OPEN METHOD

How Clear reaches an observation.

Last updated: 14 August 2026

Clear is an experimental AI system. The method below is published so its assumptions can be challenged; publishing it does not make the output authoritative or scientifically validated.

1. Read the whole post

Clear considers the caption and every available slide or video segment, then checks the central takeaway a reasonable viewer would receive.

2. Stop when no check is needed

Memes, scenery, jokes, opinions, and posts without a researchable factual assertion do not go through evidence research. If language cannot be understood reliably, Clear says so and does not charge a work unit.

3. Prefer primary evidence

Official records, original studies, regulators, direct company documents, and high-quality systematic reviews are preferred. News and expert secondary sources provide context when primary material is unavailable.

4. Apply the verdict fairly

• SUPPORTED BY EVIDENCE: the central takeaway is substantially correct; minor caveats do not change it.
• MISSING IMPORTANT CONTEXT: some basis exists, but omitted context or framing materially changes the takeaway.
• CONTRADICTED BY EVIDENCE: reliable evidence directly contradicts the central takeaway.
• INSUFFICIENT EVIDENCE: reliable evidence is insufficient, mixed, inaccessible, or genuinely disputed.
• OPINION: a judgment rather than a testable assertion.

5. Calibrate before labelling

The research model records whether the claim is testable, whether evidence is sufficient, whether the central claim is supported or contradicted, and whether omitted context is genuinely material. A deterministic policy then selects the label. Insufficient evidence cannot default to “missing context.” Negative calls receive a separate fairness review.

6. Keep wit aimed at claims

The tone is added after the evidence assessment. It may challenge wording or framing, but must not insult a person, allege dishonesty or speculate about motive.

Current publication rule

Substantial factual observations with at least two usable sources, adequate confidence, and complete reasoning enter Clear’s public research archive and sitemap. Opinions, no-claim posts, low-confidence checks, weakly sourced results, failures, and unsupported-language attempts remain excluded from search indexing. Published observations retain correction and removal controls. Clear does not publish aggregate verdict rates as scientific findings because submissions are self-selected.

Known limitations

Automated transcription, translation, visual interpretation, web search and source selection can all fail. Confidence is a model estimate, not a measured probability. Clear is building a balanced evaluation set covering supported, contradicted, context-dependent, uncertain, opinion, humour and no-claim posts.