Source manifest: Mockingbirds

Every figure in the article traces to one of the files below. Nothing is modeled from private data and nothing came from a vendor. If a number here conflicts with one you hold, the file is the arbiter, not the article.

Captured August 20 and 21, 2026.

How to retrieve any of these yourself

Every hospital in the United States must post a plain text file at https://<hospital-domain>/cms-hpt.txt naming the location of its own machine-readable file. That is the fastest and most reliable route in, and it does not depend on the hospital's own site search. Open that text file, take the URL it names, and you have the same source I used.

The nine files

All nine are CMS schema v3.0.0. "Wide" files carry one column set per payer and plan; "tall" files carry one row per payer and plan. Both encode the same fields.

Two entity notes worth keeping. Fauquier's file is filed under the 2013 joint venture entity, not under a Bon Secours identifier, which is a cheap way to see whether the transaction has closed at licensure level. Both Inova files carry the same EIN and are, at the rate level, the same file: 21,427 rows in the same order, with the gross charge differing on 24 of them.

The extracted datasets

• mrf-rate-comparison-2026-08-21.csv - 886 commercial contract rates. Columns: hospital, code, service, payer group, plan, rate, whether the dollar was published or derived from a percentage, the percentage where applicable, and the methodology the hospital stated.
• mrf-allowed-amounts-2026-08-21.csv - 264 published allowed amounts. Columns: hospital, code, service, payer group, plan, median allowed, 10th percentile, 90th percentile, claim count, the rate on file, and the ratio between the two. A claim count reading "1 through 10 (suppressed)" is the hospital's own suppression, permitted by CMS for any count under eleven.

What was filtered, and why

• Commercial only. Medicare, Medicare Advantage, Medicaid, TRICARE, CHAMPVA and VA Community Care rows were excluded throughout.
• Six payer groups. Aetna, Anthem/BCBS, CareFirst, Cigna, Sentara Health Plan and UnitedHealthcare. Rental networks, TPAs and generic commercial rows were excluded so the same six carriers are compared across all nine hospitals.
• One charge line per code. Where a hospital files more than one charge line for the same code, the plain chargemaster line was used rather than an episode estimate. UVA Medical Center files both: a lumbar MRI at $1,513 and at $5,627, an hour of infusion at $1,681 and at $25,158. The lower line is the comparable one and is the one used.
• Aliased payers folded. UVA files CareFirst and UnitedHealthcare members under the Anthem contract at identical dollars. Those are one contract, counted once, not three.
• Per diem rows use the stated cap where the hospital gives one, because a per diem is not an episode price and an inpatient stay runs past it.

Known defects in the source files

These are the hospitals' own figures, reproduced as filed.

• UVA Haymarket states a per diem of $3,094 for CPT 72148 against its own gross charge of $2,237, while the same row's minimum and maximum columns read $2,130 and $2,237, and the median it reports actually receiving is $2,237.
• Negotiated rates exceed the hospital's own gross charge on 521 Fauquier rows and 2,113 Sentara Northern Virginia rows.
• Bon Secours Rappahannock leaves 289 percentage-priced rows without either an allowed amount or the note explaining its absence. Every other filer here explains every one of them.

The same-contract control

The two Inova files share EIN 54-0620889 and carry identical negotiated rates: 21,427 rows in the same order, gross charge differing on 24 of them. That makes them a control for everything that is not price.

For each payer and service where both files publish an allowed amount, take the ratio of the higher median to the lower. Because the contract is identical, that ratio contains only case mix, length of stay and outlier behaviour. Across 21 payer-and-service pairs it runs to a median of 1.07x, a mean of 1.15x and a maximum of 1.68x.

Any cross-hospital spread materially above that floor is unlikely to be case mix alone. Any spread at or below it should not be read as a price finding, which is why the CareFirst row in the article is flagged and the Anthem row is described as travelling well rather than as evidence of dispersion.

Reproducing the arithmetic

Two figures in the article are calculations rather than published statistics, and both are labeled as such in the text.

