Small models,
explicit constraints.

Start with a space worth exploring.

Signal began with a visual constraint: a small first-person world rendered in a restrained palette, soft light and ordered dithering. We kept that renderer while developing a short sequence: discover a terminal, change the space, then reconstruct and cross suspended architecture.

The arrival, landmarks, interactions and impossible return are authored. Conventional code handles rendering, movement and puzzle correctness. Machine learning has two bounded jobs: propose an observation-room layout and rank arrangements of architectural pieces.

Better examples helped. Rules still set a strong baseline.

The existing room network has 40,994 parameters. We retrained the same architecture on 800 variations across eight authored composition families, using 192 separate same-family validation rooms for checkpoint selection. Four excluded families supplied 256 final test rooms.

Examples encode purpose, required routes, clear space and focal views. The evaluation checks those properties rather than treating every connected floor plan as equally useful.

The retrained model improved on the earlier model under this protocol. It did not beat procedural generation. The final game’s consistency comes from an engineered policy around the model, including bounded rejection and fallback.

Inspect the full room comparison and rendered examples.

Reconstruct a crossing, then test it with the real physics.

A terminal proposes eight arrangements of authored platforms, fins, columns and apertures. Before selection, a worker runs the same movement controller used by the game: early and late takeoffs, the side-terrace detour, and checkpoint restarts.

Failed candidates are rejected. The learned selector and additive-rule baseline receive the same verified pool. Comparing selectors reuses that pool; reconstructing advances the seed. The model selects an arrangement—it does not invent a jump mechanic or decide whether a landing succeeds.

The ranker is a 20 → 16 tanh → 1 network with 353 learned parameters. Its weights and normalization values occupy 1,572 bytes. It learns synthetic preferences from an explicit authored composition rubric.

These are rubric-approximation results. Pairs within a pool are dependent, and no independent human preference or enjoyment study was conducted. The full rubric is itself inexpensive and defines the labels perfectly; this experiment establishes no advantage over running that rubric directly.

Inspect the crossing comparison, protocol and per-family results.

Keep inference small and failure visible.

The two active learned artifacts total 165,548 bytes. A compact Float32 JavaScript implementation runs in workers; the current experience adds no general ML inference runtime. The original WASM prototype is retained separately for comparison and is not fetched by the current game.

Missing or corrupt ranker weights trigger the labelled rule selector. Worker failure or timeout uses an authored reference checked with the same physics. Generation has a fixed proposal and verification budget. Sound is optional, reduced motion is respected, and missed jumps return to the saved platform.

Small weights do not guarantee a fast complete experience. Real-device rendering, page loading and input behavior still need their own measurements. Simulated physics rates and Node timings are not phone frame rates.

What we would test next.

Our next useful evidence would be human ratings of composition and a short play study: can players recognize the terminal, understand the reconstruction and complete the crossing? We would compare the learned selector with the full rubric and simpler authored alternatives using the same candidate pools.

Model size is not the immediate constraint. The more useful work is improving the examples, specifying the decision the model owns, and measuring whether it contributes something beyond an inexpensive rule.

Run the experiment.

The release includes source, trained artifacts, authored data generators, manifests, reports and reproducible checks. Training can be repeated separately; playing and checking the release requires no paid inference service.