Deep Learning Portfolio Lab: evidence

Static pages generated from retained run manifests. Each page is rebuilt in CI and compared with the committed copy before it is published, so nothing here is typed by hand.

Temporal Steam Neural Recommender

Popularity, matrix factorisation, and a two-tower retriever on one frozen temporal split with a persisted candidate rule, cold-start policy, slices, and CPU retrieval latency. Preregistered hypotheses are scored from the same manifests and reported whether they held or failed.

Screening a candidate? The recruiter FAQ explains the result, the leakage controls, and which figures may be quoted; its numbers are checked against these pages in CI.

Fourteen further contracts in the repository are backlog: reviewed problem statements with verified archives and no model. The lab activates one at a time.