Themis conducts recorded witness interviews for police investigations. It follows the NICHD protocol, screens every question before speaking it, and produces court-ready records.
No polygraph / No biometrics / No lie detection
0
protocol phases, order enforced in code
0
categories of suggestive questioning screened
0
linguistic indicators in every analysis
0
languages, spoken and transcribed
By design
Most interview errors are not lies. They are suggestions, absorbed and repeated until they sound like memory. So the interviewer's constraints matter more than its abilities. There are six things it will not do.
Mention any detail the witness has not raised themselves.
Suggest a name, a face, a place, a time, or a weapon.
Praise, pressure, or reassure a witness toward an answer.
Interrupt a narrative once it has begun.
Continue past a disclosure of ongoing danger.
Overrule a request to pause or to stop.
The record
Six interview phases, from rapport to closing, in a fixed order. The engine enforces sequence and pacing in code, not by convention.
Every generated question is checked against ten categories of suggestive questioning before it is spoken. A flagged question is rewritten, not asked.
Each answer receives a safety read. A disclosure of ongoing danger stops the interview and alerts the case officer within seconds.
A break, a repeat, or an end to the interview, at any time, by voice or by typing. The room obeys the witness.
Statements are compared across every interview in a case. Differences are quoted, categorised by topic, and rated by severity.
A court report with methodology and disclaimers, and an internal report with recommended follow-up questions for the next session.
Phase 01
Ground rules first. The witness hears that "I don't know" is always an acceptable answer, that they can correct the interviewer, and that they can pause or stop at any time.
No case content is allowed in this phase. Rapport is built without touching the incident.
Phase 02
A practice round of free recall on a neutral topic, usually the witness's morning. This trains the pattern the whole interview relies on: long, uninterrupted answers.
Narrative practice measurably improves the quality of later substantive recall.
Phase 03
The witness is guided back to the scene through senses only. What could you see, what could you hear, how did it feel. No facts are suggested, nothing is named.
Context reinstatement is one of the strongest documented memory retrieval aids.
Phase 04
One open invitation, then silence: tell me everything that happened, from the very beginning to the very end. The interviewer does not interrupt, evaluate, or redirect.
Free narrative comes before any probe, always. This ordering is enforced by the engine.
Phase 05
Focused questions, but only about details the witness has already raised. If the witness never mentioned a car, the interviewer cannot ask about one.
Every question in this phase passes the ten-category screening model before being spoken.
Phase 06
A final open check: anything you want to add, anything you are unsure about. The interview ends without pressure and thanks the witness for their time.
The witness can revise or flag uncertainty about anything said, and it enters the record.
Phase 01
Ground rules first. The witness hears that "I don't know" is always an acceptable answer, that they can correct the interviewer, and that they can pause or stop at any time.
No case content is allowed in this phase. Rapport is built without touching the incident.
Phase 02
A practice round of free recall on a neutral topic, usually the witness's morning. This trains the pattern the whole interview relies on: long, uninterrupted answers.
Narrative practice measurably improves the quality of later substantive recall.
Phase 03
The witness is guided back to the scene through senses only. What could you see, what could you hear, how did it feel. No facts are suggested, nothing is named.
Context reinstatement is one of the strongest documented memory retrieval aids.
Phase 04
One open invitation, then silence: tell me everything that happened, from the very beginning to the very end. The interviewer does not interrupt, evaluate, or redirect.
Free narrative comes before any probe, always. This ordering is enforced by the engine.
Phase 05
Focused questions, but only about details the witness has already raised. If the witness never mentioned a car, the interviewer cannot ask about one.
Every question in this phase passes the ten-category screening model before being spoken.
Phase 06
A final open check: anything you want to add, anything you are unsure about. The interview ends without pressure and thanks the witness for their time.
The witness can revise or flag uncertainty about anything said, and it enters the record.
Every single turn
The witness speaks. The microphone is closed whenever Themis is talking, so its own voice can never enter the record.
Each response gets a safety read: distress level, disclosure of ongoing danger, self-harm risk. A calm request to stop is a right, not an emergency.
The engine chooses one action: ask, probe deeper, advance the phase, offer a break, confirm a stop, end, or escalate. Phase rules are enforced in code.
If the action is a question, an independent model checks it against ten categories of suggestive questioning. Failures are rewritten before speech.
Only then is the question spoken aloud, in the witness's language, and the cycle repeats.
Limits
Themis flags linguistic patterns for professional review and always shows positive indicators alongside concerns. Its output is an investigative aid, never a verdict. Recordings are stored privately and served only to the authenticated case officer.
Open the demo workspace, file a case with its own PIN, and take a live interview. Minutes later you have the transcript, the analysis, and both reports, exactly as an investigator would.