August 27, 2026

Sixteen Percent of What?

Primary source: CMS, “Calendar Year (CY) 2027 Medicare Physician Fee Schedule Proposed Rule” fact sheet, proposed rule issued July 14, 2026 (comments close September 14) · MIMI Labs: CMS Medicare Physician & Other Practitioners — by Provider and Service, calendar year 2024

G2211 is the office-visit complexity add-on: a flat ~$16, the same $16 whether the visit underneath was a 99212 or a 99215. For CY 2027 CMS proposes making it a modifier worth 16% of the base E/M instead — and 32% if you practice inside a Shared Savings Program or LEAD ACO. This plots all 45 specialties that billed it in 2024: 24.5 million real add-on claims from 138,770 distinct NPIs, positioned by the average allowed amount of the office visit the percentage would multiply against their G2211 attach rate, sized by volume. Drag the modifier and a break-even line sweeps across the field, flipping dots from navy to red. At the proposed 16% the pool grows +14.7% and only dermatology and podiatry lose. At 32% it grows +129% — the ACO-only increment alone is larger than the entire program today. The built-in critical lens is the min-billers slider: drop most of the dots and watch the dollar total barely move, because counting specialties and counting money are not the same chart.

4 min · D3 data explorer

Nobody Ordered Less Nursing on Sunday

Primary source: National Nurses United, “NNU nurses to participate in nationwide protests against Palantir”, August 2026 · MIMI Labs: CMS Payroll Based Journal — Daily Nurse Staffing, 2024 Q3 (July 1 – September 30, 2024), 14,548 facilities

Nurses are picketing an allocation model today — nurse scheduling, staffing automation, bed capacity — not a clinical model. There is no public dataset of scheduling-model outputs, but there is one place American healthcare publishes what it actually bought, day by day, from payroll: the CMS Payroll Based Journal. Here is a full quarter of it as a 92-day calendar heatmap, national, animated. Every weekend is a stripe. RN hours per resident-day fall 11.9% Saturday and Sunday; the number of residents in the building falls 0.2%. July 4th, a Thursday, staffs 9.1% below every other Thursday in the quarter. And the cut gets sharper the bigger the facility — −3.5% → −10.0% → −13.2% → −19.7% across four size bands, the opposite of a small-building-can't-fill-a-shift story. The critical lens is honest about the spread: facility by facility the median gap is +8.2% with an IQR of −9.7% to +25.6%, and a third of buildings go the other way.

3 min · animated calendar
August 26, 2026

The Escalation Gap

Primary source: Alexandra Byrne, “AI software to detect hospital drug diversion has a problem: Humans”, STAT News, August 25, 2026 · Market and disclosure facts: Brett Kelman & Darius Tahir, KFF Health News, June 3, 2026 · MIMI Labs: CMS Timely and Effective Care – Hospital (dataset modified July 22, 2026; data refreshed August 13, 2026), reporting period 2024–01–01 to 2025–06–30, joined to Hospital General Information for ownership and type

A travel nurse at Adventist Health in Bakersfield spent late September 2024 taking patients' IV opioids and charting them as given. The hospital ran machine-learning software built to catch exactly that. It fired; managers ignored the alerts, auditors found. Nobody publishes how often diversion alerts get ignored — hospitals aren't required to disclose that they run the software, let alone report when it fails. So this plots the one place CMS does publish what happened after a clinical trigger fired: all 3,084 hospitals with both a sepsis start rate and a sepsis finish rate. Horizontal, the first move — the 3-hour severe sepsis bundle, mean 79.7%. Vertical, the completed composite — mean 62.9%. Everything sits below the y = x line: 98.6% of hospitals finish worse than they start, and a third of them by twenty points or more. Both axes swap — septic shock bundle, median ED stay, left-before-being-seen, concurrent opioid prescribing — and the sample-size floor is the built-in critical lens: drag it and the hospitals pinned at 0% and 100% dissolve, spread falling from 22.8 to 14.0 between the under-30-case and over-300-case groups, while the gap barely moves. It also says what it cannot see: 1,554 of 4,660 hospitals report no SEP_1 at all, this is chart abstraction rather than behaviour, and there is no CMS measure anywhere for time-to-acknowledge or close-without-comment — every number here describes the outcome of an escalation path, none of them describe the path.

D3 data explorer

The Drop

Primary source: Alexandra Byrne, “AI software to detect hospital drug diversion has a problem: Humans”, STAT News, August 25, 2026 · MIMI Labs: CMS Timely and Effective Care – Hospital, measures SEP_1 and SEV_SEP_3HR, reporting period 2024–01–01 to 2025–06–30

One hairline per American hospital — 3,084 of them, on one screen — drawn from how often the sepsis bundle gets started down to how often it gets finished, and then animated so you watch the whole field fall. The navy band is the 3-hour bundle landing at 79.7%. The red band is the composite closing out at 62.9%. 98.6% of the lines drop, the mean drop is 16.8 points, and 1,043 hospitals lose twenty or more. Sort by the size of the drop, by completion rate, or by how many sepsis cases the hospital reported, and hover anywhere in the field for a named hospital with its two rates, its drop and its denominator. The counterweight is a single toggle: hide the hospitals reporting fewer than 30 cases and every 0% and 100% outlier disappears — standard deviation 22.8 in that group versus 14.0 above 300 cases — while the drop moves only from 19.0 to 15.4. That is the finding: the noise was denominator and the gap is not. Nonprofit 16.6, for-profit 16.6, government 17.6. Sepsis detection is a solved problem and diversion detection is a solved problem; what both stories have in common is that everything downstream of the firing was unowned, uninstrumented and invisible to the vendor.

Graphical narrative
August 25, 2026

One Sixty-Four

Primary source: Edward Graham, “VA deploys new EHR at 3 Indiana-based medical facilities”, Nextgov/FCW, August 24, 2026 · Data: VA EHR Modernization — official Deployment Schedule (tables updated August 24, 2026), all 45 sites VA has published a date for · VHA About Us (updated June 2, 2026) for the alternate denominators

Three more VA hospitals went live on the federal EHR on Saturday — Indianapolis, Fort Wayne, Marion, plus 16 associated clinics, 102,000 veterans and 6,000 staff. Deputy Secretary Paul Lawrence's line is true and worth reading twice: 11 hospitals in 2026, nearly doubling the previous five years. Every site VA has put a date on is plotted here — 17 live, 28 scheduled, 45 total — as a cumulative staircase from Mann-Grandstaff in Spokane (October 24, 2020) to Jefferson Barracks (November 2027), each dot hoverable with its city, VISN and whether its date is exact or month-only. Against it, the dashed trajectory that reaches the denominator by December 2031. The pace slider is the argument: VA's lifetime average is 2.9 go-lives a year, its best published year is 26 (the 2027 plan), and the 119 undated facilities need 28.8 a year for four straight years to keep the promise. At 26/yr the last site lands June 2032; at 13/yr, 2037. The denominator selector is the built-in critical lens — the same 17 sites are 10.4%, 10.0% or 1.2% depending on whether you use the program's 164, VHA's 170 medical centers, or its 1,380 facilities, and nothing published explains the 164/170 gap. What it says about itself: VA's own schedule page still says “14 live” in prose while its tables list 17; a “site” is not a constant amount of work (three hospitals carried sixteen clinics); 2027 dates are month-only; and everything past November 2027 is arithmetic on an empty schedule, not a plan.

D3 data explorer

The Fill

Primary source: Edward Graham, “VA deploys new EHR at 3 Indiana-based medical facilities”, Nextgov/FCW, August 24, 2026 · Data: VA EHR Modernization — official Deployment Schedule (tables updated August 24, 2026)

One dot per VA medical center, 164 of them, and a play button. Eleven years of the federal EHR rollout in twenty seconds: Spokane alone for seventeen months, four sites in 2022, the April 2023 pause that freezes the field for three years, then the 2026 restart filling in waves — Michigan in April, Ohio and Kentucky in June, Indiana on Saturday. A navy ring rides ahead of the red dots marking where the count would have to be, at that moment, to reach every facility by 2031. Then in November 2027 the fill stops — not because the program stops, but because that is the last date VA has published. 119 dots stay gray and the requirement line keeps climbing without them. Scrub the timeline or hover any dated dot for its name, city and go-live. The money moved the same week: the Oracle contract took on three option years and $17B, from the original $10B signed with Cerner in May 2018 to just under $27B, with the remaining ceiling anticipated exhausted by Q1 FY2027 — roughly $165M per medical center, arithmetic rather than accounting, since VA publishes no per-site spend. Counterweight built in: the reflex reading of a gray field is failure, and that reading is wrong — 2026 is comfortably this program's best year. The real signal is that the published plan and the public promise stop touching each other in November 2027.

Graphical narrative
August 24, 2026

Ninety-Nine Point Six

Primary source: Kevin O’Leary, “The Impacts of the No Surprises Act: a San Antonio employer goes over budget”, Health Tech Nerds, August 21, 2026 · MIMI Labs: CMS Federal IDR Public Use File — out-of-network emergency & non-emergency dispute line items, vintage 2026–01–21, reporting period 2025–01–01 to 2025–06–30

Health Tech Nerds reported that San Antonio's 28,000-life city plan is $40M over a $250M budget because a four-location freestanding ER chain wins its federal arbitrations 99.6% of the time. The federal public use file confirms the number to the decimal — West Prestige Emergency Room, 523 line items, 99.6%; all four locations together, 2,566 line items at 99.1% and a median winning offer of 3.0× the QPA. Then it shows you why that isn't the story. Across the same six months, 18,850 entities pushed 2,649,878 line items through arbitration and providers won 87.7% of them; plans offered a median of exactly 1.00× the benchmark, providers asked 4.53×, and the award landed at 3.98×. A brushable scatter of 524 entities — volume on a log axis, win share on the vertical, dot size the award multiple, Prestige's four locations ringed — with state and specialty filters. The stress test is the min-line-items slider: 9,440 entities in the file have a perfect 100% record, and 8,358 of them decided fewer than ten line items. Require 500 and the perfect share collapses from 69.3% to 0.4% while the average win rate barely moves (81.8% → 84.9%). What it says about itself: multiples aren't dollars, entity names are free text (“EAST PRESTIGE EMERENCY ROOM” is its own row, misspelling included), the window predates the surge the story describes, and CMS's ratio columns contain values up to 2.5 million.

D3 data explorer

The Ratchet

Primary source: Kevin O’Leary, “The Impacts of the No Surprises Act: a San Antonio employer goes over budget”, Health Tech Nerds, August 21, 2026 · MIMI Labs: CMS Federal IDR Public Use File, all nine published vintages, 2023 Q1 – 2025 H1

One animated graphic holding ten quarters of federal out-of-network arbitration, three series at once: volume ×21.3 (62,156 line items in 2023 Q1 to a per-quarter average of 1,324,939 in 2025 H1), provider win rate +17.9 points (69.8% → 87.7%), median award +1.29× the benchmark (2.69× → 3.98× QPA). Not one of the three has a down-trend across the full series. Underneath it, all 2,241,655 winning offers from the latest period binned against the QPA — a bimodal shape with a 255,104-item spike at 1.0–1.5× (the determinations the plan won, because the plan's offer is the benchmark) and a long right tail in which 14.5% of awards land above 12×. The 80/20 read: baseball-style arbitration was picked to push both sides to the middle, but only one side can move — the plan's number is fixed by what it already pays in-network, the provider's is whatever it wants to charge — so 87.7% isn't plans arguing badly, it's the arithmetic of a game where the only legal move is the losing one. At a mean $635 IDRE fee and 90 business days per determination, the administrative layer alone runs ~$1.7B per half-year before a single claim is repriced, and it lands on self-insured employers, not on the insurer whose name is on the card. Counterweight built in: the last bar is two quarters averaged, the award ratio is unpublished for 2024 Q3–Q4 and drawn dashed, and a QPA multiple is not a dollar amount.

