How the evidence earns the claim.

Evidence hierarchy

1. Primary documentation: an authoritative party’s description of its own system, policy, standard, or law.
2. Peer-reviewed research: published work evaluated within a stated scholarly process and scope.
3. First-party measured result: direct observations from a Bob-owned or consenting system with method and limitations.
4. Expert analysis: a reasoned conclusion connecting disclosed evidence.
5. Working hypothesis: a testable idea that has not earned a stronger label.

Content-basis labels

Each editorial page identifies whether it is documented guidance, original analysis, a first-party case study, or mixed evidence. This prevents a recommendation from quietly borrowing the authority of a platform document or research paper.

Case-study protocol

1. Define the subject, owner relationship, question, metrics, and fixed time window.
2. Preserve the pre-intervention baseline and exact comparison method.
3. Build a dated implementation timeline.
4. Reconcile sources without overlapping windows.
5. Report absolute values, denominators, and comparison values.
6. Name confounders and plausible alternative explanations.
7. Distinguish measured facts, interpretation, causation, and transferability.
8. Retain source artifacts and publish a privacy-safe evidence package.

AI-answer measurement

Prompt panels are frozen before the evaluation period. Records include exact prompt, product and visible mode, date, geography, answer, cited URLs, factual review, and repeated trials. Mention, citation, correctness, recommendation, referral, and conversion are separate measures. One response is never described as a stable global ranking.

Claim register

Material credentials, roles, service terms, prices, study metrics, client outcomes, and comparative statements should have a preferred wording, source, verification date, owner, and prohibited extrapolations. A claim is removed or narrowed if the supporting evidence becomes unavailable or stale.

Source practice

Primary sources are preferred. A source must directly support the nearby wording; popularity is not evidence quality. Access dates and archived copies are retained for volatile sources where practical. The sources section does not excuse unsupported wording in the body.

AI-assisted work

Models may help organize notes, inspect supplied code or evidence, suggest test cases, and critique drafts. They are not accepted as factual sources. Material claims are checked against primary records by a human accountable for publication. Confidential client data is not submitted without an approved data agreement.

Maintenance and correction

Every editorial item has an updated date and next review date. Material corrections change the page promptly, preserve the repository diff, and appear in the correction log. Major methods changes create a versioned update rather than silently rewriting past findings.