The price of poker: sources and method

“The price of poker” reads the Anthem and Sentara standoff from the public files. This page is the record behind it: what was pulled, the rules that decided what stayed in, how the figures were checked, what was corrected along the way, and dated captures of both campaign pages. If the piece and this page ever disagree, this page is the one that is right.

The rules the dataset was built under

Population. Commercial group products only, at seven Virginia hospitals publishing machine-readable files under the federal price transparency rule, with Anthem’s PPO as the anchor product. Excluded by rule: exchange and marketplace plans, individual and family products, federal employee plans, union and association supplementals, out-of-state Blues, hospital employee plans, workers compensation, and every government line, including Medicare, Medicare Advantage, Medicaid managed care, TRICARE and VA community care.

Unit. One row is one hospital, one MS-DRG, one payer, one plan, one setting, one billing class. The published quantity is the median allowed amount: what the hospital reports it was actually paid over a 12 to 15 month lookback, under the CY2026 rule that replaced the old estimated allowed amount. It is not a billed charge and it is not the negotiated rate on file, except where the piece says filed rate and means it.

Derivation. Every median, count and percentile is read from a single row. Every percentage, ratio and multiple is arithmetic on quantities the piece itself prints, and the piece names the inputs where each one appears.

The corpus

The Sentara files are listed at sentara.com/cms-hpt.txt and download named by employer identification number rather than by hospital; each was identified by the location name inside the file’s own metadata row. Everything else the piece relies on is enumerated in the source list under the piece itself: the CMS Hospital General Information and readmissions files, the CY2026 outpatient rule, the quality grades, both systems’ audited statements and securities filings, the trade press for the member split, and the geocoding and routing services behind the map.

How the figures were checked

The working rule for this publication is that a figure is not verified because it was checked. It is verified when it has been re-derived from the primary file by an independent second path and the two agree. For this piece that meant three extractors, written separately, none importing or calling another: one walks the files with a manual column map, one is a pandas pipeline with its own DRG detection, one reads positionally with a state machine over the code columns. The first two returned 631 keys each, with zero keys in one and not the other and zero field-level disagreements on any shared value. The third, which does not read the UVA file, returned 459 keys and agreed with the first on all 957 field comparisons. It also reproduced every multiplier on the piece’s Figure 4, from 1.000 to 1.310, identically to three decimals on all three admissions.

The rendered page is then audited by a harness that asserts every load-bearing number against the dataset it shipped from, on every rebuild. At this writing it runs more than a thousand individual checks, and it was mutation-tested: deliberate defects were planted in medians, counts, ratios and contract terms to confirm the harness actually fails. Where it did not, the hole was closed and the defect class recorded.

The same two-path rule covers the pieces of the record that are not price files. The map’s 72 hospital pins come from the CMS address file geocoded through Esri, and the seven hospitals carrying price figures were re-derived independently from OpenStreetMap; the two paths agree to under 310 meters on all seven. Road distances were re-run through a second routing service. The readmissions figures were pulled twice from the CMS file by different readers, and the 42 values agree exactly. And because a geocoded pin can still sit a few tenths of a mile from the front door, every distance claim in the piece was tested for whether that snap could flip it; the claims a pin snap could flip were restated in terms it cannot.

What the filters caught

127 distinct payer and plan string combinations appear on the three admissions the piece studies across the seven files. All of them were read, none sampled. The exclusion rules above remove 56, leaving 71 commercial group strings. Among the catches: a workers compensation plan that survived the first version of the exclusion rule, which had no term for it; a single named employer inside one plan string; behavioral health and transplant carve-out contracts that are not general commercial agreements; and, at UVA, four payers filing against one identical Anthem plan string, which is a rented network. Only the Anthem row there carries a median and a count, so the published UVA figure is uncontaminated, but it is why UVA carries an asterisk wherever it appears.

One caveat could not be filtered away and is stated instead: the product grain is not held constant between payers. Anthem files three named commercial products at every site while another national carrier files one bucket at one hospital and three strings elsewhere. Measuring that grain corrected a published figure, which is recorded below.

The trap in the unit

The methodology column, read on every kept row, is where the piece’s reframe came from. Sepsis and heart failure price off a DRG base rate at all seven hospitals, so they compare like for like. Vaginal delivery does not: two hospitals price it off the general base rate, two carry a negotiated delivery carve-out, and the Inova hospitals split it into a short-stay case rate and a per diem. So $18,428 at Norfolk General against $7,246 at Martha Jefferson is a base rate against a carve-out, and the piece frames it as a contract-structure difference rather than a price difference. The gap itself is not case mix: Norfolk General’s 10th percentile, $16,634, sits above Martha Jefferson’s 90th, $8,758. The distributions do not overlap at all.

