How human evidence is assessed

• Research volume is not evidence strength. Publications, studies, registrations, outcomes and evidence syntheses are counted as different things.
• A registered trial is not evidence of benefit, and a completed trial without published results is not a negative result.
• Claims are things to investigate. Evidence for one outcome, population or route is never carried over to another.
• Safety and effectiveness are assessed separately. No reported adverse events is not evidence of safety.
• Too little evidence is not evidence that something doesn’t work. Missing information is never filled in or treated as a negative finding.
• Animal and laboratory findings are never counted as human evidence. Related peptides, analogs, fragments and combination products are never treated as the compound itself.
• Assessments are automated by AI under these rules and published labelled as automated. A person can review any of them; a reviewed version replaces the automated one and says so.

Records come from the Research Intelligence Center’s index: PubMed records through Europe PMC (metadata and, for extraction only, abstracts that are read but never stored), ClinicalTrials.gov registrations, Crossref, OpenAlex and NCBI checks. The exact search terms and queries per compound, with their dates, are part of each research snapshot; each assessment records the snapshot it was made against and its own evidence cutoff date. Reports outside PubMed, such as meeting abstracts, are added by reviewers with a citation.

New human records of the compounds under assessment are found by rules every day: a PubMed record the research classification calls human clinical, or a registration focused on the compound, that no study record describes yet is filed for screening.

A publication is a bibliographic record; a study is one investigation, which may have several reports (the main paper, follow-ups, secondary analyses, conference abstracts, registry entries and posted results); a registration may or may not have results; an outcome is one result measured in a study; an evidence synthesis answers one question from several studies.

Every report is linked to exactly one study (only a pooled analysis may describe several), and participants are recorded once, on the study. Planned enrollment from a registry is kept apart from actual enrollment, randomization, treatment and completion counts. A registration that lists a paper as its results is filed with that paper’s study.

Each record is first screened as a human study of the compound (or excluded with a reason, such as no human participants, not a primary study, another compound, or a review covering several compounds). Each included study is then judged against each claim’s question:

• PICO: the population, intervention, comparator and outcome the claim is about, with its clinical context and follow-up.
• Eligibility: included or excluded for that question, with the reason (wrong population, intervention, comparator or outcome; an analog or combination; not human; not a primary study; a duplicate report; insufficient data; retracted).
• Relevance: direct, indirect (and in which part of the question) or not relevant, always with a justification. A short tolerability study in healthy volunteers can’t show faster tendon healing; a pain score can’t show tissue repair.

Every exclusion decision and its reason is published with the assessment.

Risk of bias is judged for a specific result (one outcome of one study), with the tool that fits the study’s design. Every domain judgment carries a written justification.

ROBINS-I is never applied to uncontrolled studies, case series or case reports: they have no comparison group to judge. Those get the structured appraisal, which lists their limitations without pretending to rate them like a comparative study. A RoB 2 overall judgment is never milder than its worst domain.

Certainty is rated for one outcome of one question, across all eligible studies, following the GRADE handbook. Randomized evidence starts at high certainty and nonrandomized evidence at low; reviewers rate down for risk of bias, inconsistency, indirectness, imprecision, publication bias, and may rate up for a large effect, a dose-response gradient or confounding that would reduce the effect, each with a written reason.

• High certainty: The true effect is likely close to the estimate.
• Moderate certainty: The true effect is probably close to the estimate, but may be substantially different.
• Low certainty: The true effect may be substantially different from the estimate.
• Very low certainty: The true effect is likely to be substantially different from the estimate.

Some outcomes can’t support a defensible rating. They are shown with an operational state, which is not a GRADE category:

• Not yet assessed: No reviewed assessment has been published for this outcome. This says nothing about whether it works.
• Assessment in progress: Reviewers are working on this assessment; nothing is concluded yet.
• Insufficient information to assess: The available human research cannot support a defensible certainty rating. This is not evidence that it does not work.

