The winner of my 16-model eval fabricated 5 times in the audit. Second place, a tenth of a point back: zero fabrications. That one ships. https://t.co/nTdxsKzuH5
Docker release builds for my GPU server were taking 50 minutes. The fix was moving two lines in the Dockerfile.
The image bakes the ML models in at build time, and the APP_VERSION and BUILD_SHA args sat above the bake layer - so every version bump invalidated the cache and rebuilt the whole 27.8GB image. Moved the two lines below the bake: 3 minutes 20 seconds.
(Also killed a stale belief: everyone remembered the image as ~10GB. Measured it - 27.8. Write numbers down.)