Graphical narrative
August 23, 2026

The Resolution Gradient

Primary source: HCAHPS Survey Instrument, V19.0 (English mail), effective for January 1, 2025 discharges and forward · hcahpsonline.org · MIMI Labs: CMS Care Compare HCAHPS Hospital file, survey period 2024–07–01 to 2025–06–30, and Hospital Readmissions Reduction Program, vintage 2026–05–01

Today's essay argues the raw material was deposited years ago by a version of you who was only annoyed — and that the habit you actually lost was writing it down. Somebody did write it down. Since 2006 CMS has mailed the same survey to a random sample of every hospital's discharged patients, and five of its questions describe the same twenty minutes at the end of the stay. They are not equally sharp. “Were you given information about what to do at home?” — 85.5% yes, and 55.0% of hospitals sit within ten points of the national maximum. That measure is on Care Compare. It cannot distinguish between almost any two American hospitals. “How often did staff describe possible side effects in a way you could understand?” — 45.1%, with 1.7× the spread. All 2,578 hospitals as a density ladder over the five questions, then a brushable scatter of any one of them against CMS's own 30-day heart-failure excess readmission ratio, with a live least-squares fit and r². The stress test is the min-HF-discharges slider, and it goes badly for the sharp question: the side-effects correlation is exactly zero below 100 discharges, −0.14 in the 100–299 band, −0.03 above 700 — and dragged cumulatively it simply wanders (−0.096, −0.071, −0.034, then back to −0.101 on the last 149 hospitals). A coefficient that will not settle as the measurement improves is the absence of a relationship plus noise. The blunt composite stays between −0.19 and −0.27 at every threshold. What it says about itself: the fan opens partly because the answer scale changes from Yes/No to a four-point “Always” bar, so 45.1% is not “half of patients got no counselling”; the survey and the readmission window are adjacent, not simultaneous; and 2,578 is a join of two suppressed files, not a census of the ~4,500 hospitals that report HCAHPS.

D3 data explorer

The Fan

Primary source: HCAHPS Survey Instrument, V19.0 (English mail), effective for January 1, 2025 discharges and forward · MIMI Labs: CMS Care Compare HCAHPS Hospital file, survey period 2024–07–01 to 2025–06–30 · HRRP, vintage 2026–05–01 · Steven Pressfield, The Arcadian (W.W. Norton, May 26 2026)

One graphic, animated, holding the entire dataset: 2,578 hospitals drawn as 2,578 thin lines across the five CMS questions that all describe the same discharge, ordered from the blunt composite on the star rating to the sharpest question in the instrument. On the left the bundle is a rope — 85.5% mean, SD 3.7 — and by the right-hand axis it has come apart into a fan: 45.1% mean, SD 6.2. Hover any line to follow one hospital across all five. Recolour by readmission outcome and the red lines fail to separate on the sharp axis; recolour by side-effect decile and the top and bottom tenths visibly cross on the way over, because a hospital in the top decile on side-effect counselling is not reliably top on the measure CMS publishes. The clause doing all the work is four words a federal survey writer chose to include — in a way you could understand — which is exactly the kind of sentence a clinician mutters in a hallway and never writes down. Built-in counterweight: against 30-day HF excess readmission the blunt composite carries r = −0.22 and the sharp question −0.09, backwards from what the specificity story predicts, and both weak enough (4.9% and 0.8% of variance) that neither is a lever. The interactive supports the claim that the specific thing is real and declines to support the claim that it matters.

Graphical narrative
August 22, 2026

“Operates in Parallel With the Standard of Care”

Primary source: Emanuel EJ, Baker-Butler A, Khosla N, Khosla V, “Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?”, JAMA, 17 Aug 2026 · counterpoint Axios, 19 Aug 2026 · MIMI Labs: FDA 510(k) premarket notification database and device classification file, vintage 2026–08–17 · 21 CFR 892.2080, 892.2070

Four authors argued in JAMA that autonomous AI will beat both physicians and physician-plus-AI hybrids by 2030 — and that the hybrid is the worst of the three, because human oversight of a good model can make the output worse. The AMA’s CEO said no on the record two days later. Neither side produced data, because the control was never instrumented. So look at what got mandated instead: all 110 FDA 510(k) clearances of imaging and cardiac AI, 2018–2025, as a brushable scatter of decision date against days in review, sized by how many clearances the firm holds and coloured by how far into the read the regulation lets the device reach — 49 that may only flag and notify, 26 that reorder the queue, 20 that mark the image, 10 that assist a diagnosis, 5 that guide acquisition. Every tier is defined by what the clinician does afterward. 21 CFR 892.2080 is explicit: the device “does not remove cases from a reading queue… operates in parallel with the standard of care, which remains the default option for all cases.” Required evidence that the clinician’s review helped: zero clearances. The stress test is the min-clearances-per-firm slider — the “FDA review is getting faster” slope is carried by a handful of repeat filers doing Special 510(k) version bumps, and a least-squares fit recomputes live as you drag until the trend you were about to put in a deck comes apart. What it says about itself: 110 devices is the regulated sliver, not the AI estate — scribes, sepsis models and every clinical LLM are mostly not 510(k) devices at all; the cohort is keyword-derived from twelve product codes; review days measure process, not rigour; and the file ends 2025–12–17, so the right edge is a data artifact.

D3 data explorer

The Repricing Ladder

Primary source: CMS, “CMS Moves to Rein In Misused Medicaid Dollars, Reward Quality Care”, proposed rule released 20 May 2026 (91 FR, 22 May 2026) · statutory basis §71116, H.R. 1 · Holland & Knight alert · MIMI Labs: CMS Inpatient Provider Specific File, April 2026 snapshot · KFF expansion tracker

State directed payments used to be ceilinged at the average commercial rate — the highest benchmark in American healthcare. They are being repriced to Medicare: 100% in expansion states, 110% in non-expansion states, with grandfathered programs falling 10 percentage points every year from the first rating period on or after 1 January 2028. CMS scores it at $775B in total savings over ten years, $510B federal. Here is who actually carries it: 52 jurisdictions and 3,437 IPPS hospital payment records in one dense range chart, each row running from the median hospital’s Medicaid share of patient days to the 90th-percentile hospital — the safety-net tail — coloured by the cap that state is held to. The finding is an inversion: the 110% concession was written for states that never expanded, but not expanding is exactly what keeps Medicaid patient-day share low, so the lower cap lands on the higher exposure — California’s median hospital at 28.5% and 175 hospitals above 30%, New York at 21.7%, Louisiana at 21.2%, all capped at 100%, against Kansas at 11.3% and Wisconsin at 8.9% capped at 110%. A step-down slider walks 2027→2035 and shows the grandfather premium draining 10 points a year to nothing. The stress test is a min-hospitals-per-state filter: Wyoming rests on a single record, Alaska and Vermont on two, and the headline gap between cap tiers is recomputed live as the thin states drop out. What it says about itself: Medicaid patient-day share is a proxy for exposure, not the SDP dollar; the IPSF retains historical providers so this is a payment-record count, not a hospital census; Puerto Rico runs on a block grant and is excluded from both averages; and the step-down slider is statutory arithmetic, not a forecast of your base rate.

Graphical narrative
August 21, 2026

It Could Be $10. It Could Be $10,000.

Primary source: Becker’s Hospital Review, Giles Bruce, “‘Tokenomics’: CIOs scramble to budget for Epic’s AI cost model”, 20 Aug 2026 · MIMI Labs: CMS Medicare Physician & Other Practitioners PUF, performance year 2024, HCPCS 99202–99215, place of service = office

Health system CIOs at Epic’s UGM reached for the same new word: tokenomics. Ochsner’s Amy Trainor, BSN, RN, cannot forecast the bill because the per-feature cost dashboard is only now being built — “It could be $10. It could be $10,000.” Four orders of magnitude is not a forecast, it is the absence of one, and the reason the range is that wide is that the bill is a product of two numbers: a price nobody has published and a volume everybody already knows. Here is the volume. 194,080,306 Medicare office visits across 550,570 clinicians in 73 specialties, every one of them an encounter a scribe or chart-prep agent fires on, as a brushable scatter of visits-per-clinician against annual meter cost on a log axis framed by Trainor’s own $10 and $10,000 rules. Two dials: price per AI call and the one nobody budgets for, calls per encounter. A second chart draws the newsletter’s own advice — the 1×/3×/10× slide — for the twelve busiest specialties, where Dermatology’s contract runs $2,416 at plan and $24,161 if the thing is actually loved. The critical lens is built in three ways: CMS suppresses cells under 11 beneficiaries, so push the clinician floor and Nuclear Medicine (n=43), Undersea and Hyperbaric Medicine (n=31) and Peripheral Vascular Disease (n=44) evaporate as small-n artifacts; the price axis is explicitly fiction, because that is the honest state of the world; and switching the y-axis to % of the visit fee flattens the scatter into a rank order, revealing that a flat per-call price is regressive — it eats the largest share of the cheapest visits, which are Nurse Practitioner ($93.27), Physician Assistant ($90.91) and Podiatry ($95.71). What it says about itself: Medicare fee-for-service only, office place-of-service only, allowed amounts not charges — a floor, not a census.

D3 data explorer

The Coder Inside the Scribe

Primary source: Healthcare IT News, Andrea Fox, “New AI capabilities from Oracle Health target workflow efficiencies”, 20 Aug 2026 (Oracle release 19 Aug) · MIMI Labs: CMS Medicare Physician & Other Practitioners PUF, performance years 2013–2024, HCPCS 99212–99215, place of service = office

Oracle Health’s clinical AI agent now reads the visit conversation and suggests professional-fee charge codes for ambulatory visits, with the clinician confirming before submission. The obvious worry is that a suggestion engine pointed at charge capture pushes levels up. It probably does — but here is what makes that hard to prove: the levels have been going up the whole time anyway. Twelve years of Medicare established office visits animate as a stacked area, and the level-4-and-above share climbs from 44.6% in 2013 to 60.0% in 2024, monotone across eleven of eleven year-over-year steps, including through the 2020 volume collapse. Then the arithmetic: Medicare allowed $85.43 for a 99213 and $119.77 for a 99214 in 2024, a $34.34 gap across 63,712,067 level-3 visits, so a slider that re-levels 5% of them adds $109M and delivers 1.3 years of drift at once. Below it, all 65 specialties at once as 100%-stacked level-mix bars sorted by headroom — Advanced Heart Failure at 93.7% level 4+, Podiatry at 13.5%. The critical lens is the spine, not a footnote: the 2021 E/M documentation overhaul is the single largest confounder for anyone attributing the next few points to an agent; a higher level is not automatically a wrong level, and claims data cannot tell you which; and the small specialties at the top of the ranking are noise in a signal’s costume — Nuclear Medicine at 86.4% on 43 clinicians, Neuropsychiatry on 108 — so push the floor to 2,000 and a genuine clinical gradient appears underneath. What it says about itself: Medicare FFS office visits only, which is the right shape and the wrong size for a feature scoped to one vendor’s ambulatory book.

