Frameworks.

The instruments

The Standard makes a claim: every analysis on this site runs through named frameworks and instruments with published criteria. This page is where that claim is kept. Four instruments do the forensic work — testing research, extracting claims, auditing citation networks, evaluating manuscripts — and two broader frameworks supply the thinking beneath them: a lens that asks who an AI serves, and a literacy that defines how a patient judges what an AI says.

01 · The instruments

REAL-PAIR Rating Ecological Authenticity Levels in Patient-AI Research

Did this study of patients using AI involve patients using AI?

Rates research across five dimensions: participant authenticity (real patients or paid proxies), question authenticity (self-generated concerns or researcher prompts), contextual contribution, stakes and motivation, and ecological validity. It exists because the dominant study designs fail most of these while drawing conclusions about "patients." REAL-PAIR makes that failure measurable instead of arguable.

ASSAY Structured claim extraction & integrity assessment

What does this paper demonstrate, as opposed to assert?

Takes any article, preprint, review, or meta-analysis and extracts its claims one by one, grading each for the evidence behind it. The output separates what the data shows, what the authors conclude, and what the abstract and press release announce. Where those diverge, ASSAY says so, claim by claim.

CLAIM-CSN Claim-Specific Citation Network audit

Is this claim's authority supported by data, or manufactured by citation dynamics?

Audits the citation network behind a single claim, in the tradition of Greenberg's citation-network analyses. It detects the mechanisms that turn weak findings into apparent consensus: citation bias, amplification, diversion, transmutation, dead-end citation, and back-door invention. The verdict states whether a claim stands on evidence or on an information cascade.

PEER-REV Structured manuscript evaluation

Is this paper sound, and is it useful to patients?

Evaluates papers for scientific integrity, statistical validity, and patient utility. It runs a structured red-flag audit, applies GRADE-style certainty rating, and calibrates its standards to what kind of claim a paper makes: an early signal held to a plausibility standard, an evidence synthesis to a systematic standard, a policy prescription to a causal one. Punishing an early signal for not being a meta-analysis, or excusing a policy prescription because the field is young, are both errors. PEER-REV is built to avoid each.

05 · The lens

CAIHL Critical AI Health Literacy

Who does this AI serve?

From Hugo Campos and Liz Salmi's National Academy of Medicine commentary, applying Paulo Freire's theory of critical literacy to health AI. Its founding move is a distinction most discussion collapses: institutional AI, deployed to serve organizational priorities, versus patient-directed AI, chosen and controlled by individuals. From there it evaluates any tool along four dimensions (credited to Vadim Dukhanin): who the primary user is, where it is hosted, whose interests it actually advances, and whether it expands or constrains the patient's agency. That last dimension is the one this site cares about most, and the reason "patient-directed, AI-assisted" sits under our name.

06 · The literacy

CLAIM Contextual Literacy for AI in Medicine

What must a patient be able to do to judge an AI's output?

Five pillars: Contextual Grounding (anchoring the answer in the patient's situation and data), Interrogative Stance (questioning rather than accepting), Associative Integration (connecting across sources and records), Judgment Activation (deliberate evaluation before any action), and Methodological Transfer (carrying the habits to the next question). CLAIM is the discipline taught throughout this site. Start here is CLAIM at its simplest; the Practice, in preparation, is CLAIM in full. Its reference implementation also exists as an open-source server exposing each pillar as a tool.

07 · How they appear in the work

Several, in sequence.

A typical analysis might run several in sequence: REAL-PAIR to test whether a study's patients were patients, ASSAY to extract and grade what it claims, CLAIM-CSN to check whether the claims it inherits are sound — all under CAIHL's prior question of whose interests the work serves. Published analyses name their instruments, and every verdict can be traced back to the criteria on this page. They live, assessed and dated, in the Casebook.

These methods are open. They are meant to be used, tested, and improved by anyone, especially patients. If you apply one and find where it fails, that is not an attack on the work. That is the work.