# LookCloser Methodology
**Version 0.1 — draft, 2026-07-18** · This document is versioned; every verdict cites the version applied. All changes to this document are recorded in the public ledger.
## 0. What we do and don't do
We verify **integrity, not efficacy**. We check whether a product's claims are consistent with the public record — label arithmetic, certifications, sourcing, manufacturing history, review behavior, and the scientific literature as it actually reads. We publish **evidence-tiered statements of documented fact** with citations. We do not give medical advice, we do not say a supplement "works," and we do not resolve scientific disagreements — we display them with their parameters and let the reader (human or AI agent) draw conclusions for their own context.
Findings can always be challenged through the correction process (§7) — with documents, not arguments.
## 1. Evidence tiers (product integrity)
Every factual statement in a product record carries a tier:
| Tier | Name | Nature | Example |
|---|---|---|---|
| 0 | Arithmetic | Label vs. mathematics/physics; conclusive on its face | Claimed 5,000mg of an ingredient in a 3,500mg serving |
| 1 | Chemistry logic | Form-stability and formulation plausibility | Creatine's known degradation to creatinine in aqueous (gummy) matrices |
| 2 | Documentary | Government and registry records | FDA warning letter to the product's manufacturer; certification claim absent from the issuer's registry |
| 3 | Statistical | Anomaly analysis of public behavioral data | Review-count increase of +214 in 48h with no corresponding sales-rank movement |
| 4 | Assay | Physical laboratory results | Published third-party HPLC test of actual product content |
Tier is displayed with every statement. Higher tiers are not "more true" — a Tier-0 impossibility is conclusive — but tiers tell the reader **what kind of checking** produced the statement.
## 2. Literature protocol (science layer)
Our evidence grading follows the public standards of evidence-based medicine — **GRADE** (certainty rating), **Cochrane Risk of Bias 2** (study-quality domains), and the **Oxford CEBM hierarchy** (design levels) — applied by a machine pipeline using an explicit, reproducible rubric.
**Search protocol (PRISMA-inspired, logged per ingredient):**
1. PubMed E-utilities queries per ingredient × outcome, filtered by publication type (meta-analysis, systematic review, randomized controlled trial), query strings logged and versioned.
2. ClinicalTrials.gov v2 for registered trials (including unpublished/terminated — publication-bias signal).
3. Citation-graph ranking (OpenAlex) to surface the studies the field itself cites most.
4. NIH ODS fact sheets as the government baseline reference.
**Per-study structured record:** design · n · duration · dose · form/delivery route · population · primary outcome · effect size (not just significance) · **funder** · **declared conflicts of interest** · author affiliations · risk-of-bias flags (randomization, blinding, attrition, selective reporting) · position relative to the claim as worded (supports / contradicts / inconclusive).
## 3. Evidence grades (per ingredient × outcome × dose-range × route)
| Grade | Meaning |
|---|---|
| **A** | Meta-analyses of human RCTs, consistent direction, adequate total n |
| **B** | Multiple independent human RCTs, largely consistent |
| **C** | Single or small human RCTs; or consistent observational only |
| **D** | Animal / in-vitro only — no human interventional evidence located |
| **X** | Higher-tier human evidence contradicts the claim |
**GRADE-style certainty modifiers (applied and displayed):**
- ↓ high risk of bias across supporting studies (Cochrane RoB domains)
- ↓ inconsistency between studies without explanation
- ↓ imprecision (small samples, wide confidence intervals)
- ↓ **all supporting studies industry-funded with no independent replication** (funding does not invalidate a study; unreplicated industry-only support lowers certainty and is always disclosed)
- ↑ dose-response gradient across studies
- ↑ large, consistent effect sizes
## 4. Claim matching rules
A product claim is evaluated against the literature on **four keys — never ingredient alone**:
1. **Ingredient** (specific form — e.g., magnesium glycinate ≠ magnesium oxide)
2. **Dose** — product dose vs. the dose range used in supporting studies, and vs. official RDA/Tolerable Upper Intake Levels; absorption-saturation points where established
3. **Delivery route** — was the outcome studied under the mechanism this product uses? (e.g., dental ingredients studied under prolonged brushing contact do not support a swallowed-powder claim)
4. **Outcome as worded** — the claim's own words, not a generous paraphrase
**Conflict rule:** when studies disagree, the record shows the disagreement with each study's parameters and funding — e.g., *"2 supporting trials (manufacturer-funded, 8 weeks, n<60, modest effect); 1 independent trial, longer, larger, null."* We never collapse conflicts into a single grade silently.
## 5. Product integrity checks (summary)
Product records draw on: label arithmetic (§1 Tier 0) · certification claims verified against issuing registries · FDA enforcement, recall, import-refusal and inspection records · manufacturer disclosure class ("manufactured by" / "manufactured for" / "distributed by") and facility trails including public customs records · trademark and corporate-registry history · review forensics (temporal, distributional, reviewer-graph, content, and solicitation signals — 17 documented checks) · practitioner-endorsement verification (licenses, disciplinary records, publication history, disclosed and undisclosed material connections).
Every check that runs is listed in the record — including the ones the product **passed**.
## 6. Scores
The summary score is a **rendering of the evidence for human convenience, not a judgment layer**. It is computed from published component weights, displayed with its methodology version, and always expandable to the complete fact list. Products with insufficient evidence show **INSUFFICIENT — here is what's missing** rather than a number. Severity classes are explicit: fraud-tier findings (arithmetic impossibility, unverifiable certification) cap the score regardless of other components; opacity-tier findings (undisclosed manufacturer — an industry-wide norm) are weighted as opacity, not fraud. Headline framing is relative (percentile within category), because absolute purity is rare in this market and a wall of identical failing grades informs no one.
AI agents receive the underlying structured record and may ignore the score entirely.
## 7. Corrections
Any brand or reader may challenge any statement by submitting evidence: documents, accredited-lab results, registry records. Challenges are evaluated under this same methodology, and **both the original finding and the correction outcome remain permanently on the ledger**. We do not accept payment to alter findings; document-authentication services purchase depth of record, never direction of conclusion.
## 8. Ledger
Every verdict, correction, and methodology version is hashed into an append-only log with externally published roots. History cannot be silently rewritten — by anyone, including us. Verification instructions are published with the roots.
## 9. Disclosures
- No advertising, no affiliate links, no paid placement, no sponsored rankings. Revenue: API access and flat-fee document authentication.
- The founder operates e-commerce brands including pet supplements. **LookCloser does not cover pet products while that interest exists**; this recusal and its trigger are permanent parts of this document.
- This site collects no user accounts for lookups, sets no trackers, and retains no personal query profiles.
## Version log
- **0.1 (2026-07-18):** Initial draft. Grading foundations: GRADE, Cochrane RoB 2, Oxford CEBM, PRISMA-inspired search. Extensions: funding/COI weighting and disclosure, dose/route/form claim-matching, conflict-display rule, product-integrity tier system, severity classes, correction process.
