# COVID-19 Claims Evidence Review — R21 protocol

Protocol version: 1.1
Site revision: R21
Protocol adopted: 2026-07-31  
Evidence cutoff for the current review: 2026-08-06

## Status and scope

This is a structured, claim-by-claim evidence review. It is not a registered systematic review or meta-analysis. The protocol applies prospectively from R5; earlier revisions did not preserve a complete candidate-source log, duplicate count, or exclusion count.

The unit of analysis is one stable claim ID. Intervention questions use population, intervention, comparator, outcome, timing, and setting fields where applicable. Policy, mortality, origins, funding, and personal-conduct questions use adapted population or record set, exposure or hypothesis, comparison, outcome, evidentiary chain, and missing-evidence fields.

## Information sources

Search and verification routes include:

- Publisher article pages and supplements
- PubMed and PubMed Central
- WHO, CDC, FDA, NIH, NIOSH, and other public-agency records
- Inspector-general audits and grant or oversight records
- Declassified intelligence assessments and official document releases
- Reference lists and citation chains from included primary sources
- Direct capture of user-supplied social-media posts, platform captions, and media metadata for claim provenance

News reports, press releases, congressional findings, and intelligence judgments may document institutional positions or unavailable records. They are labeled by evidence class and are not substituted for randomized clinical, epidemiological, genomic, or other primary evidence.

### Source-balance rule

Institutional guidance—including WHO, CDC, FDA, and NIH—is contextual evidence. It does not substitute for outcome data when randomized trials or directly relevant primary studies are available, and no clinical verdict may rest on one institution's recommendation alone.

For each material claim, the review searches for the strongest contrary result located as well as evidence supporting the proposed verdict. Null trials, favorable reviews, methodological critiques, minority analyses, integrity notices, and relevant working papers are retained when they materially affect interpretation. Conflicts are resolved by study design, directness to the tested claim, risk of bias, integrity, precision, and reproducibility—not by institutional authority or viewpoint. Sources with materially different designs or estimands are not treated as interchangeable.

## Reproducible search blocks

Each refresh records the date, service searched, exact query, candidate URL, disposition, and reason. Baseline query blocks are combined with relevant intervention, population, outcome, and date terms.

### Policies and transmission

`(SARS-CoV-2 OR COVID-19) AND (distance OR distancing OR aerosol OR airborne) AND (transmission OR outbreak)`

`(COVID-19 OR respiratory virus) AND (mask OR respirator OR N95 OR cloth mask) AND (filtration OR fit OR infection OR transmission)`

`COVID-19 AND (restriction OR school closure OR stay-at-home OR non-pharmaceutical intervention) AND (transmission OR education OR mental health OR delayed care)`

### Vaccines and safety

`COVID-19 vaccine AND (hospitalization OR critical illness OR death OR effectiveness) AND (age OR risk OR variant OR time since dose)`

`mRNA COVID-19 vaccine AND (myocarditis OR pericarditis OR anaphylaxis) AND (incidence OR surveillance OR cohort)`

### Treatments

`COVID-19 AND (ivermectin OR hydroxychloroquine OR remdesivir) AND (randomized OR placebo OR platform trial) AND (mortality OR hospitalization OR recovery)`

### Fatality

`COVID-19 AND (infection fatality OR case fatality OR excess mortality) AND (age OR country OR global)`

### Origins and funding

`SARS-CoV-2 origin AND (zoonotic OR market OR animal OR phylogenetic OR laboratory OR research-related OR intelligence assessment)`

`(NIH OR NIAID OR EcoHealth OR Wuhan Institute of Virology) AND (grant OR subaward OR audit OR oversight OR financial disclosure)`

## Eligibility criteria

Include a record when it:

1. Directly bears on the exact tested claim or a stated evidence gap.
2. Provides primary data, a reproducible synthesis, official guidance, an audit, a contemporaneous record, or a clearly labeled institutional assessment.
3. Identifies a population or record set, method, outcome, and material limitations sufficiently for interpretation.
4. Is the most complete accessible version when duplicate reports describe the same analysis.

