A few things I jotted while building.
- Core idea from day one: the sim is real physics, so you can't guess. Do the AP-C math or you don't win.
- First playable version was a build challenge, just stack track until the cart reaches the goal. Fun, but you could beat every level without doing one line of physics. Scrapped it. Locked physics objects plus a tight end-state target instead, so you have to calculate.
- Spline overshoot at piece seams threw fake g-spikes, hundreds of g out of nowhere. Fix: sample every piece at one shared arc-length spacing.
- Loops spiked curvature at the bottom seam and apex, so a cart would stall or fall off a loop it should've cleared. Fix: sub in the loop's true 1/R inside the loop span. Clothoid loop is the real fix, left it for later.
- One spring level was math-correct but literally unwinnable. The stiffness amplified compression about 28x, so the win window was narrower than the slider step. No slider position won. Widened the window and made the slider and the 3D handle share one step size.
- That one taught me math-correct isn't the same as playable. The math audits all passed and the levels still broke in the browser. So I started testing at the winning value and the ones just on either side of it.
- FRQs overshot at first, like an 8 when I wanted a 6. Read real released AP-C exams and dialed every level back.
- Built with AI help. The calls were mine, real sim over animation, which primitives, the force shapes, the exam calibration. The AI did a lot of the TypeScript.