# webgym: four new task kinds (car, restaurant, signup, filter)

Added 2026-09-26.  All kinds follow the existing conventions: `(kind, seed)` -> deterministic spec (`spec.py`, registered in
`KINDS`), rendered by `app.py` (port 8811), widgets in `widgets.js` keep committed values in `window.__gym.state`, the scripted
expert in `expert.py` acts through jev's real observations (only controls jev offers; scrolls when a control's centre is outside
the viewport; accepts the cookie banner when it covers something), and `eval_gym.py::success()` judges the real agent from page
state.  Held-out rule unchanged: `seed % 10 == 0` is never generated for training; eval seeds are `900000 + 10*i`.
Randomised per seed: field labels, widget implementation, field order (35–30 % shuffled), hero pushing the form below the fold,
cookie banner (25 %), four decoy nav links, wording of the goal.

Run the expert:  `../jev-ultrafast/.venv/bin/python finetune/webgym/expert.py out.jsonl 20 <seed0> car|restaurant|signup|filter 0.12`
Clean data:      `python finetune/webgym/clean3.py`  (gym_raw3/g_* -> gym/clean3_*, only episodes with a task_done case)
Eval:            `python finetune/webgym/eval_gym.py 10 car,restaurant,signup,filter 60`

## car — car rental search (`/t/car/<seed>` form -> `/results` -> `/results/c/<id>`)
Steps: pick-up location (autocomplete, inline or behind a trigger button; city or airport options), optional "return to a different
location" **toggle** (native checkbox or `role=switch`) that reveals the drop-off autocomplete, pick-up date + time, return date + time,
driver age, submit.  Post-submit (about half the tasks): car-type checkboxes (Economy/Compact/SUV/Van/Luxury, multi-select), transmission
radio (Any/Automatic/Manual), sort select, or the chain "filter SUV -> sort price low-high -> open the cheapest SUV" (car page,
`meta.car`).  Wrong/early submissions are recovered through "Modify search".
Widget variants: dates picker (with/without Done/Apply) or typed MM/DD/YYYY; times native select or custom dropdown listbox (16 hourly
options); age native select ("27 years") or stepper (targets within ±5 of the default 30); autocomplete typed as full name, prefix,
alias (Zurich, MUC…) and committed by click or Enter.
Sub-goals: pickup, [diff, dropoff], pdate, ptime, rdate, rtime, age, submit, [ctype | trans | csort | ctype+csort+open] (7–12).

## restaurant — table reservation (2 stages -> confirmation)
Stage 1: party size (select / stepper / popup steppers), date (calendar picker or typed), **time-slot buttons** (10 slots 17:00–21:30,
2–4 per seed disabled = unavailable, never offered by jev), seating preference (radio / segmented / select).  Stage 2: name, email, phone
(text inputs), special requests (textarea, in half the tasks), "I agree…" (checkbox or switch), confirm -> confirmation page
(`meta.done`, `meta.carried` = every committed value).  Required-field validation (`role=alert` messages) blocks the stage submit; a Back
button returns to the previous stage with values preserved.  Text values come verbatim from the goal ("for Ada Lovelace, ada@example.com").
Sub-goals: party, date, slot, [seat], go1, name, email, phone, [req], agree, go2 (9–11).

## signup — 3-page account wizard (no password fields)
Stage 1 account: full name, email, username.  Stage 2 profile: country (native select with placeholder / custom dropdown / autocomplete),
date of birth (typed MM/DD/YYYY or month+day+year selects), interests (checkbox group or toggle chips, 1–3 of 8).  Stage 3 preferences:
newsletter toggle (checkbox / switch), plan (radio / segmented / select: Free, Pro, Team).  Next/Back on every stage, validation messages
for empty required fields, final "Create account" -> welcome page.
Sub-goals: name, email, username, go1, country, birth, interests, go2, [newsletter], [plan], go3 (9–11).

