Mamadou Wane

One-Axis TVC

C++20 · Linux · real time
github.com/mamadou-wane/tvc
v0.2a complete · wire codec, SPSC ring, drain thread, campaign

A hard real-time control and telemetry stack: a 500 Hz control loop on Linux, measured as a latency distribution.

The naive version of this loop drifts seconds off schedule in ten minutes and reports itself healthy the whole time, because measuring each wakeup against the previous one hides the drift. Seven levels, applied one at a time and each measured in isolation, take p99.9 wakeup jitter from that failure to 7.5 microseconds at 500 Hz on a stock Ubuntu generic kernel, with per-core isolation and every C-state disabled.

The most useful result was a wrong number. An earlier campaign measured 88.4 microseconds on the same code, and the only difference was that it left C-states enabled: the idle driver advertises 18 and 350 microsecond exit latencies, so the old headline was measuring C-state exit. One power-management knob carried a 12x difference, and the provenance system could not see it. Governor, energy-performance preference, AC state and package temperature are recorded per run, and every one of those fields reads identically across both campaigns.

Every number traces to a committed campaign summary, and a regression gate diffs new runs against those baselines. Per-cycle telemetry is a single-producer single-consumer ring feeding a binary frame codec, with a ThreadSanitizer stress lane that forces the drop path, and it costs nothing measurable: pooled p99.9 lands within 48 nanoseconds of the quiet configuration over 2.4 million cycles per arm with zero ring drops. A Python ground station, simulation, and fault injection follow.

Wakeup jitter in microseconds, median of three repeats for p99.9.

L6 adds the telemetry ring. Its planned check, L6 within 10 percent of L5 over three repeats, failed at 9.6 against 7.5. Five more repeats of each reversed the sign, L5 at 9.7 and L6 at 8.4, because runs of both configurations intermittently carry a mode that puts a few hundred cycles in the 10 to 35 microsecond band. That mode owns the tail at this floor and which arm it visits is luck, so the cost claim rests on the pooled figure rather than on either three-repeat median.

Skeptic

Python · Docker · LLM evaluation
github.com/mamadou-wane/skeptic
M5 complete · 12-task corpus · Eval A and Eval B published

Your coding agent says the tests pass. Skeptic checks whether that means anything.

Skeptic seeds a known bug into a pinned commit of a real upstream repo, hands the broken tree to an LLM with shell access inside a network-off Docker sandbox, and then checks the diff it produces. The pinned commit is a hidden reference implementation, which turns verification into differential testing with a free oracle. The workspace ships as a git archive export with no .git, so the real fix is unreachable from inside the sandbox.

Twelve corpus tasks across two upstream repos, click and rich. A task is admitted only after seven invariants hold, among them pristine green twice, seed red exactly, gold restores baseline, and hack variants green. Ten checks read the collected observations and score them, with two more behind an opt-in paid profile: an adversarial test battery and an LLM judge. The default profile makes zero API calls.

Across the paid runs measured to date, all three hack fixtures land on SUSPECT and all four clean variants pass.