Brickpilot: Local Branches, Real Roads, Better Questions

The fork became a controlled research space for stop-creep tuning, steering handoffs, and smoother questions about what the system should do next.

Brickpilot started as a local name for the experimental fork. The name is personal, but the discipline around it is practical: separate branches for separate questions, visible version labels on the device, audits before install, and a clear rule that a successful neighborhood loop does not become a production release.

The stop-creep branch focused on the low-speed longitudinal problem. Early versions improved some signatures, especially premature re-acceleration while creeping behind a lead, while the bigger issue remained. In bumper-to-bumper traffic and at some stop approaches, the car could still slow too late, speed up too eagerly, or leave the driver to finish the stop. That matters. A pleasant highway stretch cannot compensate for a bad moment near a stopped car.

The steering branch taught a different lesson

The first steering-smoothness experiment produced a serious warning pattern: temporary steering-assist unavailable alerts, beeps, and lane-assist icon flicker. The raw analysis pointed toward an unsafe limbo state around MADS and Hyundai CAN-FD LFA handoff behavior, where lateral command continuity could persist while the main enabled state was not cleanly aligned.

That branch was quarantined. The next safe-fix iteration reverted the risky steering slew change and added conservative gating across gas, brake, manual override, disabled, and high-angle contexts. Later validation showed the big steering-assist warnings stayed away in that test, which was the right kind of boring.

Where the project is heading

The current questions are more interesting than the original ones. The project is no longer asking whether comma can work on the Tucson PHEV. It is asking how to separate model intent from controller execution, how to make low-speed behavior more decisive without making it harsh, and how to smooth steering handoffs without hiding real safety boundaries.

Keep supervised road testing boring and repeatable.
Treat driver takeovers and alerts as first-class data instead of anecdotes.
Use Brickpilot branches to create signal, then audit before trusting the signal.
Publish the story and the method, while keeping private logs and identifiers private.

That is the real fun of the project. The car is a moving laboratory, the fork is a notebook with wheels, and every drive either improves the behavior or improves the question.