What survives every rule

Every headline figure was recomputed under at least two defensible inclusion rules, and the piece only makes the ordering claim that survives all of them: on sepsis and heart failure, Sentara Norfolk General is the dearest hospital in the set and Sentara Northern Virginia the least, on paid medians and on filed rates alike. Dropping Norfolk General entirely still leaves a 1.45x spread on sepsis, 1.52x on heart failure and 1.60x on delivery across the remaining hospitals, five of them on delivery, where one files no median. The multiple-of-Medicare figures were rebuilt the same way: every Medicare payment was re-pulled from the CMS file against the verified provider numbers, and the piece confines that multiple to pairs whose Medicare denominators sit close together, because it measures wage index and teaching adjustments as well as negotiating position wherever they do not.

Corrections, named

Things this process caught in my own drafts, fixed before publication and recorded here because a correction made quietly is a correction a reader cannot check.

• A draft claimed the widest gaps over Medicare sit at the buildings with the fewest substitutes. The map’s own data contradicted it: the hospital with the fewest neighbors carries the narrowest multiple. Rewritten to what the data supports, that the widest gaps sit where the services have no substitute.
• The care delivery margin read +6.7 percent against a denominator the piece never printed. The reproducible figure is +7.1 percent, and the caption now names the denominator.
• “All files dated April 1, 2026” was wrong: UVA’s is dated March 24. The claim appeared twice and both were corrected.
• A carrier’s share of commercial rows was quoted flat at 85 percent across three hospitals. It is 85 in Hampton Roads and 90 at Martha Jefferson; the grain caveat above is why.
• A quality claim said one hospital had held its safety grade five cycles running. The cited sources carry two cycles. Cut to what the sources carry.
• A burn-center claim was attributed to a state trauma-center page that cannot contain it. The trauma claim kept its source; the burn center is now attributed to the hospital’s own description.
• An early draft of the map annotation printed each rival’s multiple of Medicare, and its first reader concluded a hospital cost twice what its neighbor does when the negotiated dollars sit closer together. The map now prints the dollars.
• A draft sentence had a hospital pair reading in the wrong direction entirely; building the map is what exposed it, and the surrounding argument was rewritten against the road distances.

The figures that are in no data file

The campaign pages, captured

Both sides of this dispute argue their case on their own websites, and pages like these come down when a dispute settles. All three pages were captured in full on August 22, 2026, read from the live sites rather than a cache, and pushed to the Internet Archive the following day. The archived copies are the durable record:

• Anthem, “Lowering Healthcare Cost Trends in Virginia”
• Sentara, contract termination press release, July 31, 2026
• Sentara, “Keep Sentara” member page

Anthem’s page names the eight facilities it says would leave the commercial network on January 1, 2027: CarePlex, Leigh, Norfolk General, Obici, Princess Anne, Surgical Suites of Coastal Virginia, Virginia Beach General and Williamsburg Regional, all in Hampton Roads. Sentara Martha Jefferson, Sentara RMH, Sentara Northern Virginia and Sentara Halifax are not on that list and fall in the December 31, 2027 tier for commercial and ACA members.

On the money, each page characterizes the other side’s position. Anthem describes Sentara’s ask as “as much as 30% over the next three years” for employer-sponsored plans. Sentara’s member page describes its own ask as “a modest, single-digit overall rate adjustment,” and trade coverage in early August put the figure at a 6.2 percent blended increase for 2027 against an Anthem counter of roughly a 1 percent decrease. Neither party has published its actual proposal, so both numbers are characterizations, and the piece treats them that way.

The standoffs counted in this piece

The rule. A dispute is counted if the provider is a multi-hospital system or a flagship academic medical center, the contract at issue covers employer-sponsored commercial coverage, the dispute was live at some point between September 1, 2025 and August 25, 2026, and at least one of the two parties published a page about it on its own website. Medicare Advantage disputes with no commercial contract at issue are excluded. Single hospitals, physician groups and clinically integrated networks of independent hospitals are excluded. Forty-two disputes clear that rule. It is a verified sample, not a census, and the census is in the paragraph below.

The census. FTI Consulting has counted publicly reported payer-provider contract disputes since 2022 by searching news databases each quarter. Its published annual totals are 51 disputes in 2022, 86 in 2023 and 133 in 2024. The four quarters of 2025 as FTI published them add to about 170, which is my arithmetic on its figures rather than a total FTI states. Two things belong with that series. FTI revised its own fourth-quarter 2025 count from 76 to 83 between two of its posts, so this is a tally of media coverage rather than a registry. And its first quarter of 2026 came in at 22, which FTI attributes to the seasonal lull at the start of a plan year.

Reproducing this

The piece names every source and this page states every rule. A reader with the seven files and an afternoon can rebuild any figure here, apply the filters, check them, or reject them. That is the point of writing from the public record: you do not have to take my word for any number on the page, and neither does anyone who disagrees with the conclusions.

Fringe Theory is independent and unaffiliated. Views expressed are my own and do not represent those of my employer. Nothing here is legal, tax, medical, or investment advice. Figures described as modeled or derived are my own arithmetic from the public sources listed, not disclosed data.