Games the engine plays
Every bot names the mechanism that moved it; every leaderboard lives only in your tab; every claim about what the engine decides is checkable in the trace panel. Pick a scenario below — or grab the keyboard.
Maze Solver & Dungeon Crawler — what you are actually watching
Every step the bot takes is a decision made the same way the runtime makes
them: cheapest reliable mechanism first. There is no neural
network in this maze. The scout checks a rule (fight is forced, exit is
one step away, a monster is too close), then falls back to breadth-first
searches over only what it has seen — exit route, loot, chest,
nearest unexplored frontier. When two frontier routes cost exactly the
same, and you have the engine checkbox on, that genuine tie becomes a real
choice request to the same WASM engine the panel below runs.
The chip next to the canvas names the rung that decided every single step.
Two map types — classic corridors, or rooms-and-hallways dungeons — and
two ways to see: a full disc, or a facing torch cone where turning matters.
How to play: pick play (WASD / arrows, or the d-pad), or versus to race a glowing ghost of the bot that runs its own identical maze with its own fog — it gains no information from you, and you none from it. Loot piles score, chests open in one second (coins 0–5… or something with teeth), monsters chase what they can see along a ray and cost you a two-second static fight on contact, and the optional timer ends the run when it expires — unless you switch on avoid monsters, which routes the walker around every threat it has actually seen (and nothing it hasn't; the fog stays honest). Found everything? The exit is the one opening in the outer wall — usually as far from the middle as it can be.
Why it is here: to show the methodology moving. Real maze games hand the AI omniscience; this one hands the bot the same darkness you get, and you can audit every choice it made in the trace panel. The leaderboard is session-only — nothing you do here is stored, sent, or scored anywhere but your tab.
Honest footnotes: vision is generalized Bresenham line-of-sight per cell,
walls stop the ray. The maze itself, the bot's exploration ladder, and the
scoring arithmetic are plain, tested JavaScript (site/maze-core.js,
70 headless tests) — only the tie-breaks go to WASM. God view is a
spectator cheat: any run finished while it is on is marked
assisted and scores nothing. Thinking about more than a demo —
progression, real combat, multiplayer — is sketched in
docs/planning/maze-solver/STANDALONE-GAME.md.
The game library
Every game runs on the same shared runtime — the Decisions SDK — under the same honesty rules: rules that never ask, scored candidates with the weights on screen, genuine ties to the engine, and refusals that fall to a named rung on the record. Filter by kind or search; click a card for the details.
nothing matches that — clear the search or pick another kind.
MicroWorld
The flagship: one tiny world where a whole colony thinks out loud.
▶ Play click the card for details →One tiny deterministic world where dozens of creatures each re-score ten actions every ~0.75 s — eat, drink, flee, fight, explore, rest, gather, go home, follow, ignore — from their own meters and an eight-cell view. Near-ties go to the engine over just the tied candidates; the fox, births, deaths and stores are rules. Click a creature to pin its decision table to the trace line.
watch a colony think →Choose Your Door
Three-to-five doors, one pick, and the whole decision on the table.
▶ Play click the card for details →One decision, fully shown: three-to-five doors, one character, and the inputs, candidate scores, confidence, and trace behind the pick. Personalities are policy, not code paths — coward, greedy, brave, and steady are the same engine with different weights, and balancing the doors makes the engine abstain on the record.
look behind the decision →RPSLS++
Rock-paper-scissors-lizard-spock on a move graph you can extend mid-match.
▶ Play click the card for details →Add a move mid-match and watch the graph reorient. The bot blends frequency and markov tables over your history into expected values, explores just enough to stay unreadable, and hands genuinely close calls to the engine with your history table on the wire.
try to out-pattern it →The Prisoner's Dilemma
The engine agent plays tit-for-tat, grim trigger, and pavlov at the table.
▶ Play click the card for details →Tit-for-tat, grim trigger, pavlov and friends on the other side, a hard per-match ask budget, refusals falling to the mood rule on the record, and transmission noise that makes opponents retaliate against moves that never happened. Editable R/S/T/P matrix and a seeded round-robin tournament.
sit down at the table →Spacecraft Emergency
A stricken ship, six scored actions, and an engine that can abstain — then you fly.
▶ Play click the card for details →The flight computer scores six actions, gates the hard ones off as rules, and — when telemetry conflicts or scores dead-heat — abstains: the sim freezes and you decide as the captain. An engine refusal escalates the same way. The refusal-is-a-feature demo, as a game.
open the cockpit →Ask the Engine
Every answer is a real engine choice request with its confidence and trace.
▶ Play click the card for details →A quiz where every answer is a real engine choice request — pick, confidence report, execution trace — scored honestly against the keyed deck. Speed-run, audit, compete; bring your own questions in the validated JSON deck editor.
Ant Colony
A thousand deciding ants, and a trail network nobody designed.