Graphical narrative
August 19, 2026

The Appeal That Isn’t

Primary source: Epic, “Payers, Providers, and Epic Launch Real-Time Prior Authorization Checks”, 18 Aug 2026 · forcing function CMS-0057-F · MIMI Labs: CMS Marketplace Transparency in Coverage PUF, 2025 release, sheet “Transparency 2025 — Ind QHP” (plan year 2023), issuer-level metrics deduplicated to one row per issuer

Four health systems switched on real-time prior-auth checks inside Epic this week, four months before CMS forces every impacted payer to expose the same API. So here is the baseline that rule is being written against. 424,978,718 marketplace claims. 85,144,260 denied — 20.0%. 376,527 internal appeals filed, 0.44% of denials. 43.8% of those appeals were overturned. Beyond internal review, 5,000 external appeals for the entire individual market for a year, of which 1,576 were overturned — a funnel drawn at true scale, where the last two bars are hairlines. Then all 166 issuers as a brushable scatter: denial rate against internal appeals per 10,000 denials on a log axis, circle area by claims volume, coloured by whether more than half of filed appeals were overturned. Drag the volume floor; flip the y-axis to overturn rate and find that 37 of 153 reporting issuers overturn more than half the appeals they receive. Then two stress tests. Oscar Insurance Company of Florida alone files 221,043 of the country’s 376,527 appeals — 59%; press Drop the top reporter and the national rate collapses from 44 to 19 per 10,000, so any headline built on the pooled figure is really a statement about one payer’s reporting convention. And the ten denial-reason categories sum to 75.7M against a 52.4M denial total — 145%, so they do not partition; the largest single bucket is the one labelled Other, at 46%. What it says about itself: this is claims adjudication, not prior-auth denials; individual marketplace only; plan year 2023; appeals are issuer-level self-reports against a different denominator than the plan-level denial counts.

D3 data explorer

No Predicate

Primary source: FDA, “FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices”, 18 Aug 2026 — docket FDA-2026-N-7874, comments close 19 Oct 2026 · MIMI Labs: FDA 510(k) Premarket Notification database, “Substantially Equivalent” decisions 1 Jan 2015 – 31 Jul 2026

The FDA proposed evaluating generative AI devices through competency assessment — benchmark, then confirm clinically, “inspired, at a high level, by how human clinicians are evaluated and credentialed.” The interesting question is not whether the analogy holds. It is why an analogy was needed at all, and the answer is visible in the machine it replaces. 20,166 substantially-equivalent clearances since 2015 across 1,573 product codes, every one of them the same sentence in a different costume: this new thing is substantially equivalent to that already-legal thing. The 300 busiest codes carry 78.5% of everything, and they are plotted at once as a dot strip chart — median review days on the x-axis, specialty rows, circle area by clearance count, coloured by filing momentum. Push the sample-size floor to 120 and 27 lineages out of 1,573 remain, holding nearly 30% of all clearances. The critical lens is built in: the review-time interquartile range barely moves as the floor rises (66 days to 42), which means the spread is structural rather than small-sample noise — but calendar days from receipt to decision include every hold clock, so it is not agency speed and not rigour either. Switch the colouring to Codes that name AI and six lineages light up, five of them radiology. Then a second chart puts a generative model in the only place the current machine can put it: outside every lineage, with nothing to be equivalent to. What it says about itself: 510(k) only, so De Novo and PMA are excluded by construction; it counts filings, not products; product codes drift.

Graphical narrative
August 18, 2026

The Counted

Primary source: The 229 Podcast (This Week Health), “50 Agents on Paper. An Audit Finds 76.” — Newsday with 229Project, 17 Aug 2026 · MIMI Labs: FDA 510(k) Premarket Notification database joined to the FDA product-code classification file, “Substantially Equivalent” decisions 1 Jan 2015 – 31 Jul 2026 under product codes naming AI, machine learning, deep learning, algorithm, or computer-aided/assisted

Sarah Richardson’s line on Newsday: ask a health system how many agents it’s running, then run the audit and add fifty percent. Fifty on the slide, seventy-six in the building. So here is the universe the federal government actually counts. Every one of 125 AI-coded 510(k) clearances since 2015 — 69 firms, 15 product codes, eleven and a half years, the entire United States — as a beeswarm of product code against decision date, hover for device, firm and K number. That one health system’s 26 uncounted agents is 21% of the national eleven-year total, and almost none of the 26 are devices at all, so none would ever appear. Then two filters take the number apart. Switch the counter to distinct product families and the 125 collapses — Aidoc alone holds 17 clearances, nearly all of them BriefCase version bumps, so the headline is a filing count wearing an innovation costume. Raise the sample-size floor and everything outside two specialties disappears: 112 of 125 are radiology, pathology has one clearance in eleven years, neurology has one and it is a hair-transplant robot. What it says about itself: 125 is a floor, not a census — De Novo and PMA pathways are excluded by construction, as is any AI device cleared under a legacy non-AI product code, and comparing a national device count to one org’s agent audit is a scale illustration, not a measurement.

D3 data explorer

The Retrieval Tax

Primary source: “Evaluating retrieval-augmented large language models for pediatric cardiology knowledge using standardized questions”, npj Digital Medicine, 17 Aug 2026 · Two figures taken from the paper (304 standardized questions; 98.4% top score without retrieval, across eight frontier models); everything else computed in-browser from binomial statistics

Eight frontier models, 304 pediatric cardiology board questions, 98.4% with retrieval switched off — and a single-textbook RAG pipeline added almost nothing. The finding is real. The sharper point is that this benchmark could not have shown you if RAG had worked. All 304 questions render as one grid, five of them red, and those five are the entire headroom retrieval has. At that ceiling and that sample size the smallest lift detectable at 80% power is +1.6 points — exactly equal to the headroom, so a retrieval layer that fixed every remaining error still would not clear the bar. Drag benchmark size, baseline accuracy and claimed lift and watch the power curve move: detecting a +1.0pp gain over a 98.4% baseline needs 1,704 questions per arm, 5.6× the npj benchmark. Then the impolite objection, which is not statistical: a board question is one pre-digested paragraph with every needed fact present and one correct answer, while a real encounter is hundreds of documents that contradict each other. Retrieval is a machine for finding the relevant thing in a pile, and a board question has no pile — so this is evidence about the benchmark, not about your chart. Per-model results are deliberately not reproduced: the paper’s table could not be independently verified at build time.

Graphical narrative
August 17, 2026

Before You Buy the Score

Primary source: Andrea Fox, “Sentara’s Epic EHR model targets MRSA infections early”, Healthcare IT News, 14 Aug 2026 · MIMI Labs: CMS Care Compare Healthcare Associated Infections – Hospital (77hc-ibv8), measure HAI_5, reporting period 1 Jul 2024 – 30 Jun 2025, CMS vintage May 2026

Sentara put an MRSA risk score in Epic, redesigned what nurses do when it fires, and cut hospital-onset MRSA bacteremia ~45%. CMS computes that exact measure for everybody — so here is what the measure can actually see. 4,035 hospitals report it; only 1,728 get a published SIR. The cutoff is visible in the file to three decimals: smallest published denominator 1.002 predicted cases, largest suppressed one 0.998. The 2,307 suppressed hospitals still had 351 real MRSA bloodstream infections across 14.8 million patient days, and none of it appears in any league table. Each of 1,728 circles is a hospital plotted against its own denominator, inside exact Poisson funnel limits: only 9.5% are statistically distinguishable from SIR 1.00 at all, 6.5% from the national 0.69. Then the lens turns on the file — of the 38 hospitals posting a SIR above 2.0, 30 have fewer than three predicted cases a year; of the 317 posting a perfect 0.00, 265 do. Drag the denominator filter to 10 and the maximum collapses while the median barely moves. It also says what the measure cannot see: blood-culture rates vary between hospitals, and a site that draws fewer cultures finds fewer bacteremias.

D3 data explorer

Forty-Five Percent, By Accident

Primary source: Andrea Fox, “Sentara’s Epic EHR model targets MRSA infections early”, Healthcare IT News, 14 Aug 2026 · MIMI Labs: CMS Care Compare Healthcare Associated Infections – Hospital (77hc-ibv8), measure HAI_5_ELIGCASES, 1,728 hospitals with a published MRSA bacteremia SIR

Ten thousand identical hospitals, two consecutive years, zero interventions — and the year-over-year change each one posted anyway. Every dot is a hospital that did nothing; the red ones would have written a press release. At the US median (3.6 predicted cases a year, about 2.5 expected infections), a ≥45% drop shows up by chance 31% of the time and the median hospital swings 50% in one direction or the other. At the 10th percentile the median swing is 100%. Press Sentara, all 11 and the red almost vanishes: their eleven hospitals with a published SIR recorded 37 observed against 48.1 predicted in this same CMS file, and at that denominator a 45% drop falls to 0.8% by chance. Underneath, the real log-scale distribution of predicted cases across all 1,728 scored hospitals, so a reader can find their own site. The counterweight is stated as loudly as the finding: this is a deliberately generous null model — real counts are over-dispersed, so the noise is wider than shown, not narrower — and it says nothing about whether the intervention was worth doing, only that most sites cannot measure it.

Graphical narrative
August 16, 2026

Nobody Keeps the Lamp

Primary source: Epictetus, Discourses I.18.15 (Oldfather trans., Wikisource) — the stolen iron lamp · MIMI Labs: FDA 510(k) Premarket Notification database, cleared submissions 1980–2026, joined to the FDA device classification file for category names and review panels

Epictetus lost his lamp and decided the fear of losing it had been the real cost. The FDA has kept a fifty-year ledger of the same lesson. Every 510(k) device sits in a product code — a category — and somebody is always first into one. 2,772 categories opened since 1980, holding 38,822 clearances. The company that got there first ends up holding 10.7% of them. Each of 1,558 circles is a category with 3+ clearances, plotted at the year it opened against how long the first mover had it alone — log scale, so two weeks and forty years fit on one screen, with a strip along the top for the categories nobody else ever entered. A survival curve underneath answers the only question that matters if something is sitting in your folder: 22% of categories had a second company inside a year, 33% inside two, 50% inside five. Then the lens turns on itself: right-censoring is the most seductive artifact in the file, so drag opened after past 2010 and watch the median gap fall for reasons that are pure calendar; push min-clearances to 15 and two-thirds of the dots vanish. It also says plainly what it cannot show — this measures clearance, not building, and applicant matching is raw text, so an acquisition reads as a rival.

D3 data explorer

One Hundred Lamps

Primary source: Epictetus, Discourses I.18.15 (Oldfather trans., Wikisource) — the stolen iron lamp · MIMI Labs: FDA 510(k) Premarket Notification database, 2,772 device categories opened 1980–2026

A hundred squares and a ten-year clock. Each square is one percent of the 2,772 FDA device categories opened since 1980; navy means the company that opened it still has it alone, red means somebody else has cleared into it. Press play and the field ignites in the order the real distribution dictates: 110 categories had company inside the first month, 612 (22%) by month 12, 923 (33%) by month 24, 1,397 (50%) by month 60, 1,673 (60%) by month 120. Scrub the clock to any month and the readout gives the real count behind the squares. Below it, a 100-bar strip carries the whole argument in one line — of the 38,822 clearances inside categories somebody opened first, the first mover holds 10.7%. The counterweight is stated as loudly as the finding: 935 categories, 34%, have never picked up a second company, and that is the largest single group in the file. The squares are quantiles, not cases — an exact rendering of the distribution and a fictional rendering of any one device.

