Personal systems experiments about concurrency, Linux I/O, distributed systems, and low-latency software.

Most repositories here are frozen snapshots kept as learning records: their READMEs, measurements, and failure notes are still worth reading, but they are not maintained products. Much of the code was written with heavy AI assistance, and each repository states its own scope and evidence limits.

• orda-book C++ - a single-threaded limit-order matching engine and replay lab. Current work stays on the baseline matcher: price-time priority, partial fills, cancellation, and tests.

• norn C++ - bounded queues, memory ordering, and hazard-pointer reclamation.
• weir C++ - Linux event gateway over epoll and io_uring with a write-ahead log. Its prominent transport numbers describe a retired architecture and are labelled as such.
• raft-kv Rust - Raft with a deterministic simulator, crash recovery, and an LSM state machine.
• loki C++ - deterministic TCP and UDP fault-injection proxy.
• charlatan C and C++ - Linux character device with a userspace harness; kernel runs are VM-gated.
• vanta CUDA - kernel validation with failure replay; GPU-gated.
• ygg Rust and Python - execution-trace analysis; it records a negative result instead of hiding it.
• kiln Rust - experiment runner used by the repositories above.
• fenrir Rust - a workload-search prototype.

Functional checks and performance evidence stay separate. Measured numbers name the machine, commit, and command that produced them, and are kept as dated artifacts rather than current claims. The conventions are in LAB_RULES and EVIDENCE; NEGATIVE_RESULTS keeps the results that argued against the original hypotheses.

_{Maintained by @wheevu.}