The Review GroupEvidence before recommendation
Browse records
Methodology version 2.0

How evidence becomes a published score.

Effective 2026-08-13 · Named AI-assisted editor: Penny

A score is publishable only when a reader can trace the exact item, source, retrieval date, weight, calculation, limitations, and correction history. If the record cannot show that receipt, the score stays private.

Three public statuses

Published evidence report

Exact identifier, linked and dated sources, visible calculation, limitations, editor, and history are complete.

Evidence review

Research exists internally, but one or more publication fields are incomplete. No score, recommendation, or affiliate button appears.

Catalog only

Basic bibliographic metadata is searchable. The item is not presented as reviewed.

Current published method

Product customer-evidence calculation

The current product reports use one source: the customer-rating aggregate on an exact Amazon ASIN. That limitation is stated on every report. Editorial articles may appear as background sources, but they do not change the number.

  1. 1
    Match the item.

    A 10-character ASIN is required. The title, brand label, and listing must refer to the same product model.

  2. 2
    Check evidence volume.

    Fewer than 5 ratings is declined. From 5–29 ratings, the item can be held as an editorial record but cannot publish a score. A scored report requires at least 30 ratings.

  3. 3
    Normalize sentiment.

    The 0–5 customer rating is multiplied by 20 to create a 0–100 sentiment signal.

  4. 4
    Apply the evidence engine.

    The engine represents up to 40 rating signals. Praise items are round((normalized signal ÷ 100) × item count); remaining items are complaints. Net sentiment is (praise − complaints) ÷ item count. The quality/performance dimension is 50 + 45 × net sentiment, rounded to one decimal and then to the displayed whole number.

  5. 5
    Publish the receipt.

    The page shows the raw rating, rating count, ASIN, retrieval date, normalized signal, calibration difference, final score, method version, and limitations.

4.5/5× 2090/100 signal→ 36 praise − 4 complaints86/100 evidence score

Configured dimensions

DimensionConfigured weightCurrent report coverage
Quality and performance30%Customer sentiment is mapped here.
Reliability and durability20%Not separately evidenced.
Value for money20%Not separately evidenced.
Design and usability15%Not separately evidenced.
Customer experience15%Not separately evidenced.

Important limitation: the engine renormalizes around the one evidenced dimension. That is why these pages are labeled customer evidence reports rather than comprehensive product tests.

Confidence and freshness

A single retailer source can never receive “high” confidence. Reports with 1,000 or more ratings display medium confidence; reports with 30–999 display low confidence. Evidence is dated on the page. A material product, listing, or score change triggers a visible correction entry rather than silently replacing history.

Books, ebooks, and audiobooks

The internal book configuration includes content quality (30%), practical value (25%), readability/craft (20%), originality (15%), and longevity (10%). Source types have different configured reliability weights. Those internal configurations are not enough to publish a score: each public record also needs direct source links, dates, edition identifiers, numerical contributions, limitations, and an accountable editor.

Current status: those receipt fields are incomplete across the catalog, so all book, ebook, and audiobook scores and recommendation labels are withheld. Exact narrator, publisher, ISBN/edition, and source matching must pass before any one returns.

What Penny does—and what the publication controls

Penny is the named AI-assisted editor and research guide. Automation can collect structured fields, normalize evidence, calculate a score, flag draft language, and draft a summary. The Review Group maintains the methodology, conflict decisions, publication controls, corrections, and removal of a report.

We distinguish source fact, model calculation, and editorial conclusion. The current product pages publish the first two and avoid a blanket purchase conclusion.

Methodology version history

  1. Added hard publication gates, common-ownership prohibition, canonical score rendering, public calculation tables, AI/human accountability, correction history, and the book/audio score hold.
  2. Initial aggregate evidence engine. Superseded for public presentation because receipt and conflict controls were incomplete.

Questions

Why are book and audiobook scores hidden?

Their current records do not yet contain linked sources, retrieval dates, edition matching, weights, and a reproducible final calculation. They remain searchable as evidence checks, without public scores.

What does a product evidence score measure?

The current published product reports measure dated Amazon customer sentiment for one exact ASIN. They are not hands-on tests and do not measure every product dimension.

Can an affiliate link make a product eligible?

No. Candidate selection happens before affiliate availability is checked, and products without purchase links remain eligible.

Can a commonly owned company receive a score?

No. Commonly owned organizations appear only on the Our projects page as factual profiles without scores, ranks, or recommendation labels.

Privacy choices

Essential storage remembers these choices. Optional services stay off until you allow them.