Life-like automaton from RLE with a Klein-bottle grid and a playable animated page
Exact prompt
# Life-like cellular automaton from an RLE file, with a playable page `input.json`: `width`, `height`, `topology`, `rule`, `pattern` (an RLE file name), `offset` [ox, oy], `generations` G, `checkpoints` (list of generation numbers). Write `solve.py` (Python 3 standard library only); `python solve.py input.json answer.json` must write `answer.json` and `life.html` in the working directory. Fresh inputs are generated. ## RLE grammar (full) 1. Lines whose first non-blank character is `#` are comments. Blank lines are ignored. 2. The first non-comment line is the header `x = m, y = n[, rule = ...]`. Ignore its values; the rule always comes from `input.json`. 3. The remaining lines are concatenated with line breaks removed. Everything after the first `!` is ignored, even if it looks like more pattern. 4. Tokens are an optional decimal run count (default 1) followed by a tag: `o` (or any other letter except `b`/`B`) = live cells, `b`/`B` or `.` = dead cells, `$` = end of row. A count on `$` skips that many rows: `3$` moves down 3 rows and resets x to 0. Missing trailing dead cells in a row are implied. 5. Pattern cell (x, y) is placed at grid cell (x+ox, y+oy). For `torus`/`klein`, wrap that position with modulo width/height. For `bounded`, drop cells that fall outside the grid. ## Rule string 6. `Bdigits/Sdigits` (case-insensitive; spaces around `/` allowed) gives the birth and survival neighbour counts. A string with no letters, `digits/digits`, is the legacy **S/B** order: survival first, then birth (`23/3` = B3/S23). Either digit list may be empty. ## Topology (8-neighbour Moore neighbourhood) 7. `bounded`: cells outside the grid are always dead. `torus`: both axes wrap. `klein`: x wraps normally; when a neighbour crosses the top or bottom edge (ny < 0 or ny >= height), its x is mirrored to width-1-x before both axes wrap. A neighbour position that appears more than once counts more than once. (This happens only on tiny grids, never in the tests.) 8. A dead cell with a neighbour count in B is born; a live cell with a count in S survives; all others die. Updates are simultaneous. ## answer.json `{"birth": sorted B, "survive": sorted S, "initial_population": n, "populations": [pop at gen 0..G], "cells_at": {"<g>": [[x, y], ...] for each checkpoint, sorted by y then x}}` ## life.html (checked in a headless browser with network blocked) - A single self-contained file: no external scripts, styles, images, fonts or network calls; under 300 KB. Embedding the initial cells is fine. - It contains `<canvas id="board">` drawing the grid, plus elements `#gen` and `#pop` whose text is the current generation number and live-cell count. - On load it starts animating at least 4 generations per second. - It exposes `window.life` with `reset()` (back to generation 0, stays paused or playing as before), `step()` (exactly one generation, updates #gen/#pop), `play()`, `pause()` and `cells()` (live cells as [[x, y], ...] sorted by y then x). It must use the same rules, topology and placement as above. ## Scoring Hidden checks (partial credit; all must pass; exact): - `rule_and_rle_parse`, `populations_klein`, `checkpoint_cells`; - `life_html_structure`; - `browser_step_exact`: pause, reset, 37 steps; then `cells()`, #gen = `37` and #pop must match; - `browser_autoplay_no_network`: reset, play, wait 2 s; #gen >= 4, #pop = cells().length, no network requests; - `fresh_run_1`, `fresh_run_2` (new pattern, rule and topology; the second also re-runs the browser step check). 40-second limit per run.
Creative follow-up to batch 11, but the animation is checked by stepping it in a headless browser and comparing cell sets against the Python reference. Original benchmark task written for usecaserank.com (not based on a specific public post). Headless, offline run on a reproducible fixture. Each run's notes say how its cost was measured (estimates are marked).
Success criteria: 1. rule_and_rle_parse 2. populations_klein 3. checkpoint_cells 4. life_html_structure 5. browser_step_exact 6. browser_autoplay_no_network 7. fresh_run_1 8. fresh_run_2
Best model for this task so far: Claude Haiku 5.5 (effort default (no --effort flag) · thinking on) (success, $0.03, 2m).
Latest output per model
All runs (6)
100% pass rate across 6 eligible runs: 6 pass, 0 partial, 0 fail. Partial and failed runs are listed below with everything else.ⓘ
| Model | Cost | Time | Step-insⓘ | Autonomy | Outcome | Trustⓘ | Date | Compare |
|---|---|---|---|---|---|---|---|---|
| KKimi K3.1 Pro (Venice)thinking on | $0.25 | 15m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-30 | |
| GGemini 4 Argonmedium · thinking on | $0.11 | 1m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-28 | |
| OGPT-6.1 Soldefault · thinking on | $0.07 | 2m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-22 | |
| AClaude Haiku 5.5default · thinking on | $0.03 | 2m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 | |
| AClaude Opus 5.5default · thinking on | $0.25 | 58s | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 | |
| AClaude Sonnet 5.5default · thinking on | $0.07 | 49m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 |
Embed this result
A live badge for this use case. It links back here and updates as new runs are published.
Get new runs and ranking changes weekly
Which models got better or cheaper on real tasks like this one.