1. The variance decomposition. For each service, one observation per hospital-and-insurer pair, price on a natural log scale. Ordinary least squares against hospital dummies alone, then against insurer dummies alone. Reported figure is the R-squared of each fit.
2. The volume elasticity. Log median allowed amount regressed on log claim count with a fixed effect for every hospital-and-service combination, using only rows with an unsuppressed count. 91 observations, coefficient -0.011, t-statistic -0.27.

Notes that did not fit in the article

The lookback windows do not line up. The CY2026 rule allows a lookback of no less than twelve and no more than fifteen months, and the nine files carry different dates (March 24 to July 14, 2026). No two hospitals here are therefore describing exactly the same period. That is the same defect the article criticizes in a carrier repricing exercise, and it applies to these numbers too.

Hospitals get their own files wrong, and that is part of the finding. On a code not shown in the article, one of these nine states a per diem of $3,094 for a lumbar MRI against its own gross charge of $2,237, while the same row's minimum and maximum columns say $2,130 and $2,237. The median it reports actually receiving is $2,237. That is why the file has to travel with the number, and it is one reason imaging was cut from the chart.

Why this runs off the hospital's file rather than the carrier's. The payer-side Transparency in Coverage files, published under 45 CFR 147.212, carry contracted rates and nothing else: no allowed amounts, no claim counts, no volume. The Congressional Research Service reports that individual files can reach a terabyte, that a single state index listed roughly 1,800 plans, and that estimates suggest a majority of the rates in them are clinically implausible (CRS R48570, June 13, 2025). The percentage-of-charges gap that the CY2026 rule closed on the hospital side remains open on the payer side.

The population filter, in full

Revised August 21, 2026, after two reader challenges. The first version of the carrier table averaged every plan a payer filed at a building. That was wrong twice over, and both corrections are recorded here rather than quietly applied.

First pass, market segment. The hospital files list every product a payer files at that building in one list. Group HMO, group PPO and POS, and also individual, marketplace and exchange. Labels seen across the nine include Aetna (IFP), Anthem Exchange, exchange hmo, Cigna Connect Exchange, Sentara Exchange, Optima (Indiv).

Second pass, everything else that is not Virginia commercial group. The first filter was too shallow. It missed ANTHEM BCBS HIX VA (HIX is an exchange product), the federal employee lines (ANTHEM BCBS VA FEDERAL, BCBS SC FEDERAL), eighteen out-of-state Blues plans filed at the Bon Secours facilities (BCBS SC, TX BCBS, CA BCBS, GA BCBS, IL BCBS, AR BCBS, PA BCBS HIGHMARK, BCBS OUT OF STATE, BLUE CROSS OF NC), the hospitals' own employee plans (ANTHEM AMERIBEN BSMH EMPLOYEES, BCBS SC TEAMMATE), a union supplemental (CIGNA NALC SUPPLEMENTAL), and a single named employer's plan (fannie-mae at both Inova buildings).

The leased-network rows. UVA files four separate payer entries against one plan, Anthem (PPO PAR HMO): ANTHEM [30001], ANTHEM CAREFIRST [30008], ANTHEM UHC [30009] and UNITED HEALTHCARE [40032]. So at UVA, what is filed as UnitedHealthcare is priced on an Anthem network product. Those rows are dropped from UnitedHealthcare, because they measure the Anthem network rather than UHC's own. This is worth naming on its own: at one hospital in this set, two carriers are buying through the same network.

CareFirst is removed from the study entirely. It appeared at only three to six of the nine buildings depending on the service, and its UVA Medical Center row came in through the ANTHEM CAREFIRST alias rather than a CareFirst contract of its own.

Net effect on the six retained services: 644 contract rates become 354, and 196 published allowed amounts become 127, of which 72 carry a suppressed count.

The networks are not held constant, and cannot be

The carrier is held constant. The network product is not. Across the surviving plan rows the files use hmo, ppo, Hmo/Pos, Ppo/Pos, NAP CHOICE POS II, OPEN ACCESS PLUS OAP, CHOICE PLUS, options and nonoptions, RPN, Optima, and bare aggregates that name no product at all (All Commercial Plans, United (All Payer), Aetna). One label, Anthem (PPO PAR HMO), names two products at once. There is no way to match one product to one product across nine hospitals from these files.

So where a carrier files more than one group product at a building, the figures average them, and the article says so on the caption.