Each assessed outcome also states what the evidence suggests (an effect, little or no effect, conflicting evidence, or not enough to tell), what the research actually measured and what remains unknown. A statistically significant result alone never makes a claim “proven”, and the absence of adequate studies never makes it “disproven”. Certainty categories are counted, never averaged or turned into a score, and no peptide-wide grade exists.

Each human study is appraised by two AI models (Anthropic’s Claude Opus and Claude Sonnet) independently, from its title, identifiers, PubMed publication types and topics, registry fields and abstract. Each model screens it, extracts the design, participants, groups, outcomes, reported results, adverse events and limitations, judges risk of bias with the tool its design calls for, and decides for every claim question whether the study is eligible and how directly it applies.

The two answers are combined cautiously: a study is included only if both models include it; the less rigorous design and the more severe risk-of-bias judgment are kept; participant counts are kept only when both agree; a claim question counts a study only if both say so. Certainty is then rated with GRADE for each question by the two models again, from the study records alone: the lower certainty and the more serious judgment per domain are kept, and if they disagree on the conclusion it is shown as “not enough evidence to tell”. A question no eligible study answers is marked “insufficient information” by rule, without AI.

Safeguards: the record is given to the models as data and they are told to ignore any instructions in it; they may cite only the study’s own reports; numbers are never estimated, and a missing value stays missing; no dose is stored or shown; every judgment needs a written reason; the database refuses anything that breaks the rules (a tool that doesn’t fit the design, a rating milder than its worst domain, a downgrade without a reason). Model and prompt versions are published with every record.

Automated assessments have not been reviewed by a person and can contain errors. They are labelled “Automated” on every page, in the API and in the data files. A qualified reviewer can review any record in Peptide Library HQ; the reviewed version then replaces it and is labelled as reviewed.

Automated records publish directly. A person’s review moves a record through: awaiting scientific review, changes requested or approved, published. Nobody reviews what they prepared. Study records and claim assessments need one reviewer’s approval; risk-of-bias and certainty assessments need two independent reviewers, and a disagreement is settled by a third, adjudicating reviewer. Reviewers declare their conflicts of interest before reviewing, and that statement is stored with every decision. A reviewer’s name and disclosed qualifications appear on an assessment only with their consent; Peptide Library never invents a reviewer, credential or affiliation.

These assessments are not peer-reviewed journal articles and are not presented as systematic reviews unless the full screening record supports that description.

Peptide Library also lists vendors, prices and partner programs. None of that can influence an assessment: no vendor, advertiser or partner can buy, request or change a judgment, inclusion criterion or grade, unfavorable assessments are never hidden, and the assessment system has no access to commercial data (this is checked automatically). Reviewers disclose any commercial ties.

• Research activity gap: Little indexed research of any kind.
• Human research gap: Laboratory or animal research exists, but little human research.
• Evidence quality gap: Human research exists, but no assessed outcome reaches moderate or high certainty.
• Clinical relevance gap: Studies exist but none directly address a claim being evaluated.
• Safety evidence gap: Safety is not established with moderate or high certainty for the populations, routes or durations studied.

Gaps overlap, and a gap describes missing research, never ineffectiveness. Quality, relevance and safety gaps are shown as not assessed until a published assessment exists.

Every published assessment is a frozen version with its methodology and tool versions, its evidence cutoff, the research snapshot it was made against, its included and excluded studies, its reviewers’ decisions and the date. Updates publish a new version with a note on what changed and why; earlier versions stay readable. A withdrawn version keeps its reason.

Every day, monitoring checks for new human studies, posted trial results, registration status changes, retractions, expressions of concern and corrections. It marks the affected assessments as needing an update and opens a reviewer task; it never changes a published conclusion by itself. Historical research snapshots, including edition 2026.1, are never recalculated.

Data: assessments and study records (CSV and JSON, CC BY 4.0, published metadata only). Corrections: support@peptidelibrary.app.

• Full texts are read only by reviewers; AI extraction uses abstracts and registry data, so it can miss what only the full paper reports.
• Searches cover PubMed and ClinicalTrials.gov; studies published elsewhere are found only when a reviewer adds them.
• The research on most peptides in humans is small. Many outcomes will be “insufficient information” or low certainty for a long time.