A tiny model.
A crossing that holds.
Terminal-driven reconstruction combines authored architectural pieces, a learned composition selector and the game's actual movement physics. The palette, concrete textures and ordered dither stay shared with the game.
Loading the frozen comparison…
What this establishes: the tiny ranker better approximates an authored composition rubric than the fixed additive baseline. The full rubric itself is an inexpensive, perfect label oracle. This does not establish an advantage over that full rule implementation, independent aesthetic preference or player enjoyment.
Compare the selected spaces
Both selectors receive the same eight physics-verified candidates in each held-out condition. Preview the actual course geometry using the game's renderer.
Held-out results
| Metric | Learned ranker | Additive rules | Full rubric reference |
|---|
Pairs with rubric-rating differences below 0.025 are excluded from pair accuracy. Top-choice agreement and regret are calculated per eight-candidate condition. Regret is the best available rubric rating minus the chosen rating; smaller is better. Pairs within a condition are dependent observations.
| Excluded motif family | Learned pair agreement | Rule pair agreement | Learned regret | Rule regret |
|---|
Physics and runtime
Runtime selection proposes eight arrangements. Each accepted candidate passes six production-controller trials: early/late takeoffs at 30/60 simulated FPS, the side-terrace detour and a middle-checkpoint restart. Extended audits add 20/120 FPS timings. Wide gaps and blocked routes are rejected in separate negative-fixture checks. The landing physics and checkpoints remain deterministic game code.
The worker verifies the ranker checksum. Missing weights use the labelled rule selector; a worker failure uses a separately physics-checked authored reference. Geometry rises into view over 900 ms while movement is held. Reduced motion skips that transition. Falling feedback and checkpoint resets are preserved.
Training and evidence
20 plan/composition features → 16 tanh units → one score. Training uses within-condition pairwise preference loss with rating anchoring. There are 2,048 training plans across eight motif families and 512 disjoint validation plans from those same families. The checkpoint is chosen using validation MSE. The final 640 plans come from four wholly excluded motif families.
Labels are generated by a versioned authored rubric covering framing, balance, silhouette, elevation, sightlines and discovery placement. No independent human preference labels were collected. The model ranks combinations of authored pieces; it does not generate raw geometry, movement rules or artwork.
Complete per-condition results · Training manifest · Automated checks