Graphical narrative
August 15, 2026

The Chokepoint, Counted

Source: Casey Ross & Bob Herman, “FTC reviewing Epic Systems’ use of NDAs, possible anticompetitive practices”, STAT News, 14 Aug 2026 · related: Commure referral-payment terminations, STAT, 13 Aug 2026 · CureIS v. Epic complaint (W.D. Wis. 3:25-cv-00650) · MIMI Labs: CMS/ONC Promoting Interoperability Hospital Public Information file, program year 2023

Investigators are asking whether Epic’s leverage over hospitals blocks rivals from reaching patient data. CMS publishes a file that lets you count the thing being asked about: 4,046 hospitals, 89 certified health IT developers, 306 products. Each circle is a state, positioned by Epic’s share of its reporting hospitals against the number of non-base-EHR certified developers the average hospital there named. Epic hospitals average 0.23; everyone else averages 1.58. Then the built-in lens does something unusual — it makes the finding stronger and then dismantles it anyway. Drag the sample-size filter to 40 hospitals and r goes from −0.66 to −0.82: this is not a small-n artifact. But switch the Y axis to Epic hospitals only and r falls to −0.13; switch to non-Epic only and it falls to −0.01. Inside either group there is no state effect at all — the whole map is a composition effect. Dropping Puerto Rico, the most extreme point in the file, moves r by 0.04. A heat grid holds the full distribution (1,487 of 1,836 Epic hospitals named no third-party developer at all; not one MEDITECH hospital is in that column), a dumbbell chart shows who actually gets in (Commure: 0 of 1,836 Epic sites, 149 elsewhere — but Clinisys runs the other way, 4.7% vs 1.0%), and the size confounder gets tested and fails: the gap is 7.3× in acute care and 6.6× in critical access. The honest ceiling is stated plainly — this file cannot distinguish product completeness from exclusion, and none of the clinical AI you actually care about needed ONC certification to begin with.

D3 data explorer

Zero

Source: Casey Ross & Bob Herman, “FTC reviewing Epic Systems’ use of NDAs, possible anticompetitive practices”, STAT News, 14 Aug 2026 · related: Commure referral-payment terminations, STAT, 13 Aug 2026 · MIMI Labs: CMS/ONC Promoting Interoperability Hospital Public Information file, program year 2023

One dot field holding every hospital in the file. 4,046 squares fade in, split into the 1,836 on Epic and the 2,210 on everything else, and then the zero column ignites: 1,487 Epic hospitals — 81% — named no third-party certified developer at all, against 552 of 2,210, or 25%, everywhere else. Mean of 0.23 versus 1.58. The obvious objection is that Epic hospitals are big academic systems that build in-house, so the graphic tests it: split by CMS hospital type and the gap is 0.23 vs 1.68 in acute care and 0.21 vs 1.39 in critical access — size does not explain it. A mirrored ranking of the fourteen most-reported non-base developers shows who gets named alongside Epic and who doesn’t: Medisolv 3.1% vs 21.7%, Inpriva 0.1% vs 13.5%, Commure 0.0% vs 6.7% — and Clinisys, the counterexample, at 4.7% vs 1.0%. Then the lens that undoes most of it: the certified-product list is a compliance surface Epic ships modules for, so “zero” is what product completeness looks like from the outside and is indistinguishable, in this file, from exclusion.

Graphical narrative
August 14, 2026

The NTAP Ledger

Source: Federal Register, FY2027 Hospital Inpatient PPS final rule (CMS-1849-F), 4 Aug 2026 · uptake study: Pelzl et al., AJNR, 24 Jun 2026 (Neiman HPI release) · also STAT, 13 Aug 2026 · MIMI Labs: CMS Medicare Inpatient Hospitals by Provider & Service, CY2023 vintage

Medicare’s FY2027 rule approved 19 new technology add-on payments, five of them AI or algorithm-based — up from one the year before. This drops those five rates onto 2,259 real IPPS hospitals and the 687,506 Medicare sepsis and 148,049 stroke discharges they actually billed in CY2023, positioned by eligible volume against what the add-on is worth as a share of the payment the case already generates. Pick Bayesian Health’s sepsis flagging at $61.84 and the median hospital’s add-on is 0.47% of the DRG payment; the whole national program at the observed 14.8% uptake is about $6.2M a year. The built-in lens is the correlation nobody states out loud: the add-on correlates with volume at r = 1.00 by construction, because it is a fixed dollar amount times a count of uses — there is no term for sensitivity, outcome, or whether a decision changed. Then switch Who adopts from uniform to the AJNR-observed pattern (comprehensive stroke center OR 1.5, Stroke Belt OR 2.0, re-calibrated to your slider) and the non-metro share of the money falls from 11.3% to 8.9%. A thinness panel names what the file cannot see: Medicare FFS only, no Medicare Advantage, no critical access hospitals, CMS small-cell suppression, and a case base that is a workload stand-in rather than a device eligibility rule.

D3 data explorer

65% of Whatever You Say It Costs

Source: Federal Register, FY2027 Hospital Inpatient PPS final rule (CMS-1849-F), 4 Aug 2026 · uptake study: Pelzl et al., AJNR, 24 Jun 2026 · STAT, 13 Aug 2026 · MIMI Labs: CMS Medicare Inpatient Hospitals by Provider & Service, CY2023 vintage

One animated, log-scaled price ladder holding the entire FY2027 AI price list: $61.84 for Bayesian Health’s sepsis flagging, $137.53 for Aidoc’s multi-triage CT, $975 for Nelli seizure monitoring, $2,171 for the Ceribell delirium monitor, $2,275 for InVision’s cardiac amyloid detection. Behind each red bar is a dashed outline — the cost the applicant claimed — and every bar ends at exactly 65% of it, because that is the rule and not a coincidence. A dashed navy line marks the $14,796 discharge-weighted average Medicare inpatient payment, so you can see the sliver. Press Show the 14× move and InVision’s figure walks from the $162.50 proposed in April to the $2,275 finalized in August — nothing about the technology changed, only the cost accounting. The coda is the only published look at who bills these: a dot field of the 2,116 Medicare stroke episodes across 1,076 facilities in the AJNR study, 313 of them lit — 14.8% — and a forest plot of what actually predicted use. Comprehensive stroke center OR 1.5, Stroke Belt OR 2.0, treated in 2022 OR 6.0. Stroke severity, age, sex, race: nothing.

Graphical narrative
August 12, 2026

The Verification Cohort

Source: Imaging Technology News, “Healthcare Systems, Aidoc Create Diagnostic AI Consortium”, 10 Aug 2026 · corroborated by AuntMinnie, 11 Aug 2026 · MIMI Labs: AHRQ Compendium of U.S. Health Systems, hospital linkage file, 2023-12-31 vintage

Twelve health systems just formed a buyers’ union for AI evidence, and the deliverable that matters is the third one — shared governance practices “any health system can adopt.” So this plots the cohort against the field: every one of the 636 health systems in AHRQ’s Compendium, positioned by total staffed beds on a √ scale against the share of that capacity sitting in hospitals under 100 beds, sized by inpatient discharges, with the twelve in red. They hold 42,646 of 776,811 staffed beds — 5.5% — and 2,120,467 of 32,173,148 discharges. The built-in stress test is an archetype filter over the hospital-level cross-tab: leave it open and coverage sits at 5.5%; select major-teaching and it rises to 8.2%; select high-uncompensated-care and it falls to 2.7%; select both — America’s safety-net teaching hospitals — and it is 326 beds of 21,418, 1.5%. Sixty-five such hospitals nationally; one belongs to a member. Then the small-n artifact, live: with all 636 systems there is a clean-looking negative correlation between system size and small-hospital share, but roughly two hundred of those dots are a single sub-100-bed hospital pinned at 100% by definition. Drag the minimum-beds slider past 300 and watch r collapse. The trend was the definition, not a finding.

D3 data explorer

Thirty-Two Blank States

Source: Imaging Technology News, “Healthcare Systems, Aidoc Create Diagnostic AI Consortium”, 10 Aug 2026 · corroborated by AuntMinnie, 11 Aug 2026 · MIMI Labs: AHRQ Compendium of U.S. Health Systems, 2023-12-31 vintage

An animated tile map of what the new consortium can and cannot see. Fifty states plus DC, equal-sized squares, shaded by how much of each state’s staffed hospital capacity sits inside a member system. The reveal runs in three beats: all 6,800 linked hospitals and 776,811 beds appear, then the 216 member hospitals across 19 states ignite, then the rest go dim — thirty-two states and the District of Columbia contain no member hospital at all. Connecticut is the most covered state in the country at 22.8% of staffed beds; New York is 17.9%, Illinois 19.2%, North Carolina 19.9%; Texas, where Houston Methodist sits, is 3.9%, and Pennsylvania is 3.7%. A second view flips the shading to where the beds actually are — California 73,448, Texas 71,316, Florida 58,276 — which is where the blanks start to hurt. Below the map, the unit reconciliation nobody does out loud: “nearly 20 million patients annually” is the consortium’s own figure and is not a discharge count. In AHRQ’s file the twelve are 2.12 million inpatient discharges out of 32.17 million, 6.6%, and 1.5% of the beds in safety-net teaching hospitals.

Graphical narrative
August 11, 2026

The Radiology Panel, Plotted

Source: FDA, “Radiology Devices; Reclassification of Digital Breast Tomosynthesis System”, proposed amendment / proposed order, 91 FR 51406–51416, published 10 Aug 2026 · comment docket FDA-2026-N-7630, closes 9 Oct 2026 · MIMI Labs: FDA 510(k) premarket notification database + FDA device classification (foiclass) file, vintage through decision date 24 Jul 2026

A device class is a description of what evidence you have to bring. On August 10 FDA proposed moving digital breast tomosynthesis across that line, so this plots the whole neighbourhood it is moving into: every one of the 212 product codes under FDA’s radiology review panel, positioned by median 510(k) review time against total clearances — 8,528 of them — on a √ scale, sized by clearances since 2016, coloured by class. Sixty-one codes have never been cleared through 510(k) at all and sit in a shaded rail on the left. One of them is OTE. Sixteen of the eighteen Class III radiology codes are in that rail; the two exceptions, KXZ and IYM, were cleared between 1976 and 1986. Then the built-in stress test. Pool every radiology 510(k) ever decided and the sixteen AI/image-analysis software codes take a 143-day median against 79 for everything else — an 81% penalty, and exactly the number that ends up on a conference slide. But fifteen of those sixteen codes were created in 2018 or later, so that compares a 2018–2026 sample against a baseline running back to 1976. Restrict the comparison group to the same era and the baseline rises to 106 days: the gap halves, +64 to +37, and year-by-year it runs +12 in 2024. Most of “FDA is slower on AI” was “FDA got slower, and AI arrived late.” Filters for device class, minimum clearances and the AI subset, plus search. Closes on four AI product codes FDA created that nobody has ever used — including QVD, machine-learning imaging software with a predetermined change control plan, zero clearances, the same PCCP mechanism that appears in today’s order’s preamble and nowhere in its codified special controls.

D3 data explorer

Zero

Source: FDA, “Radiology Devices; Reclassification of Digital Breast Tomosynthesis System”, proposed order, 91 FR 51406–51416, 10 Aug 2026 · MQSA national statistics, as of 8 Jul 2026 · MIMI Labs: FDA 510(k) database, through 24 Jul 2026

The whole mammography lineage as a single animated dot field — one dot per FDA 510(k) clearance, four lanes, 1983 to now, with a play button and a year scrubber. IZH, the mammographic X-ray system code, fills steadily through the nineties and dies out. MUE, full-field digital mammography, stays empty until 2011 and then opens — because FDA reclassified FFDM into Class II in 2010. Forty clearances since, at a 169-day median: a door, not a floodgate, and the precedent FDA cites by name in today’s order. The third lane never fills. OTE, digital breast tomosynthesis: zero 510(k) clearances, ever, because Class III meant four original PMAs and twenty-six supplements instead, starting with Hologic’s Selenia Dimensions 3D on 11 Feb 2011. The fourth lane is the sixteen radiology AI/CAD software codes — four clearances in 2018, 354 by July 2026 — every one of them reading pixels off the platform in the empty lane. Against that: 94% of the 9,107 MQSA-certified facilities have DBT units, five Class II recalls with no injuries attributed to any, 968 MDRs with about 1% reporting serious injury, 2.6 million patients across 42 studies. That is the safety case, and it is a good one. The sentence that matters is at 91 FR 51412, where phantom studies “may serve as an alternative to clinical studies,” and the codified special control lists human subjects, structured physical phantoms, or in silico methodologies. The order contains zero instances of “artificial intelligence.” Its entire software special control is nine words.

