then, so the padding airlines quietly added to the same routes is measured, not sketched. Among busy major-airport pairs the schedule has stretched roughly 18 to 22 minutes since 2010 (San Diego-Newark +22, Miami-San Francisco +20, Miami-Chicago +19). Every section on the page now computes from the real BTS record - national on-time 75.1%, cancellations 1.29%. The decade-earlier month is loaded only to build the baseline and is not counted in the records total.
○ illustrative → ◐ curated
What the Pitch Was Doing swapped to real MLB Statcast at the data layer: 2,868,564 pitches pulled a day at a time from Baseball Savant across four seasons (2015, 2019, 2023, 2024, with 711,714 in the latest). Every physical reading is now real straight from the feed, velocity, spin, whiff, movement, xwOBA and the arms-race trend, and run value per 100 is joined from Savant's pitch-arsenal-stats leaderboard. The page also doubled from nine readings to seventeen, opening a second act on the batted ball and the delivery, all real from the 2024 feed: the barrel map (launch angle by exit velocity, colored by xwOBA), batted-ball value, the softest-contact and luck leaderboards, the plate-location whiff heatmap, the count-leverage grid, the arm-slot release map, and a deception plot of perceived-velocity gain against extension. One documented stand-in holds it at curated: Baseball Savant's proprietary Stuff plus model is in no public export, so the Stuff plus index is a transparent proxy from real velocity and whiff, centered on the league-average 100 and badged, never quoted as the Savant model.
○ illustrative → ◐ curated
Who Lends to Whom swapped to real CFPB/FFIEC HMDA loan-level data at the data layer: the 2024 activity year, 12,019,590 covered records across 50 states and DC, pulled keyless from the FFIEC Data Browser. Denial rates by race, state, metro, income band, and reason now compute from the real loan-application record on a first-lien owner-occupied home-purchase slice using the FFIEC applications-acted-upon denominator. The national Black-White denial gap is 10.5 points (Black 19.7 percent against White 9.2), the reason ordering is debt-to-income, then credit history, then collateral, and the Deep South leads the metro gap table rather than the industrial North. Rates run above the 2020-2023 figures because 2024 was a high-rate, tight-credit year; the shape (a roughly two-to-one Black-White ratio, persistent across income and geography) holds. The gap-over-time and market-volume-by-year plates need multiple activity years and stay illustrative, badged, pending a multi-year ingest.
○ illustrative → ● full
The American Health Map swapped to real CDC Behavioral Risk Factor Surveillance System prevalence at the data layer: obesity, current smoking, diabetes, and cost-barrier, statewide crude prevalence across 51 jurisdictions for 2011-2024, pulled keyless from the CDC Socrata release (dttw-5yxu). The measured gradients are the ones the page always drew: West Virginia carries the heaviest adult obesity at 41.4 percent against Colorado's 25.0, and Utah's adult smoking floor is 5.7 percent. Nationally in 2024, crude adult obesity sits at 34.3 percent, current smoking at 11.5, and diabetes at 12.3, each up from its 2011 baseline. One honest seam: CDC published no national cost-barrier figure for 2011-2012, so the shared four-burdens national trend runs 2013-2024 while the state maps and the decade obesity and smoking baselines keep the full 2011 start.
○ illustrative → ● full
The dashboard swapped to real PHMSA Pipeline Incident Flagged Files at the data layer: the modern-era 2010-2025 reports across all four system types (hazardous liquid, gas distribution, gas transmission and gathering, LNG), 9,409 reported incidents ingested of which 4,595 are flagged significant and drive the page. Independently re-summed from the raw workbooks: 4,595 significant incidents, 180 deaths, 781,391 net barrels lost, and 10.6 billion dollars of property damage over the span, about 300 significant incidents a year. The commodities diverge in a real way: gas transmission logs the most incidents, high-volatility liquids spill the most barrels, and gas distribution takes the most lives. Texas leads at 1,350 incidents, then Louisiana, California, and Oklahoma. Pre-2010 workbooks reach back to 1986 but use incompatible older report forms, documented as future backfill.
