Evidence type is often used as a broad everyday label such as quantitative, documentary or testimonial. A class is a value assigned on a declared dimension. In a multi-axis system, one record can simultaneously be primary by origin, indirect by claim relation, qualitative by form, derived by lineage and historical by time.

The distinction matters because a single “type” invites false ranking. Recording several classes preserves the properties needed for later evaluation without pretending they are interchangeable.

No. Primary describes proximity to the originating process, not accuracy, completeness or relevance. A raw log can be corrupted, a firsthand account can be biased and a measurement can use an invalid construct.

A rigorous secondary synthesis may be better suited to a broad comparative claim because it evaluates many primary records through explicit selection and appraisal rules. Strength depends on the claim, method, provenance and limitations.

Yes, for different claims. A product specification is direct evidence of what the publisher officially declared. It is indirect evidence that every deployed product instance behaves as described. A ranking capture is direct evidence of an observed result position and indirect evidence of the mechanism that produced it.

Directness belongs to the relationship between a specific evidence item and a specific claim. It is not a permanent badge attached to the source.

A screenshot is a documentary visual record of a rendered state. Its evidential role depends on what is being claimed. It may directly preserve displayed wording or interface state while remaining incomplete evidence of underlying data, accessibility, source code or behavior outside the captured viewport.

Useful preservation includes the target, timestamp, device or viewport, capture method and surrounding context. The image alone should not be asked to prove more than it visibly records.

An AI output is primary documentary evidence of that output event when the run conditions are preserved. It directly supports claims about what the system produced, mentioned or cited during that run.

For claims about the external world, the generated answer is generally a secondary or derived representation. Its cited and uncited assertions require source-level verification. Model, prompt, date, interface, retrieval context and repeatability affect the evidence profile.

Only when the sources contribute meaningfully independent evidence lines. Ten publications quoting one study constitute ten distribution records but usually one originating measurement. Independence can be reduced by shared data, ownership, instrumentation, authorship or citation chains.

Map lineage before counting agreement. Diversity of observation method often matters more than the number of URLs.

There is no universal duration. Evidence becomes stale for a claim when change in the observed system could plausibly invalidate the conclusion beyond the accepted risk threshold. Volatile prices, rankings and interface states may need short windows; stable historical records may remain fit indefinitely for historical claims.

Define the validity window from system volatility, decision cost and monitoring needs. Publication age alone is an inadequate rule.

Numeric form does not eliminate judgment. Researchers choose constructs, units, thresholds, samples, transformations and aggregation rules. A precise number can still measure the wrong thing or compress important variation.

Quantitative evidence is valuable when its operational definition and measurement process are valid. Qualitative evidence may directly capture meaning, context and mechanism that a numeric proxy misses.

Treat it as derived computational evidence. Record the provider or producer, metric version, input scope, observation time and known transformation logic. Use it to support claims within the construct it actually measures, not a broader concept suggested by the score’s name.

When comparing scores, hold the measurement system constant. Changes in formula or index coverage can create apparent movement without change in the underlying asset.

Keep the mixed classification. A technical report may combine documentary declarations, quantitative tables, qualitative analysis and computed models. Classify the specific component used for each claim, or record multiple forms when the conclusion genuinely depends on their combination.

Forcing one category makes the evidence easier to label but harder to audit.

Mark origin as unresolved and lower the permissible use of the record. Search for citations, embedded metadata, identical phrasing, earlier versions and named datasets. The content can remain useful for discovery or context while being withheld from decisive support.

Unknown origin is not proof of falsehood. It is a provenance limitation that should control confidence and collection priorities.

At minimum, store the evidence object identifier, exact supported claim, origin class, directness class, form, lineage state, temporal state and the reason for each assignment. Add capture context, transformations, limitations and links to preserved material whenever possible.

The purpose is not bureaucratic labeling. It is to make later synthesis explainable: another reader should understand what the record is, what it can support and where its boundary lies.