How much that costs, measured rather than assumed. Taking every case where the same carrier files both an identifiable HMO and an identifiable PPO at the same building for the same service, 30 such pairs: the two medians differ by a median of 1.02x, a mean of 1.09x and a maximum of 2.05x. Product mix is therefore a small part of the carrier spreads and not the whole of them.

The Inova same-file control

The earlier draft described the two Inova files as carrying "byte-for-byte identical negotiated rates." That was false and the chart directly beneath it said so. The two files publish the same 21,427 rows in the same order and share one chargemaster, and 98.4575% of cells match across all 320 columns, but the Anthem negotiated-dollar column differs on 5,140 rows and the UnitedHealthcare columns on about 4,000 each.

Recomputed on the study population: of 60 payer-plan-service rate pairs, 18 are identical and the rest differ by a median of 1.17% and at most 14.8%. Anthem and UnitedHealthcare carry facility-specific carve-outs; Cigna and Aetna match.

The paid-amount gap between the two buildings, 15 payer-and-service pairs on group commercial plans: median 1.014x, mean 1.023x, maximum 1.088x. At the top end the paid gap is smaller than the rate difference, which is the sense in which case mix adds very little.

Note that the earlier reported maximum of 1.68x was the CareFirst pair. Removing CareFirst removed the widest same-contract gap, so the control is tighter partly by construction. Both readings are recorded here so the choice is visible.

The variance decomposition, and what one carrier was carrying

One observation per hospital-and-insurer pair, R-squared of log(rate) on hospital identity alone against insurer identity alone.

Building percentage first, insurer second. The echocardiogram inversion depended almost entirely on CareFirst. With CareFirst removed the appendectomy is the only service where the insurer clearly explains more than the building, and the echocardiogram sits close to even. The article now says that rather than claiming the inversion.

Checks run against the extracted data

Run after a request to confirm the data is clean. It re-derived the dataset from the nine primary files by a second, independent path and diffed the two row by row. Four things came out of it.

One contaminated row was in the published set. A CPT 96413 line at UVA Medical Center under Anthem carried modifier PN, which marks a non-excepted off-campus outpatient department. That is not the same service as the unmodified line and it is now removed.

One legitimate row was missing. A second Anthem echocardiogram line at UVA Medical Center, median $1,637 with a suppressed count, had been dropped by the original pipeline. It is now included. The two changes cancel, so the total is still 127.

Two hundred and one rows were correctly excluded, and the reason is worth recording. A naive re-derivation that took the negotiated-dollar column at face value produced a very different variance decomposition. The cause was per diem rows. UVA files Aetna sepsis at $7,700 with the methodology column reading per diem and the algorithm column reading "Per diem ($7,700). If length of stay < 4.8, ...". Comparing that against another hospital's $39,000 case rate is a daily rate against an episode rate. The original pipeline uses the stated cap where a per diem carries one and excludes it where it does not. That is the correct treatment and the second pass confirmed it the hard way.

Seventy revenue-code rows were correctly excluded. A search on the value 871 also matches revenue code 871, which at UVA is a CAR-T cell therapy blood product, with real negotiated dollars attached. Code type has to be checked, not just code value.

Also confirmed directly against the two Inova files: 21,427 rows each, 320 columns each, identical headers, 98.4575% of the 6,856,640 compared cells matching, and the gross charge column differing on exactly 24 rows.

Percentile sanity holds on every row (10th <= median <= 90th, zero violations), there are no duplicate rows in either array, and no allowed amount is zero or negative.

Sensitivity, and what does not survive it

Every published decomposition figure was recomputed with each carrier removed in turn, and with each hospital removed in turn.

• The three inpatient admissions are robust. Across all fourteen exclusions the building share runs 34% to 65% and the insurer share 3% to 27%. The direction never reverses.
• The appendectomy reverses once. Baseline 22% building against 37% insurer; drop Anthem and it becomes 37% against 27%. The article says so.
• The echocardiogram does not survive at all. Dropping Anthem takes the insurer share from 24% to 3%; dropping Sentara takes it to 46%. It is excluded from the article rather than reported.

Figures that are not in the files

Bed counts were removed from both the article and the chart. They never appear in a machine-readable file, they were not sourced, and in a piece whose argument is that the numbers are checkable, an unsourced round number is worth more to a hostile reader than every sourced one.

