The Index constitution

The rules the Index is computed under, versioned and published in full. A level nobody can reconstruct is an opinion, so every rule behind it is here rather than described.

Version 1.0 · in force for the run read 30 August 2026

A method version change breaks the series. Levels computed under different versions are never compared, and the version is published with the level, so a reader can always tell which rulebook produced a number.

The Index this rulebook governs is the Rollcall Health Index. Every superseded run keeps its own page with the roster it scored.

Objective

The Rollcall Federal Integrity Index measures how much of a published provider directory survives comparison with the federal record, across a whole market, on one scale.

It exists because a plan asking how it compares has nothing to compare against. Every plan knows its own directory has defects. None of them knows whether that is normal, and the regulator publishes no benchmark. The Index is that benchmark, built only from files anybody can fetch.

Zero is the floor and higher is worse. A constituent sits at zero when no evaluated record in its directory carries a qualified finding, and the market rate is the same measure over every constituent at once. A directory we could not read carries no score, which is never recorded as zero.

Eligible universe

The universe is defined by CMS, not by us. It is every directory filed in the CMS machine-readable URL file for the plan year, which is the register issuers are required to file into.

The register is the boundary. A directory that is not filed with CMS is outside the Index, and one that is filed cannot be excluded by us for being inconvenient to read.
The unit is the published directory, not the issuer or the company. One file can serve two dozen issuer registrations and one company can publish several files, so the file is the only unit that needs no judgement call.
Stand-alone dental is in the universe. CMS regulates those filings on the same register. They are identified separately on the record so the Index can be read either way.

Constituent selection

A directory in the universe becomes a constituent only when all six hold. Anything else is reported, never scored.

The file was retrieved and processed.
Every record-level check in the Index completed against it.
Processing was complete: every provider file the directory published was read.
The same method version was used as for every other constituent.
It was measured in the same audit period.
Findings were deduplicated before scoring.

There is no minimum size screen today. A directory of 182 providers and one of 1.9 million are both constituents and are ranked on the same scale, so a small denominator carries less evidence behind the same rate. The spread below is published for that reason: a level says nothing about whether it rests on one large directory or on many.

Weighting

The market Index is weighted by evaluated records. The constituent scores that sit under it are not weighted at all.

In the market Index, every evaluated record carries equal weight. A directory of two million records moves the Index a hundred times further than one of twenty thousand, which is what makes it a market number rather than an average of directories.
In a constituent score, size does not enter. The score is that directory's own rate against its own denominator, which is the only way two directories of different sizes are comparable at all.
Nothing is weighted by anything we chose. There is no severity weighting, no adjustment for provider mix, and no discretion in the arithmetic.

The listing-weighted market number and the median constituent answer different questions, and both are published: one is the market, the other is the typical directory.

Calculation

Constituent rate

rate = 10,000 × Q ÷ E

Q = evaluated records carrying at least one qualified finding
E = records successfully evaluated against every scored check

A record carrying two findings is one record. Counting it twice would let a single provider who is both deactivated and excluded move a directory further than a second defective record would.

Market index

index = 10,000 × ΣQ ÷ ΣE

Bands, and why there is no league table

The public reading is the band, not a position. A constituent is placed BELOW, WITHIN, ELEVATED or PRIORITY against frozen constants, so the band answers the only question a plan can act on: is this rate what a file of this size would be expected to carry. We hold an internal order to compute the bands; we do not publish it as a league table.
Bands are absolute, not relative. They are measured against constants frozen with the method version, so a constituent moves band only when its own rate moves. Nobody is demoted because somebody else improved.
Size is accounted for before a band is assigned. A small file will swing wildly run to run on chance alone, so the expected spread is widened to match the dispersion the market actually shows, and a directory below the evidence floor is marked too small to score rather than given a flattering one.
Nothing is normalised, standardised or curved. The rate is a count over a count. There is no distribution being fitted to it and no scaling applied.

Rounding is presentation only: rates display to two decimals, and every comparison behind them uses the stored value. Movement is only comparable between runs sharing a method version and the same checks.

Review and reconstitution

The Index is recomputed on every completed pass. A pass reads the register and every index it names, and publishes only when at least 99.5% of the provider files those indexes name were read. The number discovered, the number read and the shortfall are published with the run. This is a floor on the pass and not on a directory: the selection rule above is unchanged, and a directory is ranked only when every file it published was read.
A constituent enters when it first meets all six selection rules, and leaves the moment it stops meeting them. Leaving is not a score of zero, and re-entry carries no penalty.
A method version change breaks the series. Numbers computed under different versions are never compared, and the version is published with the level.
The constituent count is expected to move. It is a property of what issuers filed and what could be read, not a target.

What a constituent change does to the level

The headline number is a level, computed over whoever is in the Index that week. When a directory enters or leaves, the level moves for that reason alone, before any published file has changed. A share index behaves the same way when its membership is reconstituted, and it is not a defect in either case.

An entering directory brings its whole rate with it. A single large or badly affected file can move the market level on the day it joins, and that movement says nothing about the directories that were already there.
So the level alone cannot answer "did directories get worse". Read on its own it will sometimes say yes when the answer is no, and the other way round.
A chained series is published beside it. The chain links each run to the one before across only the directories present in both, so a change in it is a change in published files rather than a change in membership. It appears on this page as the same-constituents level.
Both are published every run, and neither replaces the other. The level is what the market looks like now. The chained series is what moved. Quoting one without the other is how a reconstitution gets reported as a collapse.

The chain carries its own method version, and the same rule applies to it: numbers computed under different versions are never compared. A run whose composition changed is disclosed with the figures rather than smoothed away.

