ds-package

ds-package — Make It Servable, Prove Parity

Overview

/ds-handoff proves the work reruns. This stage makes it serve: it wraps the pinned model behind a stable inference contract, containerizes it reproducibly, and — the point of the stage — proves the packaged model returns the same predictions it produced offline. Training/serving skew is to deployment what target leakage is to modeling: the silent failure that passes every offline check and still ships a wrong answer. This stage exists to catch it before /ds-deploy.

It ships no code itself — it guides you to generate the contract, wrapper, and Dockerfile into your own project under .last-ds-mile/package/.

When to Use

• After /ds-handoff has produced a pinned environment, a serialized model artifact, and a model card, and the model is ready to become a callable service.
• NOT for further tuning or re-evaluation — if the model isn't finished, go back to /ds-model or /ds-evaluate.
• NOT for text, vision, recommenders, or forecasting stacks — outside plugin scope.

Core Process

1. Gate check. Confirm the /ds-handoff artifacts exist: a pinned environment (lockfile or exact-version requirements.txt/environment.yml), a serialized model with its version and training-data hash/date, and a model card. If any is missing, produce it now (run /ds-handoff's work inline), say plainly that you did, then continue — this pre-check is a discipline gate; only the parity check below (step 5) is the safety gate worth stopping for.
2. Write the inference contract to .last-ds-mile/package/contract.json: the input schema (column names, dtypes, allowed ranges, known categories — derived from the training data) and the output schema (the prediction, plus probability/uncertainty if the model emits it). This is the frozen interface the service promises.
3. Write a thin, framework-agnostic predict wrapper to .last-ds-mile/package/predict.py (or the project's language equivalent): a predict(rows) -> preds that loads the pinned artifact and carries no notebook state or globals. It wraps whatever the model is — sklearn, an AutoGluon predictor, a plain function — behind the one contract.
4. Validate at the boundary. The wrapper rejects or flags rows that violate the contract (out-of-range values, unseen categories) — the serving-time analog of the /ds-data sanitization gate.
5. Parity gate (the signature check). Run the wrapper over the held/eval rows and assert it reproduces the offline predictions — exact for a deterministic model, or within a documented epsilon for a float path where hardware/library differences apply. If parity fails, stop. A mismatch means a feature is computed differently at serve time than at train time; fix the wrapper (or the feature) before continuing.
6. Containerize reproducibly. Generate .last-ds-mile/package/Dockerfile that builds the image from the pinned environment + the serialized artifact + the predict wrapper, plus a smoke test that loads the model and scores one row. Build locally and record the resulting image digest — never commit the image binary.
7. Write .last-ds-mile/stages/11-package.md: the contract summary, the parity result (tolerance used and outcome), the image digest, and the smoke-test result.