The verified gallery
Most “AI makes games” demos ask you to take the output on faith. hayao’s don’t.
Every game in this repo ships a verify.ts suite that machine-proves its
claims — and you can run the proof yourself:
npm run verify # prove the whole portfolio (solver + determinism + gates)
npm run verify -- lumen sokoban # just these — scopes every stage, not only the feel audit
npm run eval # scorecard: verified rate + proof coverage per game
npm run eval -- lumen sokoban # just theseThat is the honest version of an AI-first gallery: not “trust me, an agent built this,” but “an agent built this, a human directed it, and here is the machine proof that it works.” The proof is the marketing.
What “verified” means here
Each game asserts truth on up to six independent proof channels. npm run eval
reports which ones each game exercises:
| Channel | What it proves | How |
|---|---|---|
| winnable | every level/board is beatable | solve() / assertSolvable — a real search, not a vibe |
| determinism | same inputs → same world, bit for bit | checkDeterministic + golden replay hashes |
| ramp | the difficulty curve escalates without cliffs | assertRamp / rampIssues over the campaign |
| feel | grace windows, readable avatar, telegraphed threats | the feel gates (salienceIssues, forgivenessIssues, …) |
| generated | content was composed, not hand-authored | generateLevels / composeCampaign |
| code-as-art | it’s drawn with gradients/shapes, not placeholder squares | the art/ + paint toolkit |
The flagship: Lumen
Prompt-shaped as: “a light-routing chain puzzle that always serves a fair, winnable board, build with hayao.”
What ships: a light-routing puzzle — drop a spark, its four beams cascade through prisms — served on an Elo curve so every board is matched to the player, and not one board was drawn by hand. Each board is expressed as a seed + params and kept only if the solver proves it winnable within its spark budget; that same solver estimate is what the rating loop matches on. Boards re-derive identically on every machine.
The proof (npm run verify -- lumen):
- ✓ every generated board across four rating bands is solver-proven winnable (24 boards: 4 bands × 6 seeds)
- ✓ the difficulty estimate ramps monotonically — harder bands average higher ratings
- ✓ a scripted drag-drop playthrough lights every prism and banks the rating gain
- ✓ golden replay hash pinned; sim deterministic across runs
It leads on winnable + generated + determinism — the reference for generate-and-prove content that can’t ship a board it hasn’t proven. Source: examples/lumen.
The rest of the portfolio
Every other game under examples/ is verified too — one per major 2D
genre, plus physics and netplay showcases. They lead on different channels (a
platformer proves reachability + feel; a horde game proves determinism under load; a
tactics game proves counter-play), and the playable builds live at
hayao.dev. Run npm run eval for the current, honest scorecard
rather than a screenshot you have to trust.