🤓 Latest Submissions
DRIFT Verification Lab
A sensor driver is the software that turns a manufacturer's datasheet into precise communication with a chip: register addresses, byte order, signed arithmetic, timing and calibration. Engineers have traditionally written driver code by hand and discovered its errors on the bench. AI can now generate much of that code in minutes, shifting the central problem from writing drivers to establishing their correctness. A generated driver can compile, return a plausible value and pass tests that encode the same misreading of the datasheet. Such defects surface late, on a physical board, where they are slowest to trace. DRIFT Verification Lab (Datasheet Review, Implementation, and Fault Testing) moves that verification ahead of the hardware. I defined exact profiles for two real temperature sensors. IBM Bob extracted page-cited requirements from the manufacturers' datasheets, and I examined and approved all 23. Hashes bind each approval to the exact source reviewed. From those contracts, Bob built drivers, deterministic virtual devices, bounded fault scenarios, reports, and generators that emit executable artifacts with hash manifests. The core principle is independence between the code under test and the standard used to judge it. Every reading is checked against literal values derived from the datasheet, never from the driver. Each run records bus operations, bytes, virtual timing, expected and observed behavior, and the requirements behind each cited step. In the public workbench, an engineer can replay any scenario, find the transaction where a result diverged and download the report; every result reproduces from the repository. The sensors were chosen for contrasting demands. The TI TMP117 returns a signed temperature word. The Bosch BME280 requires the driver to read signed, little-endian calibration constants, assemble a 20-bit raw value and apply Bosch's integer compensation. Approval, bus, clock and reporting conventions carried across both; the BME280 needed its own decoder, template, model and runner. That contrast is a harder test than a second sensor with the same decoding path. The checks caught a genuine error: Bob's first TMP117 virtual model signaled a completed conversion but discarded the finished temperature. Bob corrected the model; the expected answer stayed fixed. A seeded byte-order defect shows diagnosis and repair performed entirely by Bob: given 0C 80, the correct driver reports 25.0 °C, the defective one -255.90625 °C. The fixed assertion rejects it; Bob's separate repair passes the same assertions, and the original is retained as a negative control. On the BME280, a truncated read delivers two of three bytes, and the driver correctly rejects it. The final Windows run passed 312 automated tests and 2 subtests. Thirteen recorded cases yield 11 expected passes and 2 intentional negative-control failures. Supplementary software checks covered all 65,536 TMP117 raw encodings with zero arithmetic mismatches and 1,300 scenario replays with zero normalized report differences. DRIFT is a software proof of concept on two sensors. The same approach extends to parts with no existing driver, to a virtual bench rerun on every commit, and to comparison against real hardware captures, letting teams adopt AI-written firmware on evidence rather than appearance.
27 Sep 2026
