Evidence method
Every useful claim should leave a trail.
Our method keeps product facts, reported experience, inferred consensus, and merchant offers separate—then shows the strength and limits of each.
- Version
- 1.0
- Last updated
- July 29, 2026
The short version
- Resolve the exact edition.
We compare identifiers, region, piece count, year, factory or line, material, dimensions, and packaging evidence before joining records.
- Store each claim with provenance.
Facts and observations retain a source, date, evidence class, edition-match status, confidence, and review state.
- Group genuinely independent experience reports.
Syndicated or copied reviews count once. Manufacturer claims do not become consumer experience simply because they are repeated.
- Publish agreement and disagreement.
We summarize usable facets, show how many reports and independent sources contributed, and preserve material contradictions.
- Attach offers without merging them into the product.
Merchant price, availability, region, match confidence, and observation time stay in separate, refreshable records.
Evidence classes
A label describes what a source can support. It is not a decorative badge and it does not make every claim in that source equally reliable.
| Class | What it can support | Important limit |
|---|---|---|
| Manufacturer-specified | Identity, materials, dimensions, piece count, maker claims | Does not independently establish subjective experience |
| Merchant-specified | Offer identity, price, stock, region, listed specifications | May repeat maker copy and can become stale |
| Reviewer-reported | Use experience for an exact or likely exact edition | Must retain attribution, match confidence, and incentives |
| External aggregate | A partner’s aggregate or metric | The external method and coverage must remain visible |
| Consensus-derived | A transparent synthesis of traceable observations | Cannot be stronger than its source set |
| Model-inferred | Taxonomy mapping, extraction, or match suggestions | Must be labelled and quality-controlled |
| Unknown or mixed | Signals that remain sparse, ambiguous, or contradictory | Must not be forced into a confident verdict |
Exact-edition identity
The canonical record is independent of merchant listings. We prefer GTIN, UPC or EAN, maker SKU, piece count, region, release period, and packaging evidence. Identical artwork is not enough to merge two physical editions.
A probable match can be retained for internal review, but it must not silently supply exact-edition experience claims. Where the evidence suggests a different cut, factory, board, finish, region, year, or piece count, the records stay separate.
Experience facets and consensus
The first public taxonomy is limited to things a shopper can use: reported fit and looseness, false-fit tendency, cut style, finish and glare, dust, piece thickness and feel, image clarity, difficulty, completed dimensions, handling or accessibility, and board fit.
We store structured claims rather than copies of reviews. Public summaries should link back to the original source and use only a short, compliant excerpt when one is necessary. We never synthesize a quotation.
A page may have exact-edition written reports but still lack enough independent, same-direction observations to support a fit, feel, or difficulty conclusion. In that case we publish the coverage count and keep the direction unknown.
Confidence bands
The following are provisional publication rules, not scientific truths. Thresholds will be versioned and checked against real review sets before public use.
| Band | Provisional rule |
|---|---|
| Insufficient | Fewer than three genuine experience reports or fewer than two independent domains |
| Low | Three or more reports across two domains, with weak matching, independence, recency, or agreement |
| Moderate | Five or more reports across three independent sources, mostly exact matches, without severe unresolved contradiction |
| High | Ten or more reports across four independent sources, strong matching, useful recency, and stable agreement |
A confidence band never replaces the underlying distribution. Report counts, source counts, match quality, dates, and material disagreement remain visible.
Independence, duplication, and disagreement
- Syndicated reviews count once, not once per storefront.
- Near-identical text or the same author, date, and product combination is grouped.
- Several pages that repeat one underlying discussion count as one source lineage.
- Brand-level experience does not automatically transfer to every factory, piece count, or edition.
- Difficulty remains descriptive until repeated-solver data supports a stronger adjustment.
- Minority reports stay visible when they could change a shopper’s decision.
AI assistance
AI may suggest entity matches, extract facets into a controlled taxonomy, identify duplicates, surface contradictions, draft a consensus summary, or flag missing evidence. Any public page that uses AI-assisted classification or drafting should say so and identify the method version.
AI must not manufacture evidence, invent quotations, resolve an ambiguous edition match without review, or turn merchant star ratings into our own score. Source text remains the source of a claim; a model does not make it true.
Offer freshness and affiliate links
Offers remain separate from the edition. Each record carries a merchant, region, currency, price, stock state, observation date, refresh lifetime, and edition-match confidence. A missing or stale offer does not delete the evidence record.
Prices and availability can change after observation. The merchant page is authoritative at purchase time. Monetized outbound links are marked sponsored and are covered by the affiliate disclosure.
Corrections and change control
Source records retain enough provenance to trace and correct an identity or evidence error. Material methodology changes receive a new version; correction status and review dates should remain auditable.
Product makers, rights holders, reviewers, merchants, and readers can use the corrections and takedown route. Exact-edition identity errors receive priority because they can contaminate every downstream claim.