Skip to content

[HT-70]: Analysis coach on replay — teach-from-evidence filter + verbosity settings #118

Description

@Maiz27

Parent

PRD-05: Game History & Replay — #111

What to build

Add a full-context analysis coach to the replay viewer that explains what was happening, at rest and without a clock, so a player can understand nuances they missed under live pressure. It consumes the HT-69 signals (card-death, feed-risk, collecting tells) over the reconstructed match state.

The existing coaching advisor already reads full live state, so the new work is the inverse: a teach-from-evidence filter that constrains advice so it is justified only by table-observable evidence (discards taken, melds laid, claim timing) and never cites hidden-hand information — the "judge with full information, teach only from observable evidence" boundary. Capture this boundary and the review-observation contract in a new docs/design companion document.

The analysis coach is the only coach shown while viewing (the live coach never appears in viewing — they are mode-exclusive). Add a review/coach settings category with a persisted default for analysis-coach verbosity (from narrate-everything to flag-only-clear-mistakes) and card-death warnings, plus an on-the-fly override when opening a replay, mirroring the persisted-default-plus-per-session-override pattern used by game setup.

Acceptance criteria

  • During replay viewing, the analysis coach narrates relevant moments (including dead-card and feed-risk situations) drawn from the HT-69 signals.
  • The coach's advice is justified only by table-observable evidence and never cites hidden-hand facts — verified by a test where hidden-hand info would change naive advice but the filtered output does not leak it.
  • Only the analysis coach appears in viewing; the live coach does not.
  • A review/coach settings category provides a persisted verbosity default and card-death-warning toggle.
  • Verbosity can be overridden on the fly when opening a replay without changing the saved default.
  • A docs/design companion documents the teach-from-evidence boundary and the review-observation contract.
  • Coach/filter logic stays in the domain/CPU layer; tone and correctness are validated by a human (HITL).

Blocked by

Metadata

Metadata

Assignees

No one assigned

    Labels

    ⚙️ settingsGame setup, user-configurable options, preferences screens, or rules customization controls.✅ readyWell-scoped, unblocked, and ready for implementation or review work to begin.📚 docsRepository docs, in-app help, rules documentation, contributor docs, or issue/PR templates.🔹 p2Normal-priority work that matters but does not block the immediate critical path.🖥️ uiFlutter screens, navigation, table layout, controls, feedback states, or responsive UI.🙋 hitlRequires human-in-the-loop judgment for rules, naming, design, tone, or product behavior.🧠 aiCPU move selection, difficulty, personality, reaction timing, or strategy evaluation.🧩 featureA vertical slice that delivers user-visible behavior or playable product capability.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions