1. Define the decision
We begin with a specific buyer question, not a product inventory. The intended audience, use case, constraints, comparison date, and decision criteria should be explicit before research begins.
2. Select products for relevance
Candidate products should be selected because they are plausible options for the stated buyer and use case. Partner availability may help determine what markets we research, but it does not establish which product is best. Relevant non-partners may be included and may be recommended.
3. Prefer primary evidence
For important factual claims, we prioritize:
- official pricing, product, terms, and help-center pages;
- official partner-program documentation where compensation is discussed;
- reproducible first-party tests with recorded inputs and methods; and
- calculations derived from verified inputs.
Unknown facts remain unknown. When material official sources conflict, we flag the conflict and do not automatically classify, score, or publish the affected claim.
Evidence scopes readers can inspect
Every material statement belongs to a bounded evidence scope. These labels describe what the record supports; none of them quietly implies that ChoiceMosaic is a customer of the product.
- Documentation-verified
- Supported by an identified official product, pricing, terms, or help source.
- Calculated
- Derived reproducibly from verified inputs, with the inputs and method recorded.
- Normalized-data
- Aligned across products or plans using a declared unit, billing basis, scope, and rule.
- Third-party
- A bounded external report or aggregate label, attributed with its access and sampling limits.
- First-party-tested
- Observed by ChoiceMosaic in a documented test with account, plan, steps, inputs, date, and result.
- Unknown
- Not verified by the available evidence and therefore not converted into a fact or recommendation.
4. Record direct testing accurately
A page may say that ChoiceMosaic tested something only when a direct test occurred and its scope, date, inputs, and method were recorded. Accessing a demo, reading a third-party review, or using automated analysis is not presented as hands-on product experience.
5. Use AI and automation with boundaries
AI and automation may assist with source discovery, extraction, consistency checks, drafting, and maintenance. They do not remove the evidence requirement, invent missing values, resolve a material source conflict, or supply fabricated experience. Review requirements depend on the risk and materiality of the change.
How Codex, code, and people divide the work
Codex organizes recorded sources, prepares drafts, and proposes reviewable changes. Deterministic code validates schemas, claim identifiers, freshness, evidence links, protected wording, SEO metadata, and publication gates. The human owner decides material editorial conclusions, legal and privacy language, commercial relationships, and whether a reviewable change should be merged or published. A passing machine check is not human approval.
The operator's training and experience in training design, software, project management, AWS, and security inform the evaluation frameworks and operational risk checks. That background is framework experience, not evidence of using or purchasing the products compared on a page.
6. Separate fact, inference, and compensation
Verified facts identify their evidence. Inferences explain ChoiceMosaic's reasoning from those facts. Affiliate or paid relationships are disclosed as commercial relationships. Compensation may influence research prioritization or visibly labeled commercial placement, but it does not determine rankings, recommendations, factual findings, or editorial conclusions.
7. Publish limitations and dates
Research pages should show when important information was verified, disclose meaningful limitations, and explain how to request a correction. A newer date is added only after a real review; dates are not changed merely to make a page look fresh.
Corrections
Readers and organizations can report a suspected error through the corrections process. Material corrections should identify what changed and when.