○ illustrative → ● full
Where the Tuition Goes swapped to real NCES IPEDS bulk at the data layer: the Finance (F1A/F2/F3), directory, 12-month enrollment, institutional characteristics, and student financial aid files for 5,837 Title IV finance filers, FY2021-22, with the trend reaching back to 2002 and constant-dollar prices to 2010. The education dollar breaks down to 38 cents instruction, 15 cents administration, and the rest across research, student services, and support. Measured over two decades in constant dollars, administrative spending grew 25.3 percent against instruction's 11.3, and student services grew fastest at 37.6. The real numbers overturned three of the old illustrative narratives and were rewritten: the instruction share round-tripped rather than sliding monotonically, and private net price actually fell from 25.9k to 22.7k as the sticker-to-net discount widened. Shares are of core education-and-general spending, and the multi-year trend uses the public and private-nonprofit sectors that carry a consistent expense schedule.
○ illustrative → ◐ curated
Where the Prescriptions Go swapped to a real CMS Medicare Part D DY2023 ingest (Spending-by-Drug plus Prescribers-by-Provider-and-Drug) at the data layer, driving seven of nine readings: the national headline, the opioid map, the state spend distribution, the drug leaderboard, the price-versus-volume scatter, concentration, the 2013-2023 trend, and opioid claim share by prescriber specialty. Gross 2023 Part D spend is 275.8 billion dollars, Eliquis leads at 18.27 billion, and the top ten drugs take 26 percent of spend. Two readings stay curated stand-ins: the therapeutic-class mix and the class slope, because CMS ships no therapeutic-class column and there is no reproducible branded-tier boundary; a raw all-drug rollup would put a 55 percent other bucket on top and invert the section, so it is labeled rather than forced.
○ illustrative → ◐ curated
What Sent You to the ER swapped to a real weighted CPSC NEISS ingest (2013-2017 public-use microdata, 1,865,651 sampled case records carrying statistical weights) at the data layer, driving six of eight readings: the body map, the product leaderboard, the diagnosis ranking, the age curve, the age skew, and the triage split. The weighted 2017 estimate is 14.7 million product-related ER visits, matching the known NEISS scale; floors, stairs, and beds outrank every sport and machine, and the age-rate curve is U-shaped with the 75-plus rate two and a half times the 25-to-44 rate. The ingest also corrected four mislabeled illustrative product codes against their real NEISS titles. Two sections stay stand-ins: the 20-year trend, since the obtainable export covers only 2013-2017 while the component is authored around a 2005-2024 arc, and the recall clock, which needs a separate CPSC Recalls feed.
○ illustrative → ◐ curated
Where Your Donation Goes swapped to the real IRS 2024 Form 990 e-file release at the data layer: 728,704 returns parsed from the bulk XML, 337,154 of them full 990s carrying the functional-expense split, joined to the Exempt Organizations Business Master File for category and state. The sector dollar splits 86 cents to program, 11 to management, and 3 to fundraising; the population median overhead is 12 percent, with 44 percent of full filers under 10; and the leaderboard is now real, recognizable names (Kaiser Foundation Health Plan, Mass General Brigham, UPMC, Cleveland Clinic, Mayo, Johns Hopkins, NYU). The multi-year overhead trend stays a curated stand-in because a single e-file release is a one-year snapshot, not a decade of history; pay medians are defined over filers reporting a paid officer, since including all-volunteer micro-charities collapses them to zero.
◐ curated → ● full
The last badged stand-in went real: the industry unrealized-loss-on-securities series in the Rate Trap section, previously tracking the FDIC Quarterly Banking Profile aggregate, is now rebuilt from every filer's own call-report securities marks. Aggregating FDIC Schedule RC-B across all insured institutions per quarter (available-for-sale fair value minus amortized cost, plus the held-to-maturity marks) reproduces the industry series within about 1 billion dollars of the published QBP across all twelve quarters, from the 689.6 billion dollar unrealized-loss peak in 2022 Q3 down to 481.2 billion in 2024 Q4. With that series real, every figure on the page is a real FDIC ingest and the dashboard is full: 56,482 institution-quarter call-report rows. A fetch-recipe fix also recovered 11 tiny banks that a trailing-newline bug had dropped at quarter boundaries.
○ illustrative → ◐ curated
The current BLS OEWS vintage swapped in at the data layer: May 2025, 370,172 occupation-by-area-by-industry estimate rows. The wage-spread leaders, the percentile combs, the compare-page baseline, the pay-band histogram, and the Registered Nurses industry, metro, and state readings now all read published estimates. Registered nurses show a national median of 97,550 dollars and the all-occupation median is 50,980, both consistent with BLS published figures, and the widest spread is athletes at 28 times between the 10th and 90th percentile. It stays curated rather than full because OEWS is not maintained as a time series (BLS warns against comparing vintages), so two multi-vintage panels, the nominal-versus-CPI wage trend and the p10 versus p90 growth chart, stay labeled illustrative stand-ins pending archived per-year files that are currently bot-blocked. A stale top-code assumption was also corrected: BLS now publishes uncensored upper percentiles.
