Editorial Policy — How We Research, Generate and Review

This site publishes review-derived cookware intelligence. This page explains where the numbers come from, how pages are generated, where AI is used, and how a human reviews the result.

What we publish

• Aggregated evidence, not raw reviews: counts, rates, prevalence, sample sizes and short representative quotes.
• Reproducible pages: generated offline from a versioned dataset with a fixed methodology, not written ad hoc.
• Explicit evidence chains: every insight page shows the path Origin → Signals → Evidence → Products → Claim.

Where the questions come from

We separate the origin of a question from the evidence that answers it.

• Review corpus (primary). Questions are derived from problem prevalence and mention counts per product type in the review data — e.g. rust, coating failure, warping. The pages that answer them cite the exact prevalence and sample.
• Curated taxonomy. Suitability and scenario pages start from a human-maintained taxonomy (induction, dishwasher, oven, PFAS-free, care; cookie baking, thanksgiving, …), then attach review evidence.
• External demand signals (optional, clearly labelled). Where used (for example community discussions or search-interest proxies), they only influence which questions are written. They are labelled demand_proxy and are never presented as review evidence.

We do not redistribute raw review text or scraped content; external signals are stored as aggregates with a permalink only.

How a page is generated

1. Data layer — an offline SQLite dataset is aggregated into versioned JSON bundles (counts, scores, problem prevalence, specifications).
2. Evidence chain — Python builds the Origin → Signals → Evidence → Products → Claim record for each page, with a stable data_hash.
3. Draft text — where prose helps, an LLM produces a draft through a Python pipeline (never an autonomous agent). The model may phrase explanations but may not invent numbers, products or URLs: figures are injected from the data and validated.
4. Human review — an editor reviews the page, adds an editor’s note and tips, and signs it (reviewed_by, reviewed_at). The page shows whether it is reviewed or still an AI draft.
5. Staleness check — when the underlying data changes, the page’s data_hash no longer matches its editorial overlay and it is flagged for re-review.

AI, and its limits

• AI is used offline, in Python, for bounded wording tasks with structured output, caching and validation.
• AI never computes scores, prevalence, rankings or prices — those are computed in Python from data.
• Generated pages are committed and build without any runtime AI call.

Corrections

Found an error? Use the contact link in the footer. Corrections that change a claim are logged on the changelog and the page is re-reviewed.

Disclosure

Compatibility is inferred from material composition, not from manufacturer specifications; we say so on every suitability page and recommend verifying with the product listing. Scores require at least 3 sampled reviews. Data vintage is 2023; prices are not real-time. As an Amazon Associate we earn from qualifying purchases — this does not affect the evidence we publish.

See also: Scoring standards · Methodology notes · Brand reliability .