# Pandemic fraud estimates: what each published figure measures

Published estimates of pandemic relief fraud, improper payment and recovery, each with the issuer's own term for what it counts, the amount exactly as printed, the denominator and period the issuer gives, and the caveats the issuer itself states: 22 estimates from 12 documents by 8 issuers.

Published at https://pandemicdarlings.com/data/pandemic-fraud-estimates/ . Compiled table; not part of release pandemic-darlings-data-v1-2026-07-10. Built 2026-09-21.

## Files

| file | rows | bytes | sha256 |
|---|---|---|---|
| `pandemic-fraud-estimates.csv` | 22 | 30741 | `dd142562c796a9ce0f34dc49d21d83fb048875b5487b028298e63646143b5df5` |
| `definition-classes.csv` | 15 | 2505 | `f35f845949829652c9fc90dc48f67f5273be1a4428dba527d097905bf20dc6d6` |
| `comparability.csv` | 210 | 60612 | `346c466bb83901f19cc711718f7e0cc9faf88aedf53dd21445a89f41f12c0f0b` |

`dictionary.csv` beside this file describes every column of the served table: measured type, blanks, distinct values, range, example and meaning. `SHA256SUMS` carries the SHA-256 of every file on the page, including the page itself.

## What this is

One row is one published estimate, as one document prints it. The row records what the issuer says it is estimating, a definition class derived from the documents themselves, the amount or range exactly as printed with its qualifier, the denominator and share the issuer gives, the period, the method in the issuer's words, and the caveats the issuer states.

Each row carries the path of the archive's copy of the document, that copy's SHA-256 recomputed at build time, the hash the archive recorded when the file was retrieved, the document's public address where one is established, a pin to the place in the document, and the verbatim span that carries the figure.

definition-classes.csv gives the one-line definition of each of the 15 classes. comparability.csv states, for every pair of estimates, whether the two documents measure the same thing, over the same programs, for the same period, and therefore whether the figures may be added.

compiled_label was split out of issuer_term_verbatim before publication. On six rows the delivered term joined two of the document's phrases, elided a quotation, or shortened the issuer's wording; those six strings are now in compiled_label, and issuer_term_verbatim carries the document's own contiguous wording from the same span.

## How it was built

Amounts were copied from the held documents. amount_as_printed keeps the issuer's string, amount_qualifier keeps the hedge, and amount_usd_low and amount_usd_high hold the range in dollars, with amount_usd_high blank where the issuer printed a single figure.

Every verbatim span was re-found in the held document at packaging time: 22 of 22. Each of our copies was re-hashed at packaging time; 21 of the 22 rows sit on files whose hash the archive already records, and all 21 matched. The remaining row sits on the House Oversight ID.me staff release, which this archive holds as a text capture it made of the committees' published release rather than as a downloaded file, so no acquisition manifest records a hash for it; the hash in the row was computed from our copy and is proposed for the archive manifest.

The definition classes are derived from the documents, not imposed on them: each class is the set of rows whose issuers describe the same kind of quantity. The class of every row was re-checked against its own span at packaging time.

## Read this before you use it

Do not add the figures. The rows sit in 15 definition classes. Adding across classes produces a number no issuer published. comparability.csv states for each of its 210 pairs why: exactly one pair is additive without double counting.

Improper payment is not fraud. IMPROPER-PAYMENTS rows count payments that should not have been made; the category includes error and ineligibility as well as fraud, and the issuers say so.

Charged loss is not convicted loss. CHARGED-LOSS is loss alleged in filed charges.

Each row is fixed to its document's date. Later documents may restate a figure; where a document restates its own earlier figure, both rows are present.

Two columns, because a quotation and a label are not the same thing. issuer_term_verbatim holds only what the document prints for what its figure counts; it was re-found in our copy, contiguous, for all 22 rows. Where this table needed a shorter or joined phrase to sit in a column, that phrase is in compiled_label instead, filled on six rows and written by this archive, not by the issuer.

public_url is blank where no public address is established. Four rows rest on an SBA white paper the archive holds with no recorded source address; the gap is listed, not guessed.

## What it does not measure

No row is a finding, and the table draws no conclusion about a true amount. It records what each document says and how it says it.

It is not complete. It covers only estimates for which this archive holds the document. One estimate was dropped for failing verification, and the reason is recorded in the build receipt.

It is not the site's calculator. Build your own fraud number takes its own inputs; this table is the record of what the published estimates measure, and the two are not wired together.

The actions taken against borrowers are a separate table. Pandemic enforcement and suspension events, 2026 records suspensions, referrals and enforcement summaries; it holds no estimate of the size of fraud.

## Citation

Pandemic fraud estimates: what each published figure measures, compiled from documents dated 2022-09-21 to 2026-09-14.

## Rights

Compiled from U.S. government and congressional public records; attribution to Pandemic Darlings requested for the compiled table.