○ illustrative → ◐ curated
The real FEC 2024 individual-contribution file streamed in at the data layer: 58,208,756 itemized rows, aggregated to 9.87 billion dollars of itemized individual money after de-duplicating ActBlue and WinRed conduit passthrough (12.15 billion gross). The state map, the industry leaderboard, donor concentration and size mix, the money-flow split, and the 2024 endpoint of every trend chart are now measured; per-state totals were validated against independent passes over the 11 GB file (California 1.585 billion, Texas 927 million, Wyoming 240 million). A measured surprise replaced a hardcoded claim: Wyoming, not DC, leads the nation in itemized dollars per adult at 542 dollars, a tax-residency artifact. It stays curated because the pre-2024 trend history each needs its own multi-gigabyte cycle file (a later pass) and the per-adult denominator still uses a curated 2020-Census voting-age table; those are labeled.
○ illustrative → ◐ curated
The dashboard now builds live from the USAspending REST API aggregations rather than illustrative stand-ins. Real FY2024 prime awards total 2.0 trillion dollars, contracts 740.8 billion plus grants and other assistance 1,242.8 billion; the per-capita geography is real (DC 52,128 dollars per resident, Virginia 12,730, the contractor-beltway and national-labs pattern), the sector buckets come from real PSC codes, and the contractor leaderboard is parent-rolled, with Lockheed Martin first at 63.4 billion, 8.6 percent of contract dollars. Agency totals are real too, HHS the largest at 809 billion. It stays curated rather than full because two displayed series, the recipient concentration Lorenz curve and the distinct-recipient counts, cannot be produced exactly by the aggregation endpoints and remain modeled estimates pending a full recipient-roll ingest; both are labeled modeled on the page.
○ illustrative → ● full
The whole dashboard swapped to the real Census Building Permits Survey at the data layer: state and metro annual bulk files 2010-2025 plus Census Vintage 2024 population for the per-capita denominators, 8,213 place-year rows parsed. The measured story is the one the page always argued, now on real numbers. Nationally 1.43 million units were permitted in 2025, about 64% single-family, down from the 1.74 million 2021 peak. Read per resident the map turns Sun Belt and Mountain West: Idaho leads the states at 9.0 units per 1,000 residents and Myrtle Beach leads the metros at 15.8, while the big coastal metros sit on the floor near 1.7 to 2.2 (New York, Los Angeles, Chicago, San Francisco). Texas (210k units) and Florida (178k) build the most in raw volume. Two things stay editorial and are labeled as such: which 25 metros the leaderboard tracks and the boomtown/steady/frozen archetype tag on each. The archetype trajectories now plot each cohort's real summed permits indexed to 2010; measured off that post-crash trough all three roughly doubled or more, with the Sun Belt cohort pulling to about 3.5 times its 2010 pace while the steady and supply-frozen cohorts landed near 2.2 times.
○ illustrative → ◐ curated
A validation pass found the dashboard's figures were already real, hand-transcribed NCHS/CDC mortality data - and more current than any keyless machine source - so it moves off the illustrative badge to curated. The national age-adjusted drug-overdose rate matches CDC's published series exactly, 6.1 per 100,000 in 1999 rising to 32.6 in 2022, and the 2022 final leading-causes ledger is exact: heart disease 702,880, cancer 608,371, COVID-19 186,552, drug overdose 107,941, suicide 49,476. It stays curated rather than full because a fully automated ingest is blocked - CDC WONDER, the only current and complete machine source, is query-only and needs interactive data-use acceptance - and two smaller series stay labeled curated approximations: the state-level alcohol-induced split and the years-of-life-lost shares.
◐ curated → ● full
The last missing join landed: real Census median household income for 335 metros, so the two income-decoupling sections now run on measured data and the site is full. The 2000 baseline that stalled the first attempt stayed blocked - the Census API now needs a key, and 2000 incomes sit under pre-2003 metro geography that will not crosswalk to today's codes - so income is honestly anchored to the window it does have: ACS 5-year median household income (table B19013), 2015 to 2023, pulled keyless from the Census summary-file bulk downloads. To keep the comparison fair, the price figures in those two charts are recomputed over the same 2015-to-2023 span; the rest of the site keeps its since-2000 window, and both are labeled. The measured story holds and sharpens: nationally real prices rose 33% against a 13% income gain, and 319 of the 335 metros saw prices outrun paychecks, widest in Homosassa Springs, Florida, at 85 points. The exceptions are real too - a handful of expensive coastal metros like San Jose and San Francisco where prices had already plateaued and incomes edged ahead.
