ADR-0002: Coach in two tiers: an offline sparring partner on local Gemma and a zero-knowledge judge on Gemini Argon

Accepted

Status: accepted · 2026-10-24

Context

ai-coach is the demo for a conference talk that argues against "chat with your docs" tools: they hand out summaries and stop the reader thinking. The coach must instead make the learner explain their own private notes. It must work with no network (on a train or a plane), and the notes must never leave the machine.

Decision

1. Tier 1, the sparring partner, runs offline on a local Gemma model served by Ollama. It reads the notes and runs a Socratic, Feynman-style dialogue (ADR-0005). The model comes from OLLAMA_MODEL; the host defaults to http://127.0.0.1:11434 and must be plain http on loopback, or the adapter refuses to start.
2. Tier 2, the judge, runs online on Gemini through @google/genai. It never sees a note: it gets only the challenge Tier 1 asked and the learner's answer (ADR-0004), and judges them with the blind claim-extraction protocol (ADR-0006).
3. The judge model is gemini-4-argon with thinkingLevel: 'MEDIUM' (or 'HIGH'). An optional fallback, for example gemini-3.8-flash, is used only when Argon answers 429 or 5xx or cannot be reached; a 400 still fails.
4. Authentication is GEMINI_API_KEY read by new GoogleGenAI() against the Gemini Developer API; Vertex AI is switched off.
5. Tier 1 records a session first; Tier 2 judges it later, whenever the learner is back online.

Consequences

• Sparring needs Ollama and a pulled Gemma model; judging needs network and a key.
• If GOOGLE_API_KEY is also set, the SDK uses it instead of GEMINI_API_KEY.
• Tier 1's verdict and Tier 2's verdict are recorded separately, and mastery is derived from both (ADR-0003).