Ghostline

Ghostline started as a small top-down time-trial racer. The real work is the system around it: deterministic physics, action-stream replays, reward-versioned training runs, and analytics that show why a policy improved or stalled.

The best policy runs a clean 3.948s lap on Track 1, past the human reference line. Every gain is backed by a saved run, so the progression can be checked, not taken on trust.

The core is a deterministic 2D simulator: the same action stream always produces the same lap, so every run is reproducible and every replay is exact.

Everything else hangs off that. An action stream feeds the replay viewer, the replay scores against a versioned reward, the reward drives training, and the analytics read the run back as sector traces and leaderboards.

The policy did not just get more training. It improved through a sequence of fixes, each one aimed at a specific failure.

Early policies stalled before finishing a lap, because a sparse go-fast reward was not enough signal. A valid 7.5s lap proved the sim, checkpoints, action stream, and training loop closed end to end. A warm-start time attack then reached the low 5s but exposed a ranking bug: the reward sometimes preferred a clean slow lap over a faster messy one. Matching the reward to the stopwatch fixed the objective and broke the policy under 5s.

The 4.1s plateau was information-limited: the model could see the next gate but not enough of the corner beyond it. Lookahead observations plus a higher discount factor restored reliable learning and broke the old 246-tick ceiling to reach the human benchmark. Wall-sensor training then held 3.972s across seeds, and fine-tuning cracked two more ticks to land at 3.948s.

Ghostline is deliberately a finished time-attack study, not an open-ended training chase. What matters is not just the lap time but that every step is explained and reproducible: why a policy stalled, why a reward version changed, and why the final lap holds up.