Reproducible FPR engineering experiments, agent evaluation, benchmarks, incident analysis, and local-first infrastructure.
A local-first camera-media experiment: 74 videos, 2,942 sampled frames, 132 extracted clips, 58 stills, and a cautionary lesson in AI animal-detection false positives. Private originals stayed untouched.
12 local Ornith model runs with genuine streamed TTFT, prompt/output token counts, 4/5/6 boundary scoring, and a key finding: CPU profiles hit the 160-token output cap before finishing the fix.
A real local AI reliability test: Hermes connected to Telegram but failed on a 32K/64K context mismatch. Here’s the evidence, GPU fix, benchmarks, and remaining acceptance gates.