Exclude or do not rely on a record when it:

1. Merely repeats another source without adding evidence.
2. Addresses a different virus, drug, dose, population, outcome, or setting without a clearly stated indirectness limitation.
3. Makes a causal claim unsupported by the underlying record.
4. Is inaccessible enough that the relevant result cannot be checked.
5. Has been retracted or superseded, except when retained to explain the history of a claim and labeled accordingly.

## Certainty and verdicts

Evidence certainty and claim verdict are different judgments.

- **High certainty:** the true direction or conclusion is very unlikely to be materially different.
- **Moderate certainty:** the conclusion is probably correct, but important change remains possible.
- **Low certainty:** confidence is limited and new evidence could materially change the conclusion.
- **Very low certainty:** the available evidence cannot support more than a tentative conclusion.

The certainty assessment is GRADE-informed. Risk of bias, inconsistency, indirectness, imprecision, and publication concerns can lower certainty. Large effects, dose-response patterns, or residual confounding that would reduce an observed association may strengthen some observational conclusions. Because this is not a registered systematic review, the labels must not be represented as formal GRADE determinations.

Verdicts describe the tested wording:

- True
- Mostly true
- Mostly false
- Misleading
- False
- Unresolved

## Quantitative reporting

Where an included source reports a suitable estimate, the review records the treated and comparison values, effect measure, uncertainty interval, population, timing, setting, and source ID. Relative effects are not presented without the underlying event rates when those rates are available. Estimates from different populations or eras are not pooled informally.

## Source provenance

The R21 catalog preserves 118 stable source IDs, titles, evidence types, and canonical URLs. R21 adds the separate 41-page CBP production as an official-source-only record and publishes a public-safe page/topic and duplicate index for the 1,375-page Gates collection. Both PDFs remain unmirrored after privacy review. R20 added five Senate records to the catalog and hosted two publisher-current files that passed the publication screen unchanged: NIH public-comment moderation records and Fauci and intelligence-community correspondence. The complete Slack production, its committee-selected excerpt set, and the current Daszak CBP release remain canonical-source links only after the fuller files exposed access or operational contact details. R19 added a prominent newest-first archive spotlight without changing the underlying research record. R18 clarified the provenance of publisher-applied redactions and replaced potentially agent-specific ethics language with an explicit non-determination; sources and files were unchanged. R17 added five hosted Senate HSGAC PDFs, the official July 29 hearing record and full video, and one cataloged-but-deferred collection. R16 added two NOTUS investigations and Internet Archive revision metadata for the withdrawn pandemic-diary files. R15 added the complete major 2021 Fauci FOIA email file, the two-part Fauci House transcript, the House final report, and all eight supporting-document volumes. Composite records may also expose stable subclaim IDs, legal or evidentiary status labels, and evidence-balance groupings. DOI and PubMed identifiers are retained in canonical URLs where available. The catalog snapshot date is 2026-08-06. Per-source access dates were not retained before R5; future checks are dated in the candidate-source log.

On 2026-08-06, an automated point-in-time sweep checked all 118 canonical URLs. A failed automated check is not labeled a dead source. DOI, PubMed, and known-notice metadata checks retained one expression of concern (source 58). No-notice-found is not proof of permanent integrity, and these metadata checks do not replace publisher-page review. Source-level access and Wayback results are distributed as `/source-verification.json` and `/source-verification.csv`.

The Fauci diary PDFs are treated as primary records of what was written, not as self-proving scientific, causal, intent, or legal findings. Interpretive claims are checked against complete dated entries, contemporaneous public statements, and independent evidence where available. The site hosts the publisher-current official files, retains the Senate source URLs, and publishes byte sizes and SHA-256 checksums in `/fauci-diaries/manifest.json`. When the pandemic diary was revised after NOTUS reported that earlier files exposed third-party medical information, the site removed its superseded copy. The manifest preserves only non-downloadable predecessor metadata—timestamps, sizes, and archive digests—and does not mirror or directly link the withdrawn file bodies. Publisher-current status is not treated as an independent guarantee that all sensitive information has been removed.