The file-level counts, verified against the files

The earlier passes checked the rate and allowed-amount datasets. This one checked the whole-file counts quoted inside the chart's own method panels, which no pass had touched. Every one was recomputed from the primary files. Seven were wrong.

Verified correct. All nine file row counts (Fauquier 194,061; Bon Secours St. Francis 91,084; Rappahannock 62,492; both Inova 21,427; Sentara Northern Virginia 163,403; Martha Jefferson 130,121; UVA Haymarket 482,614; UVA Medical Center 2,241,494). Fauquier's 169,018 commercial payer rows, of which 5,761 state a dollar, 3.4%. Fauquier's 143,325 percent-of-billed-charges rows. Fauquier's published sepsis medians, $9,139 to $14,587.

Corrected.

The sepsis figure had moved because the population filter removed the plan that held the low end. The rest were simply never checked.

One defect claim was wrong in both directions and has been rewritten. The chart stated that a negotiated rate exceeds the hospital's own gross charge on 521 Fauquier rows, 2,113 Sentara Northern Virginia rows and 572 Sentara Martha Jefferson rows. Recomputed every way the sentence could be read (negotiated dollar against gross, derived dollar against gross, commercial only and including government), the answer at all three of those hospitals is zero.

It does happen, at the two hospitals the sentence did not name: 32,073 rows at UVA Haymarket and 41,059 at UVA Medical Center, at a median of 2.1 times the gross charge. An Anthem case rate of $101,963.91 sits against a gross charge of $53,372.31 on the same acute leukemia line. The bullet now says that.

The second half of the same bullet was supported and stays, with its count added: Fauquier records percent-of-Medicare figures above 100% in the same column as percent-of-charges figures, on 13,832 commercial rows.

Also corrected in the chart panels: the sepsis note said 44 of 67 allowed-amount rows carried a suppressed count (it is 23 of 43); the hip fracture note said three filers and five rows (two filers, three rows); the badge explainer said seven filers and 7 of 7 (nine filers, 9 of 9 on the DRGs and 7 of 9 on the CPT codes); a chip tooltip said five filers; a reference marker labeled "Medicare outpatient rate, about $5,899, the APC price Bon Secours states in its own file" was in fact a Medicare Advantage negotiated rate at Bon Secours St. Francis, on a chart that excludes Medicare Advantage everywhere else, and has been removed; a sentence still named lumbar MRI and delivery, both cut from the chart; an unsourced bed count sat in the echocardiogram note; and the chart's disclaimer was a shortened non-canonical form, now replaced with the full house block, which matters because this same chart publishes standalone as a tool page.

The two external facts inside the chart

Neither appears in any machine-readable file, so neither is caught by any data pass. Both were checked against primary reporting on August 21, 2026 and both were made precise rather than left as a bare year.

• Bon Secours acquired Rappahannock General Hospital on December 31, 2014. Virginia Business, "Bon Secours to acquire Rappahannock General Hospital Dec. 31," published December 22, 2014, and confirmed as finalized by Becker's Hospital Review. The chart's "eleven years on" is measured from that closing.
• Optima Health was renamed Sentara Health Plans effective January 1, 2024. The chart previously said 2023, which is the announcement year, not the effective date. Sentara Health Plans states members received new ID cards to use beginning January 1, 2024. This matters because the nine files variously label the same payer Optima, Optima (Group) and sentara health, and the piece folds them together.

Chart label corrections

The hover readout composed its label as carrierGroup + " " + planName. In 25 of the 50 group/plan pairs the filed plan name already begins with the carrier name, so the readout stuttered: "Aetna Aetna (Comm. Book of Business)", "Cigna Cigna HMO", "Sentara sentara health|hmo", "UnitedHealthcare UNITED HEALTHCARE [40032]".

Replaced with planLabel(car, pl), which strips a leading run of plan tokens that either appear in the carrier group's own words or spell a prefix of the group name with punctuation removed. The second rule is what catches "UNITED HEALTHCARE" against the single-word group "UnitedHealthcare". Swept all 130 group x plan combinations: no stutter remains. No figure changed; this is a label fix only.

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.