Final writeup

One machine, one family, one check

The spine is a machine you measured on, an operation family you carried from matmul to softmax to attention, and an external check that the work was right. The GPU track is the contrast, not a second project. Nine months read as one argument only if a reader can follow the thread from the first table to the last.

The order

Claim in the first hundred words, with the number. Then the artifact index: repository, post, pull request, each with a link and a version. Then the three measurements, each with shape, dtype, band and the rig line. Then what the upstream review changed in your patch, which is the part no solo repository can show. Then the GPU contrast: the same kernel where the hardware has matrix units and a deep memory hierarchy. Then the next lever.

Link, do not restate

The phase 1 and phase 2 posts already carry the matmul numbers. Link them and say what changed since. A final writeup that re-derives the earlier table spends its length on material the reader can reach in one click, and it reads as a compilation rather than as a conclusion.

Keep the honesty structural

Every number links to results/<commit>. The GPU figures carry their own rig line and are labelled as a different hardware class, so the comparison is between designs rather than between speeds. The pull request status is stated exactly as it stands.

Things that break

Three projects stapled together with an introduction that promises a synthesis the body never delivers. A GPU number placed beside an RVV number with no note about the different machines, which is the one claim a reader can check and disprove. No next lever, which turns the ending into a summary rather than a judgment. And a first sentence that announces what the post will do instead of saying what you found.