Scenario infrastructure
Frozen scenario snapshots, survivability witnesses and deterministic schedules.
Research notebook / Gradient Gators
Better experiments.
Careful conclusions.
How reward design influences reinforcement-learning behavior and robustness. A five-person senior research project at the University of Florida.
Read the engineering case study Current stateAS OF 05 NOV 2026
Research in progress
The scenario and pilot infrastructure is in place. The next step is integrating the full reward comparison and evaluation protocol.
Training
PPO integration and an auditable pilot runner. The runner currently permits sparse-reward pilots.
Reward variants
Five reward calculators are implemented. This does not mean five completed experiments.
Automated checks
457 passed · 2 skipped · 17 warnings in the November 4 CI run.
SOURCE SNAPSHOT / 6b8e1ce · No final comparative findings published here.
The record
Engineering notes.
05 Nov 2026Pilot infrastructure
From a smoke test to an auditable pilot
What a short training run can establish—and what it cannot tell us about a learned policy.
05 Nov 2026Implemented
What a survivability witness actually proves
A connected map is only the beginning. A candidate path must survive the environment's mechanics.