How this was produced
Every material statement on this site is a typed record with a source trail and a verification status. This page explains how a document becomes a claim, how the models were used, and where the method runs out.
Analysis version v0.4.1 · evidence current through 2026-08-10 · Changelog
How a record becomes a claim
Material statements are not free prose. Every one is filed as one of five kinds, and the kinds are never blended — an interpretation cannot quietly become a fact because it was repeated often enough. Plain-language summaries and orientation copy are written by hand, but each one names the claim rows it rests on and may not state more than they support.
- Documented findingA number or finding that appears in a document. Nobody drew a conclusion from it yet.60
- Patient-reported / historySomething the patient stated. Recorded as testimony, never promoted to a measurement.17
- InterpretationA reading of what one or more findings mean. This is where analysis begins.4
- Working hypothesisA candidate explanation. Never a diagnosis.8
- Research-question claimA gap written as a question for a licensed clinician to answer.21
Counts are the claims carried in analysis version v0.4.1. The machine-readable kind names are observed_fact, reported_history, interpretation, hypothesis, and research_question.
Then it gets a verification status
Separately from its kind, each claim records how close it sits to an original instrument record. This is the number that matters most, and it is the one the project scores worst on.
- 0Verified against primary instrument record
- 88Matches public specialty summary / pack transcription
- 1Discordant sources or specialty vs commercial conflict
- 21Not checked against a closer primary
- 0No adequate source
Public wording is additionally constrained by an audited language guide, so a claim cannot be restated more strongly on a page than its record supports. In publication mode the site ships only explicitly approved claims and release-scoped routes; preview mode shows the wider inventory with accurate statuses attached. You are reading publication mode.
How the models were used
Round 1 (2026-08-05) ran three model families independently on the same records, without sight of each other's output, then turned them against each other: each family was instructed to attack the others' work for overstatement, missing counterevidence, and provenance gaps. A synthesis chair reconciled what survived. A follow-on adversarial audit on 2026-08-06 repeated the exercise against the assembled site and produced the corrections logged in the changelog.
Round 2 (2026-08-10) added a fourth build, run solo: a blinded agent re-derived its own analysis from the evidence pack alone — no sight of this project's hypotheses, ranking, or literature cards — and a second agent then attacked both that blinded analysis and the published ranking with equal aggression. Because one model family cannot supply cross-family independence, the round substituted blinding and role asymmetry for it, and its findings were re-verified against the source documents before entering the record. Where the blinded derivation landed on a published position without having seen it, that is recorded as a re-derivation; where the two disagreed, the compiled source documents adjudicated, and two of the blinded analysis's own confidence grades were overturned in the process.
Why Fable 5 was absent from Round 1 is worth recording: at that time the build declined every task touching this case material, under the deliberately broad biology safeguards it launched with. Anthropic later narrowed those safeguards (“Improving Fable 5's biology safeguards”), after which the build could work on this material; its first contributions to this project are the 2026-08-10 roles listed below. The gap reflects the vendor's safety configuration at the time, not a judgment about this material, and is why the fourth build arrived as a solo round rather than as part of the original three-way review.
- ClaudeOpus 5AnthropicRound-1 swarm analysis, design package, final go/no-go audit
- ClaudeFable 5AnthropicRound-2 solo blinded re-derivation + adversarial cross-exam (2026-08-10); community-issue adjudication
- CodexGPT-5.6 SolOpenAIIndependent findings pass, dispositions, adversarial audit
- GrokGrok 4.5xAIPrelaunch remediation, adversarial audit remediation
Multi-model agreement is not clinical validation. Models share training data and failure modes, so they can be wrong together and confident together. No AI system is listed anywhere on this site as a clinician, investigator, or medical reviewer, and adversarial AI review does not substitute for patient or licensed clinician review.
These three model families account for every AI contribution to this portfolio. Exact build identity is not recoverable for all of it: the earliest rounds recorded the tool version rather than the served model build, and providers update builds behind a stable name.