Graphical narrative
August 10, 2026

Four Percent

Today’s Big Thing turns on one statistic: 4% of Epic-based executives think a startup can win interoperability. This explorer asks where that 4% comes from, then whether it’s right anyway. The survey asked the same 112 people two questions about the same fifteen areas — a 1–5 rating (can an outsider compete here?) and a forced pick-four. Paired bar charts put both side by side with an axis toggle: zoomed, the rating chart looks like a ranking; on the true 1–5 scale it collapses into a wall, all fifteen areas inside a 0.36-point band, every one above the midpoint. Rank agreement between the two questions is only ρ = 0.19. Then the federal record, which is harder on builders than the survey is: 4,593 hospitals attesting to CMS Promoting Interoperability, one dot per state, Epic penetration against third-party certified vendors per hospital. Drag the minimum-hospital slider and — unlike the last two interactives — the correlation gets stronger, from −0.64 to −0.77. It survives its own stress test. A third graphic names the mechanism vendor by vendor: Inpriva sits at 299 non-Epic hospitals and one Epic hospital, Secure Exchange Solutions 260 versus 3, Commure 149 versus 0. A non-Epic hospital lists 1.83 third-party certified vendors; an Epic hospital lists 0.23. Ends on what the file can’t separate: displacement versus a base product that already covers every measure.

Data explorer

Four Picks, Seventeen Doors

The same finding as one animated field. 112 dots, one per executive, and a play button that walks down the ranking lighting the dots who spent one of their four picks on each area. Clinician-facing AI takes 54. Imaging 49. It ends on interoperability with four dots lit. Then the arithmetic that the headline number hides: 112 people × 4 slots = 448 picks across 17 doors, so random allocation gives every area about 26 — the line to measure against is 23.5%, not zero, and the published shares only account for 370 of the 448 because the report never breaks out four of the seventeen areas. Finally the two questions on their true scales, animated: the ratings slide out into a 0.36-point band from 3.16 to 3.52 while the picks spread across 44 points. Interoperability was rated 3.21 — above bed and OR throughput, which drew six times as many picks. The honest restatement is in a navy block: not 96% of buyers decided a startup can’t win it, but it came seventeenth in a contest for four slots of attention while being rated essentially tied with everything else — which may be the worse news, because an area nobody argues about is harder to raise money against than one they do.

Graphical narrative
August 9, 2026

The Level Four Line

Today’s essay argues that clinical judgment compresses into something faster than language, and that the real bottleneck is writing the spec. So this takes the best case: “was that office visit moderate complexity?” — the one clinical judgment America has tried hardest to specify, with a written definition, a 2021 rewrite, a decision table, audits, and money attached. 1,440 clinicians, one dot each, plotted by established-visit volume against the share they coded at level 4 or 5 (99214/99215) rather than level 2 or 3. Filter by any of sixteen specialties, switch the vertical axis, and drag the minimum-volume slider. The critical lens is the interaction: CMS suppresses every provider–code row covering ten or fewer beneficiaries, so a cardiologist with 400 level 4s and eight level 3s appears at 100% — among clinicians with 11–49 visits, 79.9% have only one of the four codes present and 85% sit at exactly 0% or 100%. Push the slider and that bimodality dissolves on cue: the rails fall to 10%, the single-code column falls to 4%. And the spread underneath does not move. Among the 11,558 clinicians billing more than 1,500 established visits a year — too busy to be a rounding artifact, too visible to skip an audit — the middle half still runs 27% to 90%. A second dense graphic ranks all 31 specialties with 1,000+ clinicians by their 10th–90th percentile band: podiatry’s median is 6.6%, endocrinology’s is 95.1%, and the shortest bar on the board is still thirty points wide. Ends where the essay does: the spec is the deliverable, and it is the part nobody can write for you.

Data explorer

Show Me the Spec

Source: Barbara Hays & Cindy Hughes, “Coding Level 4 Office Visits Using the New E/M Guidelines”, AAFP Family Practice Management, Jan/Feb 2021 · MIMI Labs: CMS Medicare Physician & Other Practitioners PUF, by Provider and Service, 2024

The same finding as a single animated graphic. 1,440 clinicians start stacked on one point — where a working specification would put them — and one button releases them into their actual 2024 numbers. The field opens to a 67-point interquartile range with a visible pile-up against both walls, and the readouts count up as the dots fly. Below it, the same clinicians split into volume bands with a step-through button: at 11–49 visits 89% are pinned to a wall, which is CMS’s ten-beneficiary suppression rule rather than clinical practice; by 500+ visits the walls are down to 14% and the middle-half spread is unchanged. National figures on all 540,972 clinicians sit underneath in a table. The piece is deliberately careful about what this is not: an older panel, a referral-heavy practice, longer slots and a good scribe all legitimately move this number, and claims data cannot separate judgment from documentation habit. But that concession is the argument — nobody can write down how much a scribe should shift your coding either. Those are reflexes too, which is exactly why the “learn to code” advice was always the wrong assignment.

Graphical narrative
August 8, 2026

Nine Point Nine Billion Dollars of Bandage

Medicare Part B paid $9.86 billion for skin substitutes in 2024 — more than it paid for ambulances or anesthesia — up from $23.9 million across 16 codes in 2013. Here are the 600 highest-billing clinicians in the country, one dot each, plotted by square centimeters of graft billed against Medicare allowed dollars per square centimeter, with the CY2026 flat rate of $127.28/sq cm drawn as a line near the floor. Filter by specialty, state, or minimum volume; switch the axes to dollars per patient or total allowed; then press “2026 flat rate” and watch the whole field drop onto the line — the same 2.9 million square centimeters, the same wounds, repriced from $4.39 billion to $369 million, a 91.6% cut with nothing clinical changed. Nurse practitioners are the single largest billing block at $1.64 billion across 372 clinicians. Critical lens built in: CMS suppresses every clinician-product row covering ten or fewer beneficiaries, so only $4.62B of the $9.86B (47%) is visible at individual level and the invisible half is by construction the low-volume half; the patient counts are summed across product codes and double-count anyone treated with more than one graft, which the page says outright rather than quietly using; and the minimum-volume slider is there to be pushed — watch the live log–log correlation readout wander once you are fitting forty points instead of six hundred. Ends on the price comparison that explains the whole thing: Apligraf, PMA-approved since 1998, has been paid $29–$39/sq cm for twelve straight years; Amchoplast was paid $4,160/sq cm in 2024. Same procedure, 138× the price, and the difference is which payment methodology the product qualified for.

Data explorer

The Price Ladder

Sixty-nine Q4‑series skin substitutes, every one a sheet of processed tissue laid on a chronic wound, every one billed to Medicare in 2024 — on a single dot field, price per square centimeter against total spend, sized by patients treated. The spread runs from $30 to $4,160 per square centimeter. One button reprices the entire field at the CY2026 flat rate and $9.84 billion becomes $905 million, a 90.8% cut that lands within a point of CMS's own estimate. Underneath, two more graphics: twelve years of Apligraf's allowed amount sitting flat between $29 and $39 while Q4205 — the largest single code in the country at $1.44 billion — jumps from $349 to $1,424 in one year, and the national spending bars from $23.9M across 16 codes in 2013 to $9,858M across 94 in 2024 with beneficiaries rising only tenfold over the same span. Critical lens: the repricing is arithmetic, not a forecast — it holds 2024 volume constant, which is exactly the thing the rule is designed to change; 25 codes below $5M are omitted (0.2% of spend); rows with ten or fewer beneficiaries are suppressed by CMS entirely; and the “patients” counts double-count anyone who received more than one product. The takeaway generalizes past wound care: any payment methodology that derives a price from the seller's own reported price gets gamed, and the tell is always spending growing an order of magnitude faster than patients.

Graphical narrative
August 7, 2026

The Ruler’s Own Ruler

Source: Aali et al., “MedVAL: Toward Expert-Level Medical Text Validation with Language Models” · npj Digital Medicine, Aug 4, 2026 (DOI 10.1038/s41746-026-03084-5) · all 90 model×task cells from Tables 2, 3a, 3b, S2, S3 · MIMI Labs: Dartmouth Atlas MEDPAR ICD‑10 principal diagnoses, 2018 · code StanfordMIMI/MedVAL, benchmark MedVAL-Bench

A Stanford group open-sourced the thing everyone said was missing: an evaluator that checks whether AI-generated clinical text is factually consistent with its input, trained without a single physician label, with a 4B model you can run on a laptop. Ten language models were scored against MedVAL‑Bench and five were also distilled — 90 model‑by‑task cells, every one plotted here, against your choice of average input length, inter-physician agreement, or task size. Set the axis to agreement and show MedVAL GPT‑4o and you reproduce the paper’s r = 0.67: the evaluator scores highest exactly where the physicians agreed with each other most. Then click query2question to drop it and the correlation doesn’t weaken — it inverts to r = −0.55. One task out of six, the one where twelve physicians reached only α = 0.560 among themselves, was carrying the whole relationship; push the minimum-task-size slider to n ≥ 135 and you are fitting three points at r ≈ 1.0, which means nothing at all. Critical lens throughout: the abstract’s headline F1 of 66% → 83% is the binary safe/unsafe judgment, while the four-class risk grade a reviewer would actually act on moves 36.7% → 51.0%; the non-inferiority claim rests on 90 of 840 examples, tests GPT‑4o rather than the released 4B, and never prints the observed Δ, the confidence bound, or the human expert’s own score. Scale check from MIMI Labs: 11,007 distinct ICD‑10 principal diagnosis codes appeared across 12.2M Medicare inpatient stays in 2018 — 840 examples would touch 7.6% of them.

Data explorer

840 Dots

Source: Aali et al., “MedVAL: Toward Expert-Level Medical Text Validation with Language Models” · npj Digital Medicine, Aug 4, 2026 (DOI 10.1038/s41746-026-03084-5) · counts from Tables 2 and S3 · benchmark MedVAL-Bench, weights MedVAL-4B

Everything the new evaluator knows about being wrong comes from 840 physician-annotated outputs. Here they are — all of them, one dot each, six task columns, built by a single animated field with no click-through. Colour them by the four-level physician risk grade (44.4% of the benchmark is graded level 3 or 4, meaning a human must review or rewrite); by which three tasks were held out of distillation entirely (395 of 840); or by input length. Then light up the 90 dots — fifteen per task — that were read by more than one physician. That 10.7% subset is the entire basis for the inter-physician agreement figures and for the headline that the evaluator is non-inferior to a human expert; the paper says so in its own limitations section. Below the field, the six tasks’ average input lengths drawn to scale: 10, 50, 69, 82, 543 and 1,497 tokens — the medication-answer task, where distillation produces its largest gains, hands the evaluator a ten-token question, and its bar is 0.7% the width of the ambient-scribe task’s. A real ED encounter is not a paragraph; it is a chart. Critical lens: on query2question the physicians reached only α = 0.560 with each other, and the paper reports r = 0.67 between how well physicians agreed on a task and how well the tuned model scored on it — the evaluator looks most expert where experts find it easiest to agree. What didn’t get distilled is the taxonomy: four risk levels, eleven named error types, twelve physicians, and a decision about what each kind of wrongness costs.