Treatment of special cases

No score

Unavailable

A directory that could not be read has no evaluable denominator, so it carries no score. It is never recorded as zero, which on this scale is the cleanest reading a file can get.

No score

Partial coverage

A directory read in part does not enter the Index. A rate over the half of a file that answered is not a rate.

Scored

No qualified findings across completed checks

Said exactly that way. It is not the same claim as clean, and it says nothing about fields these checks do not read.

Data sources and versioning

The register: the CMS machine-readable URL file for the plan year, which defines the universe.
The NPPES cumulative deactivation report, matched on NPI. A cancelled NPI is absent from the main NPPES file rather than flagged in it, so the deactivation report is the only source that carries it.
The HHS-OIG exclusion list, matched on NPI only and never on a name. Active means the record carries no reinstatement date.
The directories themselves, read from the addresses their issuers filed.

Each federal file is copied to a dated, immutable snapshot before a pass uses it, so any position in this Index can be recomputed against the exact file it was measured against. The method version is the pass together with the two federal files behind it, and it is published with the level.

Governance and publication

A computed pass is not a published one. It loads as RUNNING, and publication is a separate command a person runs. Nothing on a schedule can publish.
A pass that failed its reconciliation checks is held, and so is one with no checks at all, for the same reason: nothing verified it.
Every qualified finding keeps its dated evidence, with the federal record and the file it was read from.
Notice goes to subscribers, free, at findings@rollcall.health: every finding that names them, with the evidence, when it publishes. Organisations that have not subscribed are not privately noticed; publication is the notice. Silence is never published as agreement.
A finding can be disputed at any point, including after publication. Evidence is read when it arrives, and a finding shown to be wrong is withdrawn with the correction published in the same place.
The same rules apply whether or not a plan is a customer, and no constituent can pay to be removed, reweighted or ranked differently.

Limitations

The Rollcall Federal Integrity Index measures the share of successfully evaluated directory records with no qualified finding from the federal-status checks in the stated method version. It is not a measure of overall provider directory accuracy, regulatory compliance, provider eligibility, credentialing status, or any field Rollcall Health did not evaluate.

A position in this Index is a statement about a file on a date. It is not a statement about a plan's conduct, and nothing here establishes that anyone was paid, treated or harmed.

Two constituents are never compared across different check coverage, a check that did not run is never scored as a pass, and an unavailable directory is never treated as one with no findings.

Index announcements

What changed, and when

Every change to the panel or to the method, dated. A level that moved because the panel moved is never reported as the market moving.

2 Sep 2026: Nothing published about the current run identifies an audited directory. The page per audited host, the list of all of them, the search control on the home page and the band beside each plan on the county pages have all been removed; the level, the spread, the series and the coverage are unchanged. This also corrects the entry of 31 August below: the dated copy of the run published on 30 August was left naming its constituents, and since that was the run then current, the machine-readable file said the directories are not named while the copy beside it named all 98. That copy now matches the run it describes. The archive is deliberately not covered: every superseded run keeps its page and its dated copy exactly as published, constituents named, because a superseded measurement is a record and rewriting it would break every citation of it. Nor can a correction be complete once a file has been served for caching: the dated copies are sent with a year’s cache lifetime, so a copy already fetched stays as it was fetched.
31 Aug 2026: The Index stopped naming its constituents. The table of every directory with its rate, the ordered ring of named nodes on the home page, and the row per directory in the machine-readable file are all gone; the level, the spread across bands, the same-constituents series and the coverage funnel are unchanged. A directory’s own rate and the dated federal records behind every finding go to the organisation responsible for it, on request. Snapshots dated before today name their constituents and are left exactly as published rather than rewritten.
31 Aug 2026: The constitution was amended to state the 99.5% floor. It had said a pass reads every provider file, which the run published the day before did not: it read 1,325 of 1,331. The rule now says what the run does, the coverage fraction is published with it, and what a directory must do to be ranked is unchanged. The contradiction was live for a day and is recorded here rather than quietly corrected.
30 Aug 2026: The panel fell from 107 constituents to 98. The published level moved from 18.16 to 15.58, but the same-constituents level moved only from 13.48 to 13.22: most of the fall is which directories were in the panel, not what they published.
30 Aug 2026: Six directories that were read but not finished now show their last measurement, dated, instead of a blank, as unreadable ones already did. They are outside the level and outside the panel either way: what changes is that a directory can no longer erase its own published score by failing to serve a file, not that the panel holds it.
30 Aug 2026: Correction to the entry of 15 August 2026. Carrying a measurement forward keeps a directory's own score visible; it does not keep the directory in the panel, and a directory that cannot be read still leaves both sides of the market ratio. Nine directories carry forward this week and the panel fell from 107 to 98 accordingly.
30 Aug 2026: A run now publishes when at least 99.5% of the files it discovered were read, rather than requiring every one. What a directory must do to be ranked is unchanged: every file it published has to have been read, or it carries counts and no rate.
21 Aug 2026: Moda Health re-entered the panel after a week in which its directory could not be read, at 272.58 per 10,000. It contributes +4.73 of the +4.59 the published level moved, and the same-constituents level fell over the same days.
21 Aug 2026: The same-constituents series was published, as method 1.1. It chains each run to the one before across only the directories present in both, and is backfilled to the first run.
15 Aug 2026: Three directories carried an earlier measurement forward rather than leaving the panel, so a file that cannot be read cannot improve the market by disappearing from it.
13 Aug 2026: First publication, 106 constituents, level 13.69. That level is the baseline every later run is measured against.

The market this governs

The level, the spread across bands and the coverage of the most recent published run.

Open the Index Read the method