README
solvi arcade (runs in your browser)
Game agents built with solvi from small Python functions, a few rules and hard checks. They play well, and every move comes with its reason: the rule's inputs, the facts computed this turn, and which hard check fired, if any. No models, and no server: the page runs Python in the visitor's browser with Gradio-Lite (Pyodide, Python compiled to WebAssembly). solvi is pure Python on numpy/scipy, so it installs from PyPI straight into the browser.
The first visit downloads the Python runtime with numpy, scipy, Gradio and solvi (about 20–30 MB, once; later visits use the browser cache). Write your own decision task in the solvi playground.
| Tab | What you see |
|---|---|
| ❌⭕ XO (tic-tac-toe) | The catalog from examples/05_tic_tac_toe.py (vendored), with an extra reason rule that names the move (win, block, fork, forcing, center ...). A "you can't beat it" counter and an "explain the position" view. |
| 👻 Maze | Pac is a catalog of BFS functions plus one scoring rule. The hard check greedy_is_safe blocks any step into a ghost's reach while a safer move exists. Turn the check off, take over with the arrows, run the benchmark in the browser. |
| 🃏 Fusion | Fuse two weapon cards with BLEND or DOMINANT. A balance catalog clamps slow, lifesteal and cooldown (hard checks) and rescales damage when the power is out of band (soft check). |
| 💣 Mines | A minesweeper solver that explains every move and outlines the evidence cells; hard check: never guess while a certain-safe cell is known. 97% wins on 9×9, 84% on 16×16 (100 seeded games); 0 wrong "certain" moves. With a quick risk estimate and the check off: 5% on 16×16, with the check on: 82%. |
| 🔮 20 Q | 20 questions over 145 animals: the strategist asks the question with the highest expected information gain and says why. Truthful answers: all 145 guessed first try in 10 questions on average; with 10% wrong answers, 93% within three guesses. A hard check forbids guessing below 80% confidence. Teach it a new animal in 0.3 ms. |
| 🎭 Mafia | A detective bot cites concrete events as evidence and never accuses a player it verified innocent (0 violations in 1 000 games). Town wins 65.6% with the bot vs 29% with a random voter. Watch it, or play with it as your advisor. |
| 🕵️ Hack | Try to forge a solvi decision trace: edit a value, fix its hash, rebuild the whole chain, change an input. Replay and a published receipt catch every level and say where and why. |
| 🏟️ Arena | A tournament of maze bots (built-in and yours) with a leaderboard, replays with the reason for every tick, "why did my bot lose?", and a switch that enforces the hard safety check on every bot: Cornerer goes from 532 to 815 points. |
| 🧭 Agent and knowledge (1.0) | solvi 1.0's environment agent (solvi.Agent) and its memory (solvi.Knowledge) in a small crafting world (8 places, six goals). Run 1 explores with an empty memory: 61 steps, all decided by System 2; run 2 in the same world: 12 steps, 11 by System 1; a new world keeps the learned rules and re-learns the map. The knowledge report (items, sources, what was refuted or retracted, what each action needs), a retraction of a person's fact with what was derived from it (the store's fingerprint equals the one rebuilt without it), every decision replayed. Then protection vs justified risk (RiskBudget) on a world whose iron lies across a bridge that breaks one time in three, behind a written gate that holds in both. |
| 🛠️ Your bot | Edit the maze agent's scoring function and run 20 seeded games against the default bot, in your own browser. |
Maze benchmark (100 seeded games, same scoring rule and ghosts)
| agent | games cleared | avg score | dots eaten | lives lost (of 3) |
|---|---|---|---|---|
| hard safety check ON | 59% | 764 | 94% | 1.95 |
| hard safety check OFF | 4% | 531 | 76% | 2.96 |
Native Python computes this in about 7 s on one CPU core; the browser, on a button click, in about 20 s, with the same numbers.