Graphical narrative
August 6, 2026

The Underwater Line

Buried in the FY 2027 inpatient rule CMS published on August 4: the repeal of the alternative pathway that let an FDA Breakthrough Device designation stand in for proof of substantial clinical improvement when applying for a New Technology Add-on Payment. That reads like paperwork; it is a pricing constraint. Medicare pays one fixed amount per inpatient stay, set by the stay's MS‑DRG, regardless of what the hospital spent — so every dot here is one of the 773 national MS‑DRGs, average Medicare payment against annual discharges, and the slider is the cost of the new thing you want to put inside that stay. At $25,000 with no add-on payment, the technology costs more than the entire average Medicare payment in 620 of 773 DRGs — 6.09 million discharges, 88% of all Medicare fee-for-service inpatient volume. Turn NTAP on at 65% and the underwater count falls to 220 DRGs and 2.22 million discharges. That gap, 620 versus 220, is the add-on payment. The rule didn't remove NTAP; it removed the shortcut to qualifying for it, for applications submitted on or after October 1, 2026. Stress test: flip the axis toggle to Charges and the red mostly disappears — at a $50,000 technology only 196 DRGs sit below the line by charges against 729 by payments, because the median DRG's submitted charge is 6.0× its Medicare payment, which is why a market model built off chargemaster data invents roughly four times the headroom that exists. And drag min annual discharges up: 729 underwater becomes 482 at 1,000/yr and 243 at 5,000/yr, because 263 of the 773 DRGs see fewer than 1,000 discharges nationally and together carry 1.6% of volume. Critical lens: NTAP is calculated off the hospital's cost for the case, and this file contains payments and charges and no cost column at all — so the red-dot test is a ceiling test, not a margin test.

Data explorer

773 Prices

On July 8, FDA and CMS officials hosted an unannounced “clinical AI demo day” at White Oak; STAT reviewed the agenda and published it August 5. Ten companies, no health system, no specialty society — and the agenda covered not only whether the tools are safe but how Medicare should pay for them. This is that second question, drawn. One animated field, no click-through: it plays itself and you can scrub back. All 773 national MS‑DRGs arrive as a scattered cloud, settle into a dot histogram sorted by what Medicare actually pays, the median line lands at $12,714, and then a technology-cost line sweeps in from $300,000 down to $25,000 while the dots it can no longer fit inside turn red and a live counter runs. At $25,000 with no add-on payment that is 620 of 773 prices and 88% of all Medicare fee-for-service inpatient volume; with NTAP at 65% it is 220 and 32%. Then the last act: an empty box where autonomous clinical AI would sit — no DRG, no code, no rate, 0 of 773 rungs — because when no clinician performs the service, no existing payment category contains it, and whoever defines “a unit of service” sets the business model for the whole category. Sliders hand control back after the story ends. Critical lens: these are payments, not costs, and NTAP is computed off costs, so the red is a ceiling test not a margin test; fee-for-service only, so Medicare Advantage — roughly half of enrollees — is absent; and 263 of the 773 DRGs see under 1,000 discharges a year and carry 1.6% of volume, so a dot is a row, not a population.

Graphical narrative
August 5, 2026

Clinics Like Mine

Magan checked forty-one references against their primary sources and found every landmark ambient-scribe study ran at UCLA, Mass General Brigham, Emory, UCSF, Yale, UC Davis, Kaiser Northern California, Penn or Stanford — and none in a federally qualified health center. So here is the population that isn't in the evidence, all of it on one scatter: every HRSA grantee in the country, panel size (log) against the share of patients HRSA records as best served in a language other than English. 8,977,620 patients — 27.8% of the national panel — sit on that axis, and 190 grantees are above 50%. Filter by state, urban/rural, and language share; the readout recomputes the weighted share, the correlation, and the study count (0). Stress test: the scatter tilts at r = 0.205, which reads as “scale correlates with linguistic diversity” — drag min panel size to 20,000 and r collapses to 0.115, because the tilt was the long tail of 552 rural grantees at a 3.4% median. Critical lens: “best served in another language” is a self-reported UDS field, not an audio measurement — it says nothing about accented English, which is exactly where a scribe pipeline degrades; the grain is the grantee, not the clinic; and the zero is a hand-assembled absence (“none found,” not “none exist”).

Data explorer

Nine and 1,352

One animated field, no click-through — it plays itself and you can scrub back. 1,352 health center grantees arrive as an undifferentiated cloud, then sort themselves into a dot histogram along the share of patients best served in a language other than English, then the 499 above 25% light up (they carry 16.4 million patients, half the national panel from a third of the dots), then the nine academic settings where the ambient-scribe evidence was actually generated slide in underneath as nine navy squares with nothing above them. Alongside: the 2020 PNAS audit of five commercial speech-recognition systems — word error rate 0.35 for Black speakers against 0.19 for white speakers, a gap nobody has re-tested in any of the 1,352. Critical lens: one dot is one grantee, not one clinic and not one patient, so the tall left stack over-represents small rural organizations; the language field is self-reported and blind to accent; and the nine is an absence assembled by hand from 41 checked references.

Graphical narrative
August 4, 2026

The Accountability Object

Source: STAT First Opinion — Chaudhry & Valentine Theard, FSMB, “AI is not ready to be licensed like a physician” (Aug 3, 2026) · MIMI Labs: NPDB Public Use Data File (updated May 2026) · 193,023 paid malpractice claims, incident years 2004–2021

The Federation of State Medical Boards argued this week that a license isn't a certificate of correctness — it's a name to attach when something goes wrong. So look at the federal database of those names: 1,911,185 disclosable reports, 985,019 practitioners, zero models. Every paid malpractice claim cut two ways at once — what was alleged (11 groups) × what happened to the patient (10 severity codes) — as 110 dots on a log-log scatter of claim count against mean payment. The four groups a decision-support tool actually sits inside (diagnosis, treatment, medication, monitoring) carry 63.8% of paid claims and 60.5% of the dollars. Stress test: drag the minimum-claims filter to 200 and the expensive-looking top-right corner empties — IV & blood products → brain damage reads $875k a claim on 17 claims, sitting beside signals built on thousands. Critical lens: payments are coded range midpoints not dollars, incident years 2020–21 are hollowed by reporting lag, work-state is null in 57% of records, nine states double-file state-fund payments — and there is no field anywhere in the schema for “a model contributed to this decision.”

Data explorer

A Name to Attach

Source: STAT First Opinion — FSMB on licensing AI (Aug 3, 2026) · MIMI Labs: NPDB Public Use Data File · 1,930 dots at 100 paid claims each

193,023 paid malpractice claims arrive as an undifferentiated haze, then settle into what was alleged, then re-settle into what it cost the patient. Three beats, one animated field, no click-through required — it plays itself and you can scrub back. The four allegation columns a clinical AI tool would land inside light up red; death turns out to be the single largest severity code in the file. Alongside: 1,911,185 NPDB reports, 985,019 practitioners named, 0 models named. Critical lens: at 1 dot = 100 claims the rare categories vanish (IV & blood products is four dots across eighteen years), the dollars are coded band midpoints, and paid claims record legal outcomes filtered through two decades of tort reform — not harm.

Graphical narrative
August 2, 2026

The Calm Light

Source: CMS — Hospital Readmissions Reduction Program · MIMI Labs: HRRP public file, latest vintage, Provider Data Catalog · READM-30-HF-HRRP, discharges Jul 1 2021 – Jun 30 2024

CMS publishes two numbers for every hospital and only one of them is arithmetic. Divide readmissions by discharges and a Tennessee hospital with 31 heart-failure discharges reads 45.2%; CMS's own published rate for the same hospital, same period, is 21.5%. All 2,316 non-suppressed hospitals on one funnel plot — toggle between the number you can compute and the number that carries the money, and watch the spread fall from SD 5.55 to 2.44 without a single hospital changing. Stress test: drag the minimum-discharge slider and the “worst in America” leaderboard turns over completely — of the 50 highest raw rates, median volume is 65 discharges against 274 for the file as a whole. Critical lens: 36% of HRRP rows are suppressed outright, the denominator excludes Medicare Advantage (now >half of Medicare), and CMS conceded its own risk model was scoring poverty when it added dual-eligible peer groups in FY2019.

Data explorer

Twenty-Three Points

Source: CMS — Hospital Readmissions Reduction Program · MIMI Labs: HRRP public file, all six conditions, 8,037 hospital-condition rows

2,316 dots drop in at the readmission rate anyone can compute from the file, then red segments draw the distance to the rate CMS actually publishes, then they land on it and the left-hand chaos disappears. The median correction is 3.73 points under 75 discharges and 0.45 points at 600 or more; the largest single move is 23.68 points. Then all six HRRP conditions as small multiples — and the honest finding that the funnel only behaves in four of them: CABG and hip/knee have 363 and 253 hospitals and their percentile bands wobble instead of narrowing, so those panels are marked in red with their bucket sizes shown. Critical lens: shrinkage is a choice, and it guarantees a genuinely bad small hospital reads as average — neither published number tells you which one you're looking at.

Graphical narrative
August 1, 2026

The 3.5 Line

All 610 Medicare Advantage contracts with 1,000+ members on one brushable scatter — 2026 star rating against enrollment, sized by members, 35.7M people. Humana said the 2027 exits are "mostly plans rated 3.5 stars or lower." Below that line sit 11,165,479 members nationally, and 4,233,892 of them are Humana's — 59.6% of its own book and 37.9% of every low-star MA member in the country. The announced 600,000 is 14.2% of it; Humana's single largest contract, H5216, is four times the whole exit and stays. Stress test: star rating vs. log enrollment gives r = +0.35 across 486 rated contracts and decays to +0.22 once you drop everything under 50,000 members. Critical lens: 378 contracts (42.9%) carry no overall rating at all, and we could not reproduce Humana's own "20% of members in 4-star plans" from the public files — we get 38.6%.

Data explorer

Fourteen Percent

Source: Becker's — Humana CFO Celeste Mellet, July 29 Q2 earnings call, published Jul 30, 2026 · MIMI Labs: CMS CPSC enrollment June 2026 × 2026 Medicare Part C & D Star Ratings

1,788 dots, one per 20,000 Medicare Advantage members, settling into their star bands — then Humana's 7.1M turn red, then the 600,000 exit gets bracketed as 30 of them, then 40% fade back because that's the recapture rate Humana expects. Humana's 3.5-star band alone holds 3,590,176 people, 40.7% of everyone in the country at 3.5 stars; it has zero members at 5.0 and zero at 2.0, while Kaiser has zero below 4.0. Then the part national averages hide: 57% of every MA member in Montana is in a Humana contract rated 3.5 or lower, versus 3.9% in California. Critical lens: the star rating is a lagged contract-level bonus-payment composite, not a quality reading of your attributed lives.

