posts
- The engine got credit for the quant
open-jet beat llama.cpp 7 to 3 on a Zig agent task. On open-jet's own GGUF, llama.cpp went 7 of 9 too. Only the speed gap belonged to the engine.
- The pin template was for a different plug
Moving a Bluetooth-paired Tuya energy plug to OpenBeken without opening it, and the borrowed pin template that looked right and wasn't.
- My benchmark served the model at half speed
The launcher turned on MTP drafting for five models. The benchmark's own server script didn't, and its result rows had no field that could show it.
- My benchmark read my CLAUDE.md
An audit built to check sampler overrides found my OpenCode bench runs carrying my global ~/.claude/CLAUDE.md and skills: 53,483 characters of prompt.
- A null battery reading cleared the low-battery alarm
Building the Go observer that serves two power stations to stock upsmon: six ways the first version could fail, and the outage test that went through.
- The best-reviewed code came from the slowest Flash-Next arm
Stage 1 of the A9 agent study: the slowest Flash-Next arm reviewed best, the 62 GiB Q2_0 passed 6/6 at nearly 6x the rate. A sampler confound splits them.
- Zig needed context, not time
Four 27B local models kept hitting the one-hour wall on a Zig task. Four hours at 64k got 7 of 12 through the core tests. At 128k, 35 of 36 did, cleanly.
- The DJI's grid flag is labeled charge type
A DJI Power 2000 over plaintext BLE: the documented grid byte never changed, input watts read 0 on grid, and the link drops were a 420 ms timeout.
- Inventing the work vs promising what isn't there
Checking 71 local-model agent runs' claims against their own command logs: three models invent work, three promise atomicity they never built.
- Twenty-five percent is a countdown
How a power station's MQTT readings should reach the machines that must shut down: a Go observer on the router that speaks NUT, as designed.