○ illustrative → ● full
The real BLS QCEW annual-averages singlefiles swapped in at the data layer - every county, state, and supersector wage cell for 2015-2024, 36.0 million rows streamed. The national average weekly wage is $1,454, but the median county pays just $992 and only 152 of 3,143 counties clear the average: the mean is a short list of counties averaged over everyone else. Santa Clara ($3,789), San Mateo, and San Francisco lead; Texas border counties like Hidalgo ($825) and rural Southern counties anchor the floor, about 4.6 times apart for a week of the same length. The metro lead over nonmetro counties widened from $265 a week in 2015 to $454 in 2024. Two curated layers stay labeled: the leaderboard is scoped to counties with at least 50,000 covered jobs so a lone mine or lab cannot buy a slot, and the 'Six Americas' panels group real counties into editorial archetypes whose plotted paths are their real employment-weighted QCEW wage. Scope is the 50 states plus DC.
○ illustrative → ◐ curated
A real FDIC BankFind Suite ingest swapped in at the data layer - the current census of 4,262 insured institutions holding $26.2T, plus a per-bank call-report panel across 2022-2024 and the real failure record. The measured picture is stark concentration: 16 megabanks over $250B hold 61.5% of all assets while 3,217 community banks under $1B hold 4.1%, and JPMorgan alone ($4.0T) outweighs thousands of the smallest combined. The 2023 shakeout is real too - Silicon Valley, First Republic, and Signature failed with $532B in combined assets, and industry net income shows the 2023 Q4 special-assessment trough at $36.1B. It stays curated, not full, for one badged stand-in: the industry unrealized-loss-on-securities series still tracks the FDIC Quarterly Banking Profile rather than being rebuilt from each bank's securities marks.
○ illustrative → ◐ curated
A real NCUA 5300 Call Report ingest swapped in at the data layer - the 2024 Q4 filing for all 4,550 federally insured credit unions, 143.7M members and $2.33T in assets. The structural roll-ups are all measured now: Navy Federal leads at $180.8B and 14.3M members, the 20 credit unions over $10B hold 24.2% of assets, and 53.7% of all credit unions carry a low-income designation. It stays curated, not full, because several series a single quarter cannot produce remain badged curated stand-ins - the 1994-2024 consolidation trajectory (only its 2024 endpoint is live), the field-of-membership mix, and a penetration map whose member counts are real but whose population denominator is curated.
○ illustrative → ◐ curated
The real FHFA House Price Index swapped in at the data layer - all ~184,000 all-transactions index rows. Every price figure is now measured: the US ran 8.8x nominal since 1975 but just 1.44x real, and the state map, the 410-metro leaderboard and distribution, the decade splits, and the 2007 drawdowns are all computed straight from the file. Two windows by design - the national arc is the full 1975-2025 record, while metros and states read since 2000, because the index reaches 1975 for only 17 metros but all 410 from 2000. It stays curated, not full, because the income-decoupling sections need a real Census median-household-income join (the Census API now requires a key, and 2000 incomes sit under the old pre-2003 metro geography); those two charts are hidden until that lands rather than shown against stand-in incomes.
○ illustrative → ● full
The full NHTSA FARS census swapped in at the data layer - every fatal crash on a U.S. public road, 1975-2024 (1,872,464 of them), plus the linked person record. Two clocks run on purpose: the fifty-year arc and the occupant/non-motorist divergence run through 2024, while every snapshot is read at 2023, the latest final year. The real numbers sharpened the whole page: the national rate is 12.2 deaths per 100k, but Mississippi's roads (24.9) run more than five times Massachusetts's (4.8); rural roads carry 41% of the dead on 31% of the miles, 1.66 vs 1.07 deaths per 100M miles; and since 2010 deaths outside a vehicle are up 67% while occupant deaths moved 10%. Two approximations the stand-in leaned on became exact - alcohol now derives from driver BAC in the person file (DRUNK_DR was dropped from the modern crash file), and the pedestrian/cyclist toll is the real victim-level count. A new chart, 'the shape of the year,' shows the season of road death: about 123 a day in September against 98 in January.
