Two technical documents have reached final form, summarizing work done on two different planes of the same system. They are now available as PDFs.
Paper 1 — Notes from building a local AI inference system. A narrative engineering article documenting five hardware iterations for local inference: from the first failed test (180 seconds, 90°C, shutdown on four K80s in a standard tower) to the fifth revision validated over twelve continuous hours of inference at full utilization. Eight GPUs active under real load, thermal range stabilized between 35 and 52°C, zero drift across the full window. Acoustic profile compatible with conversation at one meter. CAD design, custom-shaped aluminum parts, 3D-printed components, cable paths constrained by airflow. The Engineering context sections pair the empirical narration with the physical explanation of why the observed phenomena are consistent with known thermodynamics.
Paper 2 — Deterministic Artifact Verification Pipelines for AI-Generated Software Systems. A formal paper modeling the verification pipeline as a deterministic function over the Cartesian product of artifacts × configurations. Five ordered stages (structure, diff/policy, test, runtime, external exposure), VERIFY_ARTIFACT algorithm with early-exit gate, HARD/SOFT/THRESHOLD semantics, threat model, cryptographic attestation. Concrete implementation operationally validated over more than one hundred iterated releases of the same system, with 2,446 test functions distributed across nine blocks.
Why together. They cover the two ends of the stack the method presupposes: the physical at the bottom, the logical at the top. Same principle in both — output must be validated empirically, not assumed. The evidence (thermal on one side, verification on the other) is part of the process, not a byproduct. Neither was conceived as marketing: they are operational documentation built during real work.
Relationship to existing articles. Parts of this content already live here on the Lab in shorter form. For the hardware side: My first local inference test lasted 180 seconds. and How I test vibration and cooling before trusting expensive AI hardware. For the verification side: Epistemic Software Engineering. The PDFs are the consolidated version, in English, with complete data and extended formal model.
Download.
→ Notes from building a local AI inference system (PDF, ~2.8 MB · 13 pages)
→ Deterministic Artifact Verification Pipelines for AI-Generated Software Systems (PDF, ~50 KB · 17 pages · Bilotta, Draft v1.0, March 2026)
The documents are in English, the original format of the work. Future iterations will be added as they come.