# `ui/` — AI usage evaluator UI shell

Self-contained Gradio front-end. Runs standalone (no backend / ML deps) so it can be developed
and deployed independently of the `prompt_card` workstream.

## Run
```bash
python -m ui.app          # serves the Gradio app
```

## Layout
```
ui/
  app.py            # gr.Blocks, event wiring, the staged-reveal generator
  theme.py          # design tokens, dark CSS, CDN head (fonts + Chart.js + flip/radar JS)
  data.py           # CardData contract + stub provider  ← shared data shape
  components/card.py# render CardData -> flip-card HTML + accordion bodies
  screens/          # intake / processing / result column builders
  parsing/          # Task 2: ChatGPT+Claude .json/.zip parser + zero-dep language gate
  scoring/          # Task 3: Scorer protocol + DummyScorer (heuristic placeholder)
```

## The three adoption seams (where the real backend plugs in)
1. **Parsing** — `ui/parsing/parser.py::parse_export(path) -> ParsedExport`.
   Mirrors the real export shapes; the backend's `prompt_card.adapters` are the reuse target
   when the workstreams merge.
2. **Scoring** — `ui/scoring/interface.py::Scorer`. Implement `score(parsed) -> ScoreResult`
   with a real model and pass it where `DummyScorer()` is used in `app.py`. The UI is unchanged.
3. **Card data** — `ui/data.py::CardData` is the stable render contract. `score_to_card` maps a
   `ScoreResult` onto it.

## Status (stubbed vs real)
- Real: export parsing, language split, processing facts (counts/date range/busiest slot).
- **Scoring ADOPTED** (`ui/scoring/observable.py::ObservableScorer`): the 5 axis scores, overall/tier, and
  per-axis confidence (measured Cohen's κ) now come from the backend `prompt_card.observable_pipeline`
  (base MiniCPM4.1-8B + the locked per-category LoRA hybrid). `app.py` calls `get_scorer()`, which returns
  the real scorer when `OPENBMB_BASE_URL`/`OPENBMB_TOKEN` are set, else falls back to `DummyScorer` (so the
  UI still runs standalone with zero ML deps). A live scoring failure degrades to the dummy mid-demo.
- Now REAL too: per-axis **evidence quotes** are the actual user turns where each axis fired (pipeline
  `data["evidence"]`; falls back to a representative turn only when an axis had zero positives), and the
  **critical breakdown** shows true per-type **counts** (pipeline `data["critical_type_counts"]`).
- Still heuristic/static: the per-axis **tips** and the **improvement** line (advice, not measurements).