ADR-0005: Run Tier 1 as a Socratic dialogue with banned buzzwords and bounded follow-ups

Accepted

Status: accepted · 2026-10-24

Context

The coach must make the learner explain mechanisms in their own words, not recite terms. A local model left to itself tends to explain, summarise or quote.

Decision

1. A round opens with one challenge that asks for the central mechanism, then alternates the learner's answer with either one follow-up or a conclusion (pass or fail, gaps, feedback).
2. Banned buzzwords are checked locally, ignoring case and diacritics (including ł) and matching whole words only. An answer with one sends the same question back for a rephrase without calling the model. After maxRephrases (default 2) the answer is accepted, and the buzzwords used are added to the gaps.
3. At most maxFollowUps (default 3) follow-ups; after that the round concludes even if the model wants to probe.
4. A challenge that quotes the note is asked again with the rejected text, up to maxChallengeAttempts (default 3); then the round refuses to start. A follow-up that keeps quoting is replaced with a fixed neutral probe.
5. The Socratic rules are given to the model as instructions: one question per turn, never explain or hint, demand the mechanism, define terms plainly, probe the edges, challenge hand-waving.
6. The buzzword list, the rules, the limits and the question language are configurable in coach.config.json (sparring), with defaults in the domain.
7. The recorded answer is the first answer plus each follow-up and its answer.

Consequences

• These rules live in the SparringRound domain object, so they hold whichever model runs Tier 1.
• The first real run on gemma3:latest gave vague gaps and empty feedback; prompt tuning is expected.