Everything before this was practice under your own referee. The capstone changes the referee: a kernel is done when it is correct under a differential suite, competitive against both the XLA floor and the hand-tuned ceiling, and reviewed by people who maintain kernels for a living. Upstream review is the only test that cannot be gamed, and what it changes about your code is part of the artifact.

Pick a gap that is real: an attention variant no fast public TPU kernel covers, a block-size autotuning pass for an existing Tokamax kernel with measured wins across a shape sweep, or a JAXBench entry on an operator where published baselines trail the expert bound. Sized right, this is four weeks: one derivation, one forward, one backward, one review cycle.

The write-up matters as much as the merge. Derivation, schedule decisions with their measured consequences, the full benchmark record with provenance, and what review changed: that document is the proof that the fourteen weeks produced an engineer, and it closes the track in public, where it started.

• One kernel closing a real gap: an uncovered attention variant, an autotuning pass, or a benchmark entry
• Forward and backward, differential-tested, corner cases enumerated
• Benchmarked against the XLA floor and the hand-tuned ceiling, recorded in bench/
• Sent upstream, review discussion linked from this site

• JAXBench · the benchmark and the gaps its baselines leave open; capstone target candidates live here
• Tokamax · the natural upstream venue; its contribution surface is the autotuning option