Layout
index.html: loads@gradio/lite@5.45.0from jsDelivr, lists the requirement (solvi==1.0.0, bumped with each release) and mountsapp.pyandgames/*.pyby URL.app.py: the Gradio 5 UI (theme and CSS ingr.Blocks(...)).games/tictactoe.py,games/maze.py,games/fusion.py: pure game logic and solvi catalogs, the same as in the server version of this Space.games/_ttt_catalog.pyis a vendored copy of the tic-tac-toe example catalog.games/crafting.pyandtabs/agent.py: the "Agent and knowledge (1.0)" tab — the toy crafting world ofexamples/25_environment_agent.py(vendored, with a step log for the page) and its UI; the agent and its memory live in agr.State, so each visitor has their own.games/explain.py: renders a solviResponseas a "why" card.games/sandbox.py: runs the visitor's bot code in-process, in a fresh namespace, under asys.settraceguard: a call that runs more than 200,000 lines of the bot's code (an infinite loop), or a run longer than 30 s, is stopped. It protects against runaway Python loops, not against a loop inside C code, and it is not a security sandbox (the code runs in the visitor's own browser tab).
Differences from the server version (../arcade)
- No threads or subprocesses in Pyodide: the maze benchmark runs on a button click (20 or 100 games) instead of at startup, and the bot runs in-process with the guard above instead of a resource-limited subprocess.
- The tic-tac-toe counter is per tab (there is no shared server to count everyone's games).
- "Play 50 ticks" uses an async generator with
asyncio.sleep(a blockingtime.sleepwould stall the worker).
Workaround in index.html
Gradio-Lite 5.45.0 (the latest release) installs gradio 5.45 with micropip, whose resolver is greedy. With today's PyPI
it picks huggingface-hub 1.x/2.x for gradio_client and then fails on gradio's huggingface-hub<1.0 pin (also
anyio<5, ...), so the app does not start. index.html wraps the web worker and filters PyPI's simple index to files
uploaded before the Gradio-Lite release (2025-09-10), except solvi. If a future Gradio-Lite release fixes this, the
wrapper script can be removed.
Run locally
cd solvi/spaces/arcade-lite
python -m http.server 8080 # then open http://localhost:8080
Similar repos
README
solvi arcade (runs in your browser)
Game agents built with solvi from small Python functions, a few rules and hard checks. They play well, and every move comes with its reason: the rule's inputs, the facts computed this turn, and which hard check fired, if any. No models, and no server: the page runs Python in the visitor's browser with Gradio-Lite (Pyodide, Python compiled to WebAssembly). solvi is pure Python on numpy/scipy, so it installs from PyPI straight into the browser.
The first visit downloads the Python runtime with numpy, scipy, Gradio and solvi (about 20–30 MB, once; later visits use the browser cache). Write your own decision task in the solvi playground.
| Tab | What you see |
|---|---|
| ❌⭕ XO (tic-tac-toe) | The catalog from examples/05_tic_tac_toe.py (vendored), with an extra reason rule that names the move (win, block, fork, forcing, center ...). A "you can't beat it" counter and an "explain the position" view. |
| 👻 Maze | Pac is a catalog of BFS functions plus one scoring rule. The hard check greedy_is_safe blocks any step into a ghost's reach while a safer move exists. Turn the check off, take over with the arrows, run the benchmark in the browser. |
| 🃏 Fusion | Fuse two weapon cards with BLEND or DOMINANT. A balance catalog clamps slow, lifesteal and cooldown (hard checks) and rescales damage when the power is out of band (soft check). |
| 💣 Mines | A minesweeper solver that explains every move and outlines the evidence cells; hard check: never guess while a certain-safe cell is known. 97% wins on 9×9, 84% on 16×16 (100 seeded games); 0 wrong "certain" moves. With a quick risk estimate and the check off: 5% on 16×16, with the check on: 82%. |
| 🔮 20 Q | 20 questions over 145 animals: the strategist asks the question with the highest expected information gain and says why. Truthful answers: all 145 guessed first try in 10 questions on average; with 10% wrong answers, 93% within three guesses. A hard check forbids guessing below 80% confidence. Teach it a new animal in 0.3 ms. |
