Blind driver on a random track · controlled proof, replayed step by step
Same track, four drivers. At every turn, the dashed gold ray is the driver's own guess — drawn from
where it's standing, before the real turn is revealed. The road then extends in the real direction,
green if the guess matched, red if it didn't. Nothing here is live-simulated; every driver is the
actual champion genome trained in evolution/blind_driver_track_v1.py, replayed against one
fresh, fixed track per variant. Toggle "drive by its own guesses" to switch from
accuracy-scoring to a real physical drive: the car actually steers by its own belief instead of staying
glued to the real road, and can wander within a tunable lane margin before it's flagged off-track.
The 5 "Primitive" tabs are a different question entirely: those tracks have REAL structure (5 named,
recurring patterns, not random turns), and the driver has to recognize which one is playing from an
8-step lookback — each one shown here is a track it never saw during training. Turning on lane-margin
mode for those tabs reveals a real limitation, not a bug: high per-step turn accuracy (85-92%) still
drifts badly once actually driving, because a single missed turn creates a permanent heading offset
nothing ever corrects. The 5 "Combined" tabs fix exactly that — same lookback, PLUS real drift feedback,
together: on-lane time jumps from 14.5% (lookback alone) to 86.8% overall.