Graphical narrative
July 31, 2026

The Ex Parte Line

Source: Tradeoffs — "Meet the Man Launching Trump's Medicaid Work Requirements Months Early" (Jul 30, 2026) · CMS-2454-IFC, effective Jul 31 · MIMI Labs: CMS Medicaid & CHIP Eligibility Processing, state-month renewal outcomes, Mar 2023–Apr 2025

All 51 reporting jurisdictions on one brushable scatter — ex parte renewal rate against procedural coverage loss, sized by volume, 69.1M renewals. Texas clears 11.8% of renewals without asking anyone for a document; Washington clears 80.6%. Nationally 8.49M people lost Medicaid on paperwork versus 4.69M found actually ineligible — 1.81 paperwork losses per ineligibility finding. Stress test: filter to states under 300k renewals and 2024–25 gives r = −0.84, a gorgeous finding that collapses to r = −0.08 when you toggle to the unwinding. All 51 states hold steady at about −0.49 in both periods. Critical lens: the dataset has 22 columns and not one of them is the medically-frail exemption code list that decides who is too sick to work.

Data explorer

The 26%

Source: Tradeoffs — Nebraska Medicaid director Drew Gonshorowski, published Jul 30, 2026 · MIMI Labs: CMS Medicaid & CHIP Eligibility Processing, renewal outcomes May 2024–Apr 2025

1,000 dots per state, drawn to the real CMS renewal ledger, animating into the split that decides everything: 55 cleared from data the state already held, 45 who had to act. Of the 31.0M renewals the automation could not clear, 27.4% ended in a procedural termination; of the 38.0M it cleared, none did, by definition. Then the trap — Washington has the country's highest ex parte rate (80.6%) and the third-highest loss rate inside its un-cleared cohort (54.7%), while Pennsylvania is second-worst on automation and best on that measure. Automate harder and what's left is the hard cases. Rank states on the bar chart alone and you rank them backwards.

Graphical narrative
July 29, 2026

Cleared Before Measured

Source: Chalouhi et al., npj Digital Medicine — multireader fetal ultrasound study (Jul 28, 2026) · MIMI Labs: FDA 510(k) release, imaging-AI product codes, decisions through Jul 17, 2026

Every FDA 510(k) clearance in the imaging-AI product codes — 369 dots, 184 companies, 2016 to July 2026 — plotted against review time, colored by what the device is allowed to claim. Only 20% sit in a detection/diagnosis code; more than half just segment and measure. Stress test: raise "min clearances per company" to 5 and the median review drops 140 → 123 days; filter to each company's first submission and it climbs to 172. Critical lens: nothing in the 510(k) record says whether anyone ever ran the unassisted arm — Sonio Suspect cleared in 91 days and published its 522 days later.

Data explorer

The Unassisted Arm

Source: Chalouhi et al., npj Digital Medicine, published Jul 28, 2026 · device: Sonio Suspect, FDA K243614 · study funded by Sonio, a Samsung company

All 750 fetal ultrasound stills from yesterday's multireader study on one field — 250 abnormal, 500 normal — with the AI switching on and off. Unassisted, 13 U.S. physicians caught 136 of 250 abnormalities and agreed with each other 26% of the time; with the assistant, 221 and 72%. Then the stress test the study didn't run: drag prevalence from the enriched 33.3% down to a real 3% anomaly screen and watch PPV fall from 78.5% to 18.4% — roughly four false alarms for every real finding, even in the assisted arm.

Graphical narrative
July 28, 2026

The Blast Radius Ledger

Source: Becker's — health IT vendor breach exposes 425,000+ patients (Jul 2026) · AnMed systems disruption (Jul 26–27, 2026) · MIMI Labs: HHS OCR breach portal, filings of ≥50,000 people, Oct 2009–May 2025

Every large breach ever reported to HHS — 939 filings, 586 million records — on one brushable scatter, vendor-linked breaches in red. Raise the floor from 50k to 5M and watch the everyday provider breaches vanish while the vendor detonations remain: red's share of affected people grew 18% → 63%. Critical lens: the vendor is never named in the file, and the biggest breach in history (Change Healthcare, ~190M) is missing from the snapshot entirely.

Data explorer

One Vendor, Many Letters

Source: Becker's — health IT vendor breach (Jul 2026) · AnMed update (Jul 27, 2026) · MIMI Labs: HHS OCR breach portal, ≥50k filings, 2009–2025

Sixteen years of breach filings replayed as a monthly dot field: a hospital breach makes one dot, a vendor breach makes a tower. Watch three detonations — AMCA (14 letters, one month), Blackbaud (40 letters, one month), MOVEit (a seven-month tail) — surface under other organizations' names, with this week's AnMed and Unlimited Technology Systems dots drawn where they'll land. The vendor's name appears in the ledger zero times.

Graphical narrative
July 27, 2026

The Predicate Lane Economy

Source: Federal Register, 91 FR 46719 — FDA final order codifying the diabetes digital behavioral therapeutic device (Jul 24, 2026) · MIMI Labs: FDA device classification (foiclass) + 510(k) databases, decisions 2016–Jul 2026

FDA maintains 6,987 product-code lanes; 78% carried zero 510(k)s in a decade. All 846 lanes with real traffic on one brushable scatter — traffic vs. median FDA review days, one dot per lane, the 13 digital-therapeutic lanes in red with QXC (Friday's new diabetes lane) on the zero shelf. Critical lens: drag the sample-size floor and watch the "fastest lane in FDA" dissolve into small-n noise.

Data explorer

Thirteen Lanes, Fifteen Cars

Source: Federal Register, 91 FR 46719 (Jul 24, 2026) · MIMI Labs: FDA 510(k) database, all "Substantially Equivalent" decisions through Jul 2, 2026

Every digital-behavioral-therapeutic lane FDA has opened, drawn as a lane — and every 510(k) that ever drove through one as a dot. Thirteen lanes, fifteen clearances ever, five lanes empty, including the diabetes lane codified Friday. For scale: one radiology software lane (LLZ) carries 442 — 29× all thirteen combined. A lane is not a market.

Graphical narrative
July 25, 2026

Billing the Algorithm

Source: AAPC — "AMA Posts CPT Early Release Codes" (effective July 1, 2026) · AMA — CPT codes for AI-enabled services · MIMI Labs: Medicare Physician & Other Practitioners, national rows, CY2018–CY2024

The mid-year CPT release is where AI codes land first — so which ones does Medicare actually pay? 27 AI-adjacent codes checked against seven years of Part B physician claims, one dot per code, animated by year: the whole 2024 "billable AI" market is $10.0M, 12 of 25 AI codes never clear the 11-patient floor (including 0691T, today's Money Plumbing example), and the algorithmic-ECG code earned $3.53 a run — once. Critical lens: a small-n filter that dissolves most of the "growth," plus the physician-claims blind spot stated out loud.

Data explorer

The Code Exists. The Check Doesn't.

Source: AAPC — "AMA Posts CPT Early Release Codes" (effective July 1, 2026) · MIMI Labs: Medicare Physician & Other Practitioners, national rows, CY2013–CY2024

Twelve years of AI billing codes as one animated chart: the six-year Category III valley (FFR-CT never topped 16k family-wide services), the 2024 conversion spike to 29,270 services and $7.3M as 75580, autonomous retinal AI's price falling $32→$27 while the human-read version still outbills it — and a ghost shelf of 12 codes that never registered eleven national patients. Conversion, not code creation, is the revenue event.

Graphical narrative
July 24, 2026

The Quarter-Billion Reshuffle

Source: CMS — CY2027 Physician Fee Schedule proposed rule (comment window open) · MIMI Labs report + queries: Medicare Physician & Other Practitioners, CY2024 extract + FFS Part B enrollment extract 2025-06-30

CMS proposes to reshuffle the remote-monitoring codes — five years after standardized payment grew from $5.36M to $259.1M (48×, 389k device-supply patients, 17k management clinicians). Every state's RPM footprint as one dot: adoption per 10k FFS beneficiaries (log) vs. management intensity, sized by dollars. Connecticut: 856 per 10k — 3× California — from just 122 clinicians (~190 patients each). Critical lens: the min-clinicians slider dissolves per-capita "leaders" into a handful of monitoring programs.

Data explorer

The Curve Payment Built

Source: FDA press announcement — first TEMPO participant selected: Dexcom (Jul 22, 2026) · MIMI Labs report + query: Medicare Physician & Other Practitioners, CPT 95249/50/51, 2013–2024

FDA just fused device oversight with a CMS payment path, and Dexcom is participant #1. Twelve years of Medicare CGM claims explain why that matters: the curve draws itself from 26,397 interpretation patients to 368,447 (14×), and every inflection lands on a coverage ruling — 2017 DME classification, 2021 fingerstick rule dropped, 2023 all-insulin expansion — never on a sensor launch. Meanwhile clinic-owned CGM peaked in 2018 and fell 66%. Critical lens: professional-fee sliver, FFS only — the levels understate, but the bends don't lie.

Graphical narrative
July 23, 2026

The Story Outran the Ledger

Sources: Olive funding and acquisition announcements · Axios · Fierce Healthcare · Becker's · Healthcare Dive · Waystar SEC filing

Walk two sourced tracks through Olive AI's expansion and contraction, then classify your own twenty-case shadow-mode sample to expose the exceptions a platform demo hides.

Interactive autopsy · ~4 min
July 20, 2026

The Map Under Subpoena

Source: HIT Doc (John Lee, MD) — "The Texas v. Epic story isn't the state's" (Jul 2026) · CMS Promoting Interoperability hospital attestations, PY2023; extract 2025-05-01 · MimiLabs SQL

Five Epic competitors refused discovery in unison — and while the real map of EHR data control gets assembled under subpoena, its public v0 is explorable now: 4,593 attesting hospitals, one dot per state, vendor share vs. market size, switchable across seven vendors. Epic: 40% of buildings, #1 in 34 of 56 states — and California is decided by one hospital (124 vs. 123). Critical lens: buildings ≠ beds ≠ records, and the min-hospitals slider dissolves Delaware's "57% Epic" (n=7) on contact.

Data explorer

The Breach Beat Is Background Noise

Source: Reuters — US companies face rise in cyber attacks (Jul 17, 2026) · HHS OCR breach portal, extract 2025-12-10 · MimiLabs SQL

Two more breach disclosures landed in one wire-service factbox this week and barely registered. Here's the numbness, quantified: 6,501 large-breach reports since the wall of shame opened in 2009, replayed month by month — from one every 44 hours (2010) to one every 12 (2023), 625M individuals, ~1.8× the US population. Critical lens: the biggest breach in history (Change Healthcare, 192.7M) still isn't in the confirmed archive — toggle the pending list on and "quiet" 2024 becomes the worst year ever recorded (~289M).

Graphical narrative
July 19, 2026

1.6 Million Former First-Years

Source: Daily Stoic — "6 Stoic Rules To Beat Ego" (Jul 18, 2026) · CMS Doctors & Clinicians National Downloadable File, extract 2026-06-01 · mimi_ws_1.provdatacatalog.dac_ndf via MIMI Labs

Companion to today's essay on why looking stupid is the price of building: every Medicare-enrolled clinician in America by medical-school graduation year — 1,608,842 people, one histogram, brushable by cohort, filterable by the 30 largest specialties (cardiology's median clinician is 27 years out; NPs, 8). Median clinician: 15 years past year one; 16.4% are in their first five years right now. Critical lens: the right-edge cliff is Medicare enrollment lag, not fewer graduates — brush 2024–2026 and don't believe your eyes.

Data explorer

Beginners Ship

Source: Daily Stoic — "6 Stoic Rules To Beat Ego" (Jul 18, 2026) · FDA 510(k) database, extract 2026-07-13 · mimi_ws_1.fda.device_510k via MIMI Labs

Since 1976, 25,351 organizations have cleared an FDA device — and every year ~500 more clear their first. In 2024, 43% of all clearing companies were first-timers. Animated 50-year chart of first attempts, plus the "beginner tax" by decade: first-timers pay ~32 extra days of median review in the 2020s, and ship anyway. Critical lens: name-based matching inflates "first-timers," 510(k) is the substantially-equivalent lane, and cleared-only data hides the beginners who never made it.