▶ Play click the card for details →Up to a thousand ants deciding at one hertz — RETURN and GATHER are rules that never reach the engine, the rest is scored and near-ties go to the engine under a batching budget. A trail network nobody designed emerges from the pheromones; the decisions/s counter keeps the scale honest.
start the colony →AI Civilization in 60 Seconds
A tiny city of deciding citizens; inject a shortage and watch the cascade.
▶ Play click the card for details →Sleep, urgent eating and childbirth are rules that never reach the engine; farm vs work vs build vs trade vs explore is scored, and near-ties go to the engine under a batching budget. Inject a shortage and watch the cascade: hunger up, farming up, building down, homelessness growing.
found the city →Traffic Simulator
Every car is a rule; the signal lights are the deciders.
▶ Play click the card for details →Stop, go, wait, turn, reroute are rules. The signal lights are the decision agents: near-tie phase calls go to the engine over keep/switch under a city-wide ask budget, and emergency preemption is a rule that is never a question. Throughput and wait percentiles make the congestion slider honest.
open the city →Tower Defense
Enemies and towers both decide out loud, on seeded deterministic waves.
▶ Play click the card for details →Enemies run the rule ladder (advance, divert, retreat, regroup, sabotage) while towers score six targeting policies per shot and hand genuine near-ties to the engine. Seeded deterministic waves, a decisions/s counter, and click-to-take-over tower policies.
raise the towers →Maze Solver & Dungeon Crawler
A dungeon crawler where the search sees only what it has seen.
▶ Play click the card for details →Classic corridors or room-and-hallway dungeons with two vision models, monsters with a real line-of-sight ray, loot to gather and an exit to find. The solver solves only from what has been seen, genuine ties go to the engine, and the god view is a declared cheat that scores nothing.
enter the maze →Rubik's Cube
A real 3D cube with a layer-by-layer solver that shows its plan step by step.
▶ Play click the card for details →Drag to orbit, turn faces by hand, scramble, and ask for the plan: a deterministic layer-by-layer solver where every move is simulation-verified before it plays. The genuine ties (which equal-length setup) go to the engine, and each step's rung chip says rule, roster, or engine.
scramble it and ask for the plan →Game of Life
Conway's automaton with variants — and the engine proposes the next seed.
▶ Play click the card for details →Rule variants and a pattern library: the engine reads the board (still, oscillator, spaceship, exploding) and proposes the next interesting seed; ties on equal-interest seeds escalate to it. Paint cells while it runs.
run life →Pac-maze
Pellets, power pellets, and six ghost personalities deciding every turn.
▶ Play click the card for details →The maze generator reused as an arcade: pellets, power pellets, and six ghost personalities whose chase/ambush/flee weight vectors decide every turn — genuine ties escalate to the engine, fright and eaten-eyes are rules you can watch run. Play yourself or flip on the bot.
enter the maze →Pong
Two engine-driven paddles; edit the reward weights and watch the incentives.
▶ Play click the card for details →The paddles score candidate targets — intercept, edge-spin, rest-center, mirror-the-opponent — against five reward weights you edit live, and watch the incentives play out. Genuine ties escalate to the engine.
play pong →Apex Line
A 2.5D racer where AI opponents score four candidate lines every stretch.
▶ Play click the card for details →Seeded, hilly tracks: AI racers score apex, inside, outside, slipstream with personality weights; genuine ties go to the engine, and your suggested-line chip shows which rung answered it. Traffic is a rule.
take the flag →GridWars
A twin-stick neon arena where every geometry decides out loud on a warping grid.
▶ Play click the card for details →Diamonds jink off your fire line, squares flee and split into three, rockets charge and brake, snakes eat your trail, wells swallow fire until they burst. Every enemy picks a labelled tag five times a second and the state panel counts each one; waves are seeded and deterministic, and the engine names which geometry enters next.
enter the grid →Sparring Partner
An opponent that keeps a decayed histogram of your habits and predicts your next attack.
▶ Play click the card for details →A side-view fighter whose opponent predicts from your recorded moves per situation and escalates near-tie answers to the engine — with your habits in the question. Inspect and reset the learning anytime.
enter the ring →One runtime under all of it
The games share a small SDK — seeded RNG, fixed-timestep loop, config
normalization, rubric scoring, the engine adapter, and the UI shell — so a
new game is mostly its own map, rules, and rendering. The same SDK doubles
as the honest way to use the engine outside games: deterministic seeds as
test fixtures, verdict gates in CI, and plain typed decisions with
calibrated confidence anywhere a choice needs an audit trail. The maze's
methodology is documented end to end in
docs/planning/playground/DECISIONS-SDK-PLAN.md.
The Decide console lives on its own page
The full request → decision → trace loop — the real WASM engine, honest refusals, byte-checkable traces, a declarative graph runner — with room to breathe. These presets load and decide the moment the console boots.
Prefer the terminal? The same decisions from
cargo run --bin opencodifier-cli -- decide, or serve the HTTP API and hit
the documented endpoints.