Methodology
The research the system stands on
Themis is built on the two most validated frameworks in investigative interviewing, with an independent screening layer derived from the eyewitness-suggestibility literature. This page lists exactly what is enforced and how.
Interview structure
Interviews follow the NICHD Investigative Interview Protocol (Lamb, Orbach, Hershkowitz, Esplin and Horowitz, 2007) combined with the Cognitive Interview (Fisher and Geiselman, 1992). Six phases run in fixed order, and the engine enforces minimum and maximum exchange counts per phase in code. The interviewer cannot skip free recall, cannot probe ahead of the witness's own account, and cannot keep a witness in a phase indefinitely.
Question screening
Before any generated question is spoken, an independent model screens it against ten categories of suggestive questioning, informed by Loftus and Palmer (1974) and the misinformation-effect literature that followed. Flagged questions are rewritten; if no safe rewrite is found after two attempts, a neutral open prompt is used instead.
Presupposition
Assumes a fact the witness has not established.
Forced choice
Offers options that may exclude the witness's actual answer.
Detail insertion
Introduces a person, object, colour or time the witness never mentioned.
Verb intensity loading
Uses loaded verbs implying severity or intent not established.
Confirmation pressure
Phrasing that pushes the witness to confirm rather than recall.
Authority pressure
Invokes legal consequence or urgency to compel an answer.
Temporal compression
Forces a sequence before the witness has offered one.
Acquiescence trap
Phrased so that agreeing means confirming the interviewer's version.
Misinformation embedding
Uses case-file details as if the witness had said them.
Closure pressure
Rushes the witness or signals the interview is nearly over.
Safety monitoring
Every witness response is assessed for distress level, disclosure of ongoing danger, and self-harm risk. Elevated distress prompts an offer to pause. Critical distress, ongoing danger, or self-harm risk stops the interview: the witness sees emergency contact information and the case officer's dashboard is flagged. A calm request to stop is honoured as a right, not treated as an emergency.
Post-interview analysis
Completed transcripts are analysed for eight linguistic indicator types, and statements are compared across all interviews in the case for contradictions, which are quoted verbatim and rated by severity. All findings are computational indicators derived from language patterns. They are not clinical assessments, they carry no claim about truthfulness, and they require review by qualified forensic professionals before any use.
Hedging language
Qualifiers concentrated in specific parts of the account.
Over-specificity
Unusual precision where generic recall would be expected.
Under-specificity
Conspicuous vagueness where detail would be natural.
Timeline inconsistency
Internal contradictions in the sequence of events.
Formulaic language
Scripted-sounding phrases repeated verbatim.
Emotional mismatch
Tone inconsistent with the reported events.
Spontaneous corrections
Unprompted self-correction. A positive indicator.
Unprompted detail
Volunteered peripheral detail. A positive indicator.
References
- Lamb, M. E., Orbach, Y., Hershkowitz, I., Esplin, P. W., and Horowitz, D. (2007). A structured forensic interview protocol improves the quality and informativeness of investigative interviews with children. Child Abuse and Neglect, 31(11-12), 1201-1231.
- Fisher, R. P., and Geiselman, R. E. (1992). Memory-enhancing techniques for investigative interviewing: The cognitive interview. Charles C Thomas.
- Loftus, E. F., and Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. Journal of Verbal Learning and Verbal Behavior, 13(5), 585-589.
For the full architectural write-up, read the design notes.