A new dashboard, shipped fully real from birth on the EFF and UNLV Atlas of Surveillance - the whole public CSV ingested and counted directly: 15,071 documented surveillance-technology deployments across 7,747 U.S. law-enforcement agencies. The lead finding is a near-monopoly: 64% of the 4,084 documented automated-license-plate-reader deployments are a single vendor, Flock Safety, with the next-largest (Vigilant Solutions) a distant 130. Deployments turn up in 49 of 50 states plus DC, and the deepest documented stacks - Chicago and Houston police, 10 distinct technologies each - run plate readers alongside body cameras, drones, face recognition and cell-site simulators. Badged full for a complete ingest of the source file, with a standing caveat the source itself under-counts: it is documentation, not a census, so every figure is a floor.
○ illustrative → ● full
The real DOL Wage & Hour enforcement file (the WHISARD case system) swapped in at the data layer - all 365,393 concluded federal compliance actions from FY2005 forward, pulled case by case through the department's Open Data API. Every figure is now measured: $4.7B in back wages found owed to 4.8M underpaid workers. The real numbers sharpened the thesis and broke the illustrative axes both. Recovery peaked early, at $305M in 2011, then drifted down through the 2020s to the thinnest enforcement in two decades; the pandemic dip in FY2020 is visible but is not the whole decline. The ledger is savagely lopsided: 44% of concluded cases close with no back wages at all, while the heaviest 2.2% of cases carry more than half of every dollar owed (a concentration index of 0.88). Construction and food service lead the industries, unpaid overtime and sub-minimum wages under the FLSA are 78% of the money, and 9% of all back wages trace to employers already flagged as repeat violators.
○ illustrative → ● full
The real U.S. Department of Education College Scorecard institution file swapped in at the data layer - every net-price, debt, earnings and graduation figure now pulled straight from the government's cells. The featured cast stays a curated roster of 20 recognizable schools (MIT to Cal State LA), each matched to its real row by federal UNITID; the sector shares are computed over all 5,127 institutions the Scorecard reports earnings for, and they sharpened the thesis: 59% of for-profit colleges leave their typical student earning less than a high-school graduate, against 7% of publics. Real outliers broke the illustrative axes - MIT's $143k median earnings had overrun the old $130k ceiling - and at the wealthiest schools a low-income student's net price runs to zero or below, grant aid beating the entire cost of attendance.
○ illustrative → ● full
The full CFPB Consumer Complaint Database swapped in at the data layer - all 16,650,879 complaints, streamed row by row so the 9 GB file never lands in memory. The real data reframed the page's finding: 'relief' is overwhelmingly a corrected credit file, not money. Only about 1% of complaints close with money back, and the three credit bureaus - which draw four in five of all complaints - return money effectively never, so the company charts now measure monetary relief. Credit reporting's annual complaints turn out to have blown past every axis the illustrative stand-in drew (4.8M in 2025 alone), and the in-progress current year is held off the trend lines so they don't dip.
○ illustrative → ● full
Wave 2, swapped in at the data layer. storm-events now reads all 2,030,218 NOAA Storm Events records back to 1950 - where heat, not tornadoes, turns out to be the deadliest hazard; recalls the 86,419 openFDA enforcement actions across drug, food, and device; hazmat-transport 122,229 PHMSA Form 5800.1 incidents from 2020-2024; mine-safety 273,607 MSHA accident records against 2.7M quarterly employee-hour filings since 2000; air-quality the full EPA AirData county series, 2000-2024, over which bad-air days per county fell 74%. Each chart was rescaled to the real ranges its illustrative stand-in never reached, and captions that still called real numbers illustrative were corrected.
○ illustrative → ● full
Real bulk data swapped in at the data layer, components untouched. disasters now reads all 70,048 OpenFEMA disaster declarations; rail-accidents the 224,704 FRA Form 54 records since 1975; nursing-homes the 14,695 CMS-certified facilities; toxic-releases the full EPA Toxics Release Inventory - 3.2M facility-chemical rows across 1987-2024, so the decline from 18.7B to 3.3B pounds is measured, not sketched.
The collection went public at limn.kvlak.com. Three shipped on real bulk ingests from day one (wildlife-strikes, hospital-compare, tracker-atlas); every other dashboard is badged curated or illustrative so you always know how real each figure is before you trust it.