Verbatim disclosure record for v0.4.1
- Grok (M4 / M4R / E2.x / Checkpoint F / F.1 / F.1.1 / G.1 / G.2 / G.2.1 / G.2.2 / H.1 / I / I.1 / I.1.1 / J.1 prelaunch remediation; 2026-08-06 adversarial audit remediation COR-0022…COR-0032)
- Claude (design package / Round-1 swarm / G audit / H final audit / J go-no-go; 2026-08-06 overnight adversarial multi-model audit)
- Codex (Checkpoint E / E2 / F dispositions / G.2 independent findings; 2026-08-06 overnight adversarial multi-model audit; 2026-08-07 paired pre-publication review)
- Grok (2026-08-07 paired pre-publication review)
- Claude (2026-08-07 plain-language landing rewrite, ruled-out register, review triage and remediation)
- Grok (2026-08-09 overnight research run over the compiled source documents; findings register and unknowns)
- Claude (2026-08-10 audit of that run, source re-verification, DEC-0039 scope decision, v0.4.0 promotion)
- Codex (2026-08-10 paired pre-publication review of the v0.4.0 content and language gates, three rounds)
- Grok (2026-08-10 paired pre-publication review of the v0.4.0 content and language gates, three rounds)
- Claude Fable (2026-08-10 adjudication of community issue
- Codex (2026-08-10 paired review of the issue
- Grok (2026-08-10 paired review of the issue
- Claude Fable (2026-08-10 Round-2 solo blinded round — blinded independent re-derivation from the evidence pack, adversarial cross-examination of the blinded output and the published ranking, chair synthesis re-verified against the compiled source documents; CLM-0112/UQ-0020 aromatase-inhibitor exposure wiring; COR-0043, COR-0044; layered into v0.4.0 pre-packaging)
- Codex (2026-08-10 paired pre-packaging review of the Round-2 layering, two rounds; caught the lit-0308 polarity misfiling and the launch-citation attestation gap)
- Grok (2026-08-10 paired pre-packaging review of the Round-2 layering, two rounds)
- Claude Fable (2026-08-11 orchestration of the unified questions surface; Claude Opus design survey and implementation of /questions/, the extended public-language gating, and the fail-closed cross-links)
- Codex (2026-08-11 paired review of the questions surface; caught the partial-approval dangling-link class)
- Grok (2026-08-11 paired review of the questions surface; caught the determiner-grammar defect in drug-name generalisation)
Internal milestone codes are retained here unedited so the disclosure is not smoothed over in summary.
What this process cannot do
- Cannot diagnose or prescribeLicensed clinicians interpret records and decide care. Nothing here is an instruction.
- Cannot read a primary recordNo original laboratory report, DXA printout, radiology report, or clinic note has been reviewed by this project. 0 of 110 claims are verified against an original instrument record.
- Cannot invent missing recordsWhen a primary instrument report is absent, the question stays open rather than being filled in.
- Cannot upgrade evidence by consensusAgreement among models does not change a verification status. Three models liking a claim is not a source.
- Cannot publish excluded filesLocal papers, raw community dumps, and patient source PDFs never reach the build artifact.
Sources and literature
Targeted specialty search passes used a research cutoff of 2026-08-05. The catalog holds 339 cards covering 329 unique works. Cards may be abstract-only; access and license fields are recorded on each one, and duplicate DOI/PMID records are aliased rather than deleted so a citation never silently disappears.
Patient source documents live in a Google Drive folder under Drift0r's own control — not this project's. This site describes and cites those materials but never rehosts them: no original patient PDF is shipped in the repository or the static build.
- Externally controlledContents may change, be restricted, or be removed at any time. This site does not mirror them, and access is Drift0r's to grant or revoke.
- Treat as private medical informationThe folder may include sensitive records, including mental-health specialty summaries.
Patient-hosted source folder (external Google Drive)
Drift0r granted permission to publish this research portfolio. Permission is not endorsement of any finding or of the folder's contents, and clinician review has not been performed. Full consent and privacy terms are on the legal page.
Where the rest lives
- AboutWho built this, and the independence statement — not a clinician, not affiliated with Drift0r.
- LegalMedical disclaimer, consent and privacy, copyright, and withdrawal of consent.
- How this could be wrongSeven named failure modes, with published errata.
- ChangelogEvery material change by version, including 44 logged corrections.
- How to contributeCorrection, privacy, and removal requests. Do not post private medical records.