A clean QEMU run ruled out the instruction mismatch
QEMU ran the shipped llama stack with a Sandy Bridge instruction profile. The result did not reproduce the crash. It pointed to the next problem being elsewhere.
- testing
- ci
- qemu
Writing
Production AI, debugging black boxes, offline-first mobile, and the checks that keep software honest. Short, opinionated, occasionally useful.
QEMU ran the shipped llama stack with a Sandy Bridge instruction profile. The result did not reproduce the crash. It pointed to the next problem being elsewhere.
GCash ads put valid peso amounts beside the receipt total. FloatWatch uses the relationship between amount, fee, and total to reject that noise.
FloatWatch runs receipt OCR on-device with ML Kit. Its batch wrapper preserves every input path and isolates a failed image instead of dropping the run.
FloatWatch keeps SQLite behind repository interfaces. That gives a cloud implementation one boundary, but it does not make synchronization a one-file job.
FloatWatch uses a weighted completeness score and required-field checks to send uncertain receipts to review. It does not pretend to predict accuracy.
A llama.cpp wheel failed on a judge CPU without AVX2. Disabling AVX2 was not enough because the non-native build still enabled BMI2 and emitted pdep.
A black-box judge that returns a single accuracy figure per attempt. No logs, no stderr. Five days of probing it, one wrong conclusion I lived inside for days, and the guard that hid the crash it was preventing.
I judged an image pipeline for months by whether the output looked right, until the metric that said PASS started quietly disagreeing with the pixels.