The FOIA email, House transcript, final report, and supporting-document PDFs are likewise treated as primary documentary records rather than self-proving conclusions. Release dates were checked against the originating publication pages: the major newsroom FOIA publications were in June 2021, the January 2022 House release presented excerpts rather than a new 3,000-page tranche, and the Select Subcommittee's final supporting collection was published in December 2024. Complete hosted-file metadata and SHA-256 checksums are distributed in `/primary-records/manifest.json`.

The 2026 Senate investigation records receive the same treatment. R17 hosts Investigation Records — Volume I, the Baric interview and awards correspondence whose Senate filenames label them redacted, and both July 29 hearing opening statements; it links the complete official Senate video. R20 hosts the NIH moderation and litigation records and the Fauci and intelligence-community correspondence. It catalogs but does not mirror the Slack or current Daszak CBP files. R21 completes a page/topic and duplicate index for the broad 1,375-page Gates collection but keeps the PDF official-source-only because the publisher-current file includes a date-of-birth identifier and operational contacts. It also catalogs but does not mirror the separate 41-page CBP production after every page was visually reviewed and the file was found to contain direct contacts and sensitive immigration, family, demographic, employment, financial, and law-enforcement records. The originating FOUO, LES, and USPER labels are described as source handling markings; they do not by themselves establish that Senate publication was unlawful. The site did not add, remove, or choose any blackout and hosts only byte-identical publisher-current source copies. When a current source still exposed details the site would not republish, the record was deferred rather than edited. Committee summaries and member statements remain attributed. Private messages establish what participants discussed, including uncertainty and disagreement; they do not independently resolve origin or establish conspiracy, concealment, misconduct, or legal intent. The NIH records document moderation and legal deliberations but do not adjudicate constitutionality or legality. The CBP records do not independently establish scientific causation, intent, wrongdoing, ethics violations, or legal liability. The archive makes no independent determination about compliance with ethics, disclosure, conflict-of-interest, or other legal requirements. File hashes, extractable-text coverage, duplicate results, topic ranges, and privacy-screening status are distributed in `/senate-records/index.json`; the Slack and Gates productions have public-safe indexes at `/senate-records/slack-page-index.json` and `/senate-records/gates-page-index.json`. The non-modifying site screen is quality control, not a redaction process, privacy certification, legal review, or ethics determination; keyword searches may also miss image-only text.

Public availability is not treated as automatic permission to mirror a file. R20 quarantines the withdrawn 65,336,851-byte diary, a superseded 5,070,880-byte Daszak CBP revision, and every text, OCR, corpus, screenshot, or other derivative made from either body. Only non-downloadable hashes and revision metadata are retained. The publisher-current Daszak CBP file blacks out identifiers exposed in the retired copy but remains deferred because separate operational contact details are still visible. The complete Slack production remains deferred because access-bearing URLs are visible; the excerpt file is deferred with it to avoid selection without the fuller context. R21 retains the Gates and separate 41-page CBP bodies only at their official publisher links and publishes no sensitive values or document-text derivative. Hosted R20 records disable unchecked publisher-source fallback so a subsequently changed upstream body cannot be served under a previously verified manifest entry. Publisher-current status remains necessary but is not treated as a guarantee that no sensitive information remains.

The nine supplied media records have SHA-256 provenance receipts covering the captured file, canonical post URL, post text, available metadata, and caption-file hashes. Platform captions were available for seven of eight videos; the unavailable caption is disclosed rather than reconstructed. The receipts are distributed as `/media-provenance.json`; captured media itself is not republished by the site.

## Update and correction process

A new candidate source is logged with its query, discovery date, claim ID, and disposition. Material changes receive a new site revision and, when they change a claim record, a new claim revision. Contradictory evidence is retained and described. Correction requests follow the public correction template distributed with the site.

The current candidate and disposition record is distributed as `/candidate-source-log.csv`. Social-media posts are primary records of what was posted or said, not independent proof of the scientific, legal, or causal claims they contain. Edited excerpts are checked against captions and available full-context records; missing context is disclosed rather than reconstructed. Allegations of intent, censorship, records crimes, perjury, genocide, or other legal liability are labeled as allegations unless a cited adjudicative record establishes otherwise.