Every entry here is a moment I was about to be wrong in public.

Refinement showed the residual falls by exactly 4.00x per halving of the timestep, which is second-order convergence. I derived the closed-form law it should follow and it matched to within 1%: predicted 3.95e-3 against 3.9625e-3 measured. The solver was right and my threshold was a number with no theoretical basis.

Three separate faults were live at once: a mesh URI using a scheme the consumer could not resolve, a mesh pointing at a file that was never modelled, and link and joint names that collide. The second was hidden by the first, because an unresolvable scheme is silently skipped by the existence check.

A boolean intersection test, asking whether the two solids actually share volume, found the hinges had been interfering the whole time.

It was separating them by magnitude alone. Real normalised features span about -2.78 to +5.47, and a bare Tanh output only reaches -1 to +1, under 40% of that range. The critic never had to look at how a swarm moves.

I measured when collapse actually arrived. With them, epoch 55. Without them, epoch 145. On a 560-sample dataset of 10 features by 99 timesteps they were a clear net negative.

A matched-capacity null model that sees only the input, registered as the primary comparison before any results, beat every trust signal tested.

Under a detection rule fixed before I looked at the results, weight_norm fired on 0 of 8 seeds. The Fourier signal, under the same rule, led on 8 of 8 by a median of about 2,600 steps.

Under the same pre-registered rule it fired on 5 of 8 seeds, but after grokking had already happened. An early warning that arrives late is not an early warning.

A test aimed only at the frame convention, which axis set the angular rates are measured in, came back 187 times off.

A test at a 1e-10 tolerance in float64 failed. torch.fft.rfftfreq follows the default dtype, so it handed back float32 wavenumbers even for a float64 field, and squaring them cost two significant digits.

Running the rover traversability model over the same cells found 888 that the rover cannot cross and the drone can fly over without noticing. 427 of them fail on step height alone, which a slope-only check cannot see.

I checked whether the estimated sensor biases converged to the true biases I had injected into the simulation. They did.

Comparing the commanded speed against the intended survey speed showed the lookahead controller was flying 3.4 times too fast.

One call to the clock anywhere in the engine would make every build different, and no ordinary test would catch it because nothing else would change. A test parses the engine source into a syntax tree and asserts that time, datetime, calendar, random and uuid are never imported.

Both files were sitting in public/. The check resolved its directory from import.meta.url, which points into dist/ once Vite has bundled the module, so it was testing a path that never exists. The rendered HTML still said "Photo to add" with the photo in place.