Graphical narrative
July 18, 2026

The Denial Economy's Public Ledger

Source: Healthcare Dive — W&M advances MLR Transparency Act 42-0 (Jul 17, 2026) · CMS Part C & D MLR filings, CY2023, extract 2023-12-31 · mimi_ws_1.partcd.mapd_mlr via MIMI Labs

Congress voted 42–0 to force MA plans to show how much revenue reaches patient care. The ledger it wants to open already has a public v0: every MA and Part D contract's 2023 MLR filing — 635 contracts, one dot each, enrollment (log) vs. adjusted MLR, sized by revenue, red below the 85% floor. $537.8B in, $52.4M paid back: 0.01%. Critical lens: the extreme ratios dissolve above 10,000 enrollees (small-contract artifacts) — and a denied claim is structurally invisible in this file, which is why denial-rate disclosure is the dataset that would actually change behavior.

Data explorer

The Appeal Nobody Files

Source: KFF — MA prior-auth denials for post-acute care (Jul 2026) · denial 65%/54%/12%, appeals 18%, overturns 95%

Congress voted on prior auth twice this week and took both sides on AI — but the decisive number sits in a KFF data brief: 95% of appealed SNF denials are overturned, and only 18% are ever appealed. Watch 100 denials play out as an animated dot field — 18 turn navy, 17 flip green, and dashed rings mark the ~78 of the silent 82 that would likely have been approved. Critical lens: filed appeals are plausibly the strongest cases, so ~78 is an upper bound — the honest claim is still damning.

Graphical narrative
July 17, 2026

The Landing Zone

Source: CMS — Proposed Transformational Medicare Reforms, CY2027 PFS rule (Jul 2026) · CMS Shared Savings Program PY2024, extract 2025-09-29 · MimiLabs SQL

CMS proposed giving traditional MIPS a 2029 expiration date and pointing clinicians at ACOs. Here's the destination: all 476 PY2024 MSSP ACOs, one dot each — beneficiaries (log) vs. savings rate, sized by earned dollars, navy for two-sided risk, red for one-sided. Filter by risk model, search any ACO, and drag the size floor. Critical lens: the 22% "miracle" savings rates dissolve above 20,000 beneficiaries (small-n variance), and 35 ACOs share an identical assigned quality score of 77.05.

Data explorer

The Report Card CMS Just Tore Up

Before the MIPS eulogies, the program's own report card: half a million clinician score records in one morphing distribution, told in four chapters — everyone scores 87, small practices carry 5× the penalty rate, 27% of solo/small records never reported (automatic −9%), and the average earned bonus was a rounding error. Critical lens: rows are TIN/NPI records, not people, and a saturated score can't tell a good clinician from a good billing department.

Graphical narrative
July 16, 2026

The Vendor Fingerprint

Source: STAT — CMS moves to ban remote patient monitoring vendors (Jul 15, 2026) · CMS Medicare Physician & Other Practitioners, CY2024 · via MIMI Labs

CMS proposed keeping the RPM codes and banning the third-party vendors who deliver the monitoring. Every state plotted by two fingerprints of that business model: billing concentration (patients per provider) vs. intensity (device-months per patient), sized by dollars. Toggle to $/patient and drag the sample-size floor. Critical lens: Connecticut's 189 patients/provider looks like a vendor factory but bills at low intensity — one big system, not churn. Traditional Medicare FFS only.

Data explorer

The Code That Outgrew Its Model

Source: STAT — CMS moves to ban remote patient monitoring vendors (Jul 15, 2026) · CMS Medicare Physician & Other Practitioners, CY2018–2024 · via MIMI Labs

Watch the RPM billing hockey stick draw itself — $1.2M in 2018 to $256M in 2024, a 210× run — with the RTM “escape valve” climbing off a tiny base, then the 2027 vendor ban. Critical lens: this FFS line is a floor (industry estimates near $500M add MA and cost-sharing), and the 2027 segment is a projection, not observed data.

Graphical narrative
July 15, 2026

The Public Road

Source: Delaware's Smart Health Network — the neutral state hub · ONC/ASTP AHA Annual Survey IT Supplement, 2024 wave · MimiLabs report + SQL

Delaware bet the health-data hub is public infrastructure; TEFCA is the national version of that bet. Fifty states plotted by hospitals surveyed vs. the share already live on the federal “connect once, reach everyone” network. Toggle to planning/any-network and drag the sample-size floor to watch the small-state leaders (Hawaii, 58% on 12 hospitals) dissolve. Self-reported survey awareness — not measured live-exchange volume.

Data explorer

Nine Billion Faxes, or One Road

Source: Delaware's Smart Health Network · ONC/ASTP AHA Annual Survey IT Supplement, 2024 wave · MimiLabs report + SQL

Watch the tangle of point-to-point “dirt tracks” (N×(N−1)/2 private lines) collapse into a single neutral hub (N). Then the real early road: 461 hospitals live on TEFCA, while 1,106 are still only planning. Critical lens: planning isn't plumbing, and Delaware itself is n=4.

Graphical narrative
July 14, 2026
July 13, 2026

The Machine That Saves Billions

Source: MedCity News — “Prior Auth Is a Fight and AI Won't End It” (Jul 2026) · CMS Medicare Advantage MLR 2023 · via MIMI Labs

Today's Big Thing says AI can't “end” prior auth because no payer unplugs a machine that saves it billions. Here's the machine: 571 real MA contracts, plotted by enrollment vs. medical loss ratio — the share of premium kept out of claims. The giants cluster right above the 85% federal floor and kept $40B between them; the industry kept $57.3B. Light up the billion-dollar valves, then drag the sample-size floor and watch the “some plans keep 25 cents” outliers dissolve into small-n noise.

Data explorer

The Denials Nobody Fought

In 2023 MA plans denied 3.2 million of ~50 million prior-auth requests. In 1,000 dots: reveal the 6.4% denied, then re-base the field to the denials and follow them — 88% were never appealed, and of the 12% that were, 82% got overturned. The denials that stuck, stuck mostly because nobody fought them. Critical lens: the 82% overturn rate ranges 42% (Kaiser) to 94% (Centene), and the 6.4% average hides far higher post-acute denials.

Graphical narrative
July 12, 2026

The 2 AM Distribution

Today's essay: the imagination that builds the tool is the same one that watches it kill someone at 2 AM — and it always renders the failure as a death. Each dot is one real drug in the FDA's 2026 Q1 adverse-event intake: serious-coded reports vs. the share that include death. Flip on the “2 AM version” (a red line at 100%) to see the gap, then drag the sample-size floor and watch the thin-sample outliers dissolve. Median reported death share: 11.4%.

Data explorer

What Anxiety Renders

1,000 real serious-coded drug-safety reports, laid out as a field. The 2 AM version is all red — every report a death, the way your imagination paints it. Press “show what actually gets logged” and the dots drain to their real outcomes while the counter falls from 100% to 14%. The critical lens: this view filters to outcome-coded cases, and spontaneous reporting still over-weights catastrophe.

Graphical narrative
July 11, 2026

Cleared, Then Unwatched

The FDA has cleared 1,451 AI-enabled devices; zero carry a mandatory post-market monitoring rule. Each dot is one real AI 510(k) clearance — year cleared vs. FDA review time, colored by specialty. Flip the switch to see what watches the device after go-live, and drag the sample-size floor to watch the “AI is diversifying beyond radiology” story dissolve into 1–3-device noise.

Data explorer

The 40% Nobody Logged

A cleared, validated CDS tool ran for six months with physicians silently overriding 40%+ of its recommendations — and no one noticed. Watch 100 recommendations resolve, navy for accepted, red for overridden, then reveal the monitoring layer that logged exactly zero of the reversals. Adoption is solved; the audit trail is the unbuilt product.

Graphical narrative
July 10, 2026

The Incidental Heart

Source: European Heart Journal – Digital Health (2026) · Dr. Ashley Beecy (Sutter Health)

34,000 paired chest CTs and echocardiograms. A model that finds reduced ejection fraction in scans ordered for lung nodules, cancer staging, and trauma — not the heart. Filter the scatter by scan indication, see where the model flags correctly, and explore the critical lens: this is a detection model, not a diagnostic. The business isn't the model — it's the “now what?” routing layer.

Data explorer

9.5 Million Patients, One Clearance

The Defense Health Agency just put ambient AI in every clinic it runs — 9.5M beneficiaries, ~400 clinics, one EHR. Documentation time at Wilford Hall dropped from 30–45 min to 5–10 min per note. Compare that to the JAMA multi-site average and your hospital's two-site pilot. The most security-obsessed buyer in American medicine just made ambient documentation infrastructure, not an experiment.

Graphical narrative
July 9, 2026

The Payable Sliver

Source: CMS SaMS interim payment policy (CY2027 OPPS) · CMS Medicare Physician & Other Practitioners

CMS just proposed the first payment lane for “Software as a Medical Service.” Before you build for it, see the one already open: 34 real HCPCS codes — remote monitoring, chronic-care management, e-visits — plotted by reach, pay-per-service and dollars. Step 2020→2024, drag the reach floor, and watch how few “software” lanes exist that don't just rent a clinician's minutes.

Data explorer

The Lane That Grew

Medicare remote-monitoring spend went 38× — $6.75M to $256M in five years — the moment the service got a code. Watch the curve climb, then meet the gap: 97% of doctors review wearable data, ≤6% have it integrated, because a consumer watch isn't the FDA device the lane requires. France had the code and still stalled.

Graphical narrative
July 3, 2026

The Override Nobody Uses

Source: KFF — ACA denials & appeals, 2024 · CMS Transparency in Coverage PUF

Every dot is one insurer: denial rate vs. how often the denial gets overturned on appeal, sized by claims volume. Drag the sample-size floor and watch the 80%+ "fairness" outliers dissolve into small-denominator noise. Insurers deny ~1 in 5 claims; patients appeal ~1 in 600 — but ~42% of appeals win. The override exists; almost nobody reaches it.

Data explorer

The Effect Vanishes

The most rigorous LLM decision-support trial yet cut clinician errors on chart review (−16% diagnostic, −13% treatment) — then showed no significant change in what happened to the patient 14 days later. Watch the gain decay, step by step, toward the bedside, and see the 19.5% action rate where the signal drains out.

Graphical narrative
July 2, 2026

The Real POCQi Explorer

A condensed scatter of all 30 specialties (drag the sample-size floor and watch the fake trend dissolve), plus a live OpenEvidence run showing why the hard half of point-of-care — building the question from a 1,200-document chart — never reaches the test set.

Data explorer
July 1, 2026

The Common Answer Trap

Companion to “Primary care declares independence”

A naive model reaches for the answer the internet gives most often. Play the model, then meet the patient the common answer would have hurt.

~3 min

The Verification Layer

Companion to “Primary care declares independence”

Build the checker that flags a confidently-common wrong answer before a human signs it. Discover why model confidence is the wrong signal to trust.

~3 min
June 30, 2026

Can You Survive the Rephrase?

Companion to “Health AI flunks the stress test”

Frontier AI models ace medical benchmarks — then break when the question is rephrased. See if you can do better.

~3 min

The Readiness Gap Simulator

Companion to “Health AI flunks the stress test”

Frontier AI models top the medical benchmark — then collapse under stress. Toggle the perturbations and watch the leaderboard reshuffle.

~3 min

The PERC Consistency Test

Experiment companion

20 runs. 3 models. 2 temperatures. One patient on Camila. Watch LLMs struggle with the estrogen trap hidden in prior visit notes.

~5 min