Share a prediction from a reviewed Run

Expose Run B's prediction for an application and an MCP client, keeping training and feature helpers private.

Expected result and prerequisites

REST and MCP return the same predictions object. Invalid nested input is refused. Complete Compare fraud experiments; retain Run B, the unchanged notebook, matching artifacts/model.joblib and scientific wheels. The complete environment procedure defines TARGET, SCIENCE_WHEELHOUSE, BUNDLES and SERVING_PYTHON, including the separate serving runtime. Files/data are the synthetic GPL-3.0-or-later example.

Select the function, model and libraries

Prepare the exact-root grant, then build:

RUN_B=$(python -c 'import json; print(json.load(open("run-b.json"))["run_digest"])')
CAPABILITY=python:fraud_detection:predict
apizr experiment expose "$RUN_B" --root . --operator-policy operator.json \
  --capability "$CAPABILITY" --interface rest --artifact model \
  --dependency scikit-learn --dependency joblib \
  --dependency-wheelhouse "$SCIENCE_WHEELHOUSE" --allow-conditional \
  --output-dir "$BUNDLES/rest"
apizr experiment expose "$RUN_B" --root . --operator-policy operator.json \
  --capability "$CAPABILITY" --interface mcp --artifact model \
  --dependency scikit-learn --dependency joblib \
  --dependency-wheelhouse "$SCIENCE_WHEELHOUSE" --allow-conditional \
  --output-dir "$BUNDLES/mcp"

Conditional opt-in acknowledges opaque external wheel providers; it does not waive required evidence or unsupported contracts. pandas remains a training library; joblib and scikit-learn are explicit selected serving dependencies.

Make and check real calls

Install both generated requirements files and start REST using steps 8–9. Send:

{"transaction":{"amount":845.0,"transaction_velocity":11,"foreign_transaction":true,"account_age_days":45}}

Read classification/probability in predictions. The complete MCP client initializes a real stdio session, asserts exactly fraud_detection.predict and calls the same payload without an AI account. Qualification compares both actual responses and checks private training/helpers. artifacts/trained.txt is not a selected service resource.

The qualified macOS arm64 / Python 3.11.14 execution returned this actual object through both transports:

{"predictions":[[1,0.8127568160788873]]}

The label is 1 for this synthetic transaction, with the example model's reported probability. This is a model result, not a calibrated real-world risk estimate.

Useful failures

Missing amount, a string amount or an extra transaction field produce HTTP 422 and MCP tool errors. Run A with B's current model must fail before publishing a bundle. Restore exact reviewed bytes or record another legitimate Run; do not relabel or disable an identity check.

Limits and cleanup

Resource bounds are 16 MiB per file / 32 MiB combined. Current Run-bound resources use direct service bundles, not governed per-call workers. A working prediction is neither scientific validation nor deployment approval. Stop REST with kill "$REST_PID"; MCP closes with its client context. Remove only the example's generated bundles/runtimes after preserving any evidence needed.