Built on a social contract.

Local models, local data.

Transceve runs all AI models on private infrastructure inside the EU (Germany/Finland, on Hetzner): speech recognition, a specialised analysis model, and a vision model for body language. No conversation data is ever sent to an external AI provider or third-party API. This means your data is never used to train external models and stays within infrastructure under our direct control.

Structured, constrained prompts.

Every BPN analysis uses a tightly structured prompt grounded in the Transceve Analytical Specification, a formal document defining exactly how each Basic Psychological Need should be scored and evidenced. The model is not asked open-ended questions. It is asked to return a specific JSON structure with scores, confidence levels, rationale, and qualitative extractions. This structure-first approach significantly reduces the space in which the model can confabulate.

Participant voice attribution.

One of the highest-risk areas for AI confabulation in conversation analysis is attributing statements to the wrong speaker. Transceve uses a post-processing heuristic filter that identifies the most likely participant (service user) voice using question ratio, pronoun usage, turn length, and professional language markers. Once identified, all qualitative extractions are constrained to that speaker's utterances, with worker speech excluded from evidence. This filter operates outside the analysis model, not as a prompt instruction the model could override.

Built for social impact, not general chat.

The Transceve analysis model is purpose-built for scoring conversations against the Self-Determination Theory framework, not adapted from a general-purpose chatbot. This narrower focus reduces the risk of the model applying incorrect or generic interpretations of psychological constructs, and ensures every analysis is grounded in the same analytical specification regardless of who is using the platform.

Limitations.

• Transceve scores are not a substitute for clinical assessment or professional psychological evaluation.
• BPN scores represent a structured AI interpretation of conversational evidence: they should be treated as indicators, not definitive measurements.
• Confidence levels are provided per score to indicate how well-supported each score is by the conversational evidence. Low confidence scores should be treated with additional caution.
• Structural flags are raised when the conversation contains features (very short sessions, dominant worker voice, insufficient participant speech) that may reduce the reliability of the scores.
• Transceve's grant report drafting feature uses AI to draft responses. All output should be reviewed, edited, and approved by a human before submission.

What to check.

• Always read the full transcript alongside the BPN scores: the scores are a summary, not a replacement for the conversation.
• Check the confidence level for each score. If confidence is low, review the rationale carefully before relying on the score.
• Review structural flags. A flag indicating "insufficient participant speech" means the score is based on limited evidence.
• Treat qualitative extractions as indicative quotes and verify them against the full transcript before including them in reports.
• All grant report drafts generated by Transceve must be reviewed and edited by a qualified person before submission. Never submit AI-generated content without human review.