Why we scrapped the third-generation Garde navigation stack.

After eleven months in the hangar we shelved a working autopilot — and rebuilt targeting from a two-page memo Kunato left taped to the coolant riser.

The third-generation stack shipped last winter with a benchmark I still cannot fault: 3.2 ms from sensor tick to thruster command, edge cases folded into a lookup the size of a paperback. On paper it was the cleanest thing Sidonia Engineering has released since the Type-17 refit. In practice, three pilots came back off the training range with the same complaint — the Garde felt correct, and none of them trusted it.

I spent two weeks last winter rewriting the acceleration curves in a sub-basement that smelled of coolant and old solder. What I found was not a bug. The stack was doing exactly what we had asked it to. The problem was that we had asked the wrong question, and no benchmark on the deck was going to tell us so.

§ 02The pilots were right, and the numbers were also right.

Kunato’s memo, taped above the riser in his usual block hand, read only: “A Garde is not an aircraft. Stop tuning it like one.” Six lines beneath it, a targeting model — no lookup, no lattice, one integrator and a small amount of shame. We tore out the third-generation stack the following morning and the hangar was quieter for it.