Identify the output type
- Raw measurement: a value such as stroke width, baseline angle, timing or signal residual.
- Similarity or distance: a model-specific comparison, not automatically an authenticity probability.
- Screening flag or probability map: a lead for manual examination, with possible false alarms and missed signals.
- Likelihood ratio: relative support for defined propositions under the stated comparison model.
- Cryptographic verification: a technical check of a signed record and file integrity within that record’s scope.
Ask what was compared
Record questioned and known specimens, reference provenance and the features the method assessed. A writer ranking is closed-set: it returns a nearest candidate even if the actual writer is absent. A recording-content comparison asks whether files share recorded material; it is a different question from whether they contain the same speaker.
Keep evidential support separate from probability
A likelihood ratio does not, by itself, give the probability that a person wrote a document or committed an offence. The propositions, background population, model assumptions and broader evidence matter. Do not invent a percentage from an uncalibrated score or present a distance as a match verdict.
Treat absence carefully
A clean metadata check does not establish an unaltered file. Missing content credentials, printer patterns, hum traces or visible watermarks can reflect unsupported devices, unsuitable capture or signal loss. An absent indicator is often inconclusive rather than proof that the suspected event did or did not occur.
Investigate disagreements
Inspect the original alongside outputs. Consider natural variation, disguise, reuse, acquisition artifacts, compression and coincidental similarity where relevant. Explain why a particular observation changes the assessment. An attractive visual map should not override assessability limits or a conflicting measured result.
Report with appropriate weight
State the output, its scope, validation basis and material limitations. Distinguish automated output from the examiner’s reviewed opinion. Identify uncertainty and further work. See reporting guidance and the analysis notice.