Solo AI CoachBack to Solo AI Coach

AI System Card

Last Updated: August 14, 2026

Effective Date: August 2026

This document describes the AI system that powers session analysis in Solo AI Coach, for transparency toward coaches, their clients, and regulators (EU AI Act Article 13 transparency obligations for limited-risk AI systems; general AI-system documentation good practice). It reflects the system as implemented in code as of the Last Updated date below.

System Purpose

Solo AI Coach's analysis engine reads a coaching session transcript and identifies communication patterns — linguistic patterns (meta-model violations), beliefs, emotional signals, and generates coaching-relevant suggestions — to help a coach reflect on and prepare for their next session with a client.

It is a reflection and preparation aid for the coach who uploads the session. It is not a diagnostic, clinical, or decision-making tool, and its outputs are not validated for use as one.

Inputs

  • Text only. The system processes transcript text — either pasted/uploaded directly, or produced by a separate transcription step (Deepgram, for audio uploads) before analysis. The analysis model itself never receives audio or video.
  • No biometric, voice-tone, or video data is analyzed at any stage. Transcription converts speech to text; nothing about vocal characteristics (pitch, pace, biometric identifiers) is passed to or used by the analysis step.

Processing

  • Model: Claude Sonnet 5 (claude-sonnet-5), via the Anthropic API, called once per analysis with the full transcript as input.
  • Method: A fixed system prompt (versioned separately in docs/analysis-kit/) instructs the model to detect meta-model linguistic patterns (a coaching-methodology framework for identifying deletions, distortions, and generalizations in language) and report findings through a structured tool-call schema — the model cannot return free-form prose instead of the defined fields.
  • Two modes: "Conversational" (general coaching) and "Hypnotherapy" — each uses a distinct prompt and output schema tailored to that coaching context.
  • No fine-tuning, no training on customer data by Solo AI Coach itself. Whether Anthropic uses API input to improve its own models is governed by Anthropic's API terms, not by Solo AI Coach — see the Anthropic Privacy Policy linked from our Privacy Policy, Section 4.

Outputs

Each analysis returns structured findings: detected belief statements, emotional signals, meta-model pattern instances (with short evidence quotes from the transcript), and suggested coaching questions or interventions. A "session summary" narrative field is included for the conversational mode.

Outputs are NOT:

  • A diagnosis of any medical, psychological, or psychiatric condition
  • A clinical or mental-health assessment
  • A recommendation to take any real-world action without the coach's own professional judgment
  • A statement of fact about the client — they are the model's interpretation of patterns in language, which can be wrong (see Limitations)

Limitations & Accuracy

  • No published accuracy benchmark. The prompt and schema were developed and iterated against a working set of transcripts, but we have not run a formal accuracy evaluation against a labeled synthetic-transcript test set, and no accuracy or precision/recall figure is published. Any claim that the system's findings have been "tested" or "validated" against ground truth would be inaccurate — treat every finding as the model's interpretation, not a verified fact.
  • Language and cultural scope. The prompt and meta-model framework were developed primarily against English-language coaching conversations. Accuracy on other languages, dialects, or culturally-specific communication styles is untested.
  • Transcript quality dependency. For audio uploads, transcription errors (misheard words, misattributed speakers) propagate directly into what the analysis model sees — the analysis has no way to know a transcript is wrong.
  • Hallucination risk. Like any LLM, the model can occasionally state a pattern or belief that isn't well-supported by the transcript. The evidence-quote requirement in the output schema is a mitigation (it forces the model to ground each finding in transcript text a coach can check), not a guarantee.
  • No temporal or cross-session memory. Each analysis considers only the single transcript submitted — the model does not track a client across sessions, and any apparent "progress" a coach sees is the coach's own interpretation across multiple independent analyses, not a system-generated trend.

Human Oversight

Every output is delivered to the coach for their own review — the system does not act on its findings (e.g. it doesn't message a client, flag an account, or trigger any automated consequence). The coach:

  • Reviews the transcript and findings together before treating anything as accurate
  • Can edit the transcript and re-run analysis if the transcript itself is wrong or incomplete
  • Confirms, before every upload, that they have lawful authority to upload the session and that the client has consented to AI analysis (logged in consent_log — see the Record of Processing Activities, docs/RECORD-OF-PROCESSING-ACTIVITIES.md)

There is no fully-automated decision loop anywhere in this system — a human (the coach) is always the one interpreting and acting on the output.

Monitoring

  • Every analysis run is logged (analysis_audit_log, data_use_decisions) with the tier used and whether the resulting data is eligible for research use — this is an operational/audit log, not a model-quality monitoring pipeline.
  • There is currently no automated system that scores or flags individual analysis outputs for quality or safety after generation — quality feedback today comes from coaches reporting issues via support, not automated monitoring. This is a documented gap, not a built capability — see the Incident Response Workflow, docs/INCIDENT-RESPONSE-WORKFLOW.md, for how a reported problem is handled.

Prohibited / Out-of-Scope Uses

This system is not built, tested, or intended for:

  • Clinical diagnosis, treatment planning, or any mental-health decision-making
  • Employment decisions (hiring, performance evaluation, termination) about any individual, coach or client
  • School or educational placement/assessment decisions about any individual
  • Voice or video biometric analysis of any kind
  • Training a shared model on customer data without that customer's explicit opt-in (see tier-based research eligibility in the Record of Processing Activities, docs/RECORD-OF-PROCESSING-ACTIVITIES.md, Activity C)
  • Any use requiring HIPAA-covered data handling — no Business Associate Agreement (BAA) is in place with Anthropic, Deepgram, or Supabase for this Service as of this document's date, so PHI subject to HIPAA should not be uploaded

If you are using this tool in a context that resembles any of the above, stop and contact privacy@soloaicoach.com before continuing.

Contact

Questions about this system card: privacy@soloaicoach.com

Version: 1.0 Last Updated: August 14, 2026