| 🎭 Mafia | A detective bot cites concrete events as evidence and never accuses a player it verified innocent (0 violations in 1 000 games). Town wins 65.6% with the bot vs 29% with a random voter. Watch it, or play with it as your advisor. |
| 🕵️ Hack | Try to forge a solvi decision trace: edit a value, fix its hash, rebuild the whole chain, change an input. Replay and a published receipt catch every level and say where and why. |
| 🏟️ Arena | A tournament of maze bots (built-in and yours) with a leaderboard, replays with the reason for every tick, "why did my bot lose?", and a switch that enforces the hard safety check on every bot: Cornerer goes from 532 to 815 points. |
| 🧭 Agent and knowledge (1.0) | solvi 1.0's environment agent (solvi.Agent) and its memory (solvi.Knowledge) in a small crafting world (8 places, six goals). Run 1 explores with an empty memory: 61 steps, all decided by System 2; run 2 in the same world: 12 steps, 11 by System 1; a new world keeps the learned rules and re-learns the map. The knowledge report (items, sources, what was refuted or retracted, what each action needs), a retraction of a person's fact with what was derived from it (the store's fingerprint equals the one rebuilt without it), every decision replayed. Then protection vs justified risk (RiskBudget) on a world whose iron lies across a bridge that breaks one time in three, behind a written gate that holds in both. |
| 🛠️ Your bot | Edit the maze agent's scoring function and run 20 seeded games against the default bot, in your own browser. |
Maze benchmark (100 seeded games, same scoring rule and ghosts)
| agent | games cleared | avg score | dots eaten | lives lost (of 3) |
|---|---|---|---|---|
| hard safety check ON | 59% | 764 | 94% | 1.95 |
| hard safety check OFF | 4% | 531 | 76% | 2.96 |
Native Python computes this in about 7 s on one CPU core; the browser, on a button click, in about 20 s, with the same numbers.
Layout
index.html: loads@gradio/lite@5.45.0from jsDelivr, lists the requirement (solvi==1.0.0, bumped with each release) and mountsapp.pyandgames/*.pyby URL.app.py: the Gradio 5 UI (theme and CSS ingr.Blocks(...)).games/tictactoe.py,games/maze.py,games/fusion.py: pure game logic and solvi catalogs, the same as in the server version of this Space.games/_ttt_catalog.pyis a vendored copy of the tic-tac-toe example catalog.games/crafting.pyandtabs/agent.py: the "Agent and knowledge (1.0)" tab — the toy crafting world ofexamples/25_environment_agent.py(vendored, with a step log for the page) and its UI; the agent and its memory live in agr.State, so each visitor has their own.games/explain.py: renders a solviResponseas a "why" card.games/sandbox.py: runs the visitor's bot code in-process, in a fresh namespace, under asys.settraceguard: a call that runs more than 200,000 lines of the bot's code (an infinite loop), or a run longer than 30 s, is stopped. It protects against runaway Python loops, not against a loop inside C code, and it is not a security sandbox (the code runs in the visitor's own browser tab).
Differences from the server version (../arcade)
- No threads or subprocesses in Pyodide: the maze benchmark runs on a button click (20 or 100 games) instead of at startup, and the bot runs in-process with the guard above instead of a resource-limited subprocess.
- The tic-tac-toe counter is per tab (there is no shared server to count everyone's games).
- "Play 50 ticks" uses an async generator with
asyncio.sleep(a blockingtime.sleepwould stall the worker).
Workaround in index.html
Gradio-Lite 5.45.0 (the latest release) installs gradio 5.45 with micropip, whose resolver is greedy. With today's PyPI
it picks huggingface-hub 1.x/2.x for gradio_client and then fails on gradio's huggingface-hub<1.0 pin (also
anyio<5, ...), so the app does not start. index.html wraps the web worker and filters PyPI's simple index to files
uploaded before the Gradio-Lite release (2025-09-10), except solvi. If a future Gradio-Lite release fixes this, the
wrapper script can be removed.
Run locally
cd solvi/spaces/arcade-lite
python -m http.server 8080 # then open http://localhost:8080