## filter — faceted product listing (`/t/filter/<seed>?f=1&…` -> `/p/<id>`)
120 products with brand (8), colour (6), size (5), price, rating.  Facets: price min/max text inputs + Apply (or Enter), rating radios
(any / 3+ / 4+), brand and colour checkbox groups with **Show more** revealing the hidden options, size checkboxes, applied-filter chips
with remove buttons (+ Clear all), sort (select or dropdown), pagination (numbers / Next / both).  35 % of tasks start with a pre-applied
brand or colour filter; 60 % of those require REMOVING it (chip or unchecking), the rest keep it.  Goals combine 2–4 constraints, 25 %
end on page 2, 40 % end by opening the first result (the expected product id is computed from the same listing logic, `apply_listing`).
Sub-goals: [rm_brand|rm_color], brand, color, size, rating, price, sort, [page], [open] (2–6; a facet with two values is two clicks).

## DART-style noise and recovery
* 12 % per step: a harmless detour (click a decoy link or focus a field) that the expert then recovers from.
* car (25 % of episodes): one premature submit -> wrong results -> "Modify search" recovery (`skill=recover`).
* restaurant/signup (35 %): one premature Next while a required field is empty -> validation messages -> the expert fills the field
  and retries.  The stage is only skipped early when validation is guaranteed to reject it, so carried values are never wrong.

## Expert success rates (20 episodes each, noise 0.1, seeds 1…)
| kind       | smoke test (20 ep, noise 0.1) | verification after fixes (noise 0.12)            |
|------------|-------------------------------|--------------------------------------------------|
| car        | 20/20                         | 59/60 (seeds 700001+); the miss (submit covered by an open dropdown) is fixed, re-run 16/16 |
| restaurant | 20/20                         | 15/15 (seeds 710001+, mixed run)                 |
| signup     | 20/20                         | 15/15                                             |
| filter     | 20/20                         | 15/15                                             |
| flight/hotel/shop (regression) | –             | 45/45 (seeds 700001+); a "Modify search" link was added to the hotel results page so the early-submit recovery works there too (previously 2/24 hotel episodes died with "wrong results and no way back") |

## Training data (noise 0.12, 2 workers x 220 episodes per kind, seeds 300001… in steps of 20000)
Raw: `finetune/out/gym_raw3/g_<kind>_<seed0>.jsonl`; clean (episodes with a task_done case): `finetune/out/gym/clean3_<kind>_<seed0>.jsonl`.

| kind       | seed ranges          | clean episodes | cases  | failures during generation |
|------------|----------------------|----------------|--------|----------------------------|
| car        | 300001, 320001       | 431 / 440      | 20 501 | 9: covered date trigger / dropdown option (fixed afterwards, see above), 3 StalePage |
| restaurant | 340001, 360001       | 440 / 440      | 16 517 | 0 |
| signup     | 380001, 400001       | 439 / 440      | 16 426 | 1 (covered dropdown option) |
| filter     | 420001, 440001       | 440 / 440      |  6 973 | 0 |
| total      |                      | 1 750          | 60 417 | |

Cases per episode: car ~47, restaurant ~37, signup ~37, filter ~16 (each gold step yields a full-goal case and a sub-goal case,
plus a DONE case per completed sub-goal and one task_done case).  Add `$O/gym/clean3_*.jsonl` to the `cat $O/gym/*.jsonl` step of
the training script (it already globs the whole directory).

## Files touched
`spec.py` (car, restaurant, signup, filter_ generators + helpers `slug`, `person`, `staged`, `car_list`, `filter_items`,
`apply_listing`), `widgets.js` (toggle, text, slots, multi, date "selects", initial values for Back navigation, placeholder
selects, `error`/`empty`/`serialize`), `app.py` (form_page generalised with `show_if`; `car_results`, `stage_page`, `done_page`,
`filter_page`, `filter_product_page`; hotel results "Modify search"), `expert.py` (kind-generic `same`, `satisfied`, `wrong_results`,
`recover`, `early_submit`; `field_step` shared by forms and wizards; `filter_step`; public signatures unchanged), `eval_gym.py`
(`success()` for the four kinds), `clean3.py` (new).