Your assignment
Original fictional queue-priority scorer. Define an evaluated release as artifact plus input/output contract, runtime/dependency versions, preprocessing/configuration, data/evaluation references and owner roles. A model file alone is insufficient evidence. Prerequisites: Python functions, JSON and basic hashes. No real cloud model service is deployed by this lab.
The Maple bundle
The offline project uses JSON-authored linear parameters, not a trained or calibrated model. Schema queue-v1 expects units(integer 0–10) and late_ratio(finite 0–1). Stable score=clip(0.1*units+0.4*late_ratio,0,1); candidate changes the units weight to 0.2. These scores are not probabilities or real queue decisions.
Your attempt
For requests(units 2,ratio 0.5),(3,0.5),(10,1), calculate stable and candidate scores. Does passing the same feature schema prove unchanged behaviour? What would you require before claiming the candidate improves predictive quality?
Worked solution
Stable scores 0.4,0.5,1; candidate 0.6,0.8,1. Schema compatibility does not preserve semantics. Golden checks establish these authored calculations and rejection rules, not predictive accuracy, calibration, fairness or deployment readiness. A real candidate needs evaluation tied to the exact artifact/configuration and intended population.
Identity and integrity contract
The teaching manifest pins allowed JSON filenames, SHA-256 digests and version/schema. Verify the path remains inside the bundle, verify bytes, then validate metadata/numeric parameters. Do not deserialize untrusted executable model formats. A checksum compared with a trusted local manifest checks integrity; it does not authenticate origin if an attacker can replace both. No signing or real model registry is implemented here.
Input boundary exercise
Reject missing/extra features, bool units, units 11, ratio NaN, invalid parameter types and non-finite arithmetic. Do not silently coerce units="2" into an accepted request. Decide what a real service would return and log without storing unnecessary raw personal inputs. The local reference raises an error; no HTTP endpoint is created.
Rollback rehearsal
Resolve stable parameters, then candidate, then stable again for the same request. Restored stable must reproduce 0.4 for the first request. This simulates selecting a pinned artifact, not real traffic rollback. Runtime, feature definitions and dependent schemas must remain compatible, and already executed decisions/data writes may not be reversible by restoring weights.
Review and explain
Score 0–2 for artifact identity, path/hash limits, schema/input boundary, golden arithmetic and realistic rollback. Explain version identity versus a movable alias: record the immutable version resolved when evaluating or serving. No course credit or credential is created.
Authoritative reference
MLflow contributors, Model Registry Workflows. Compare version/alias concepts with our local plain-JSON exercise; no MLflow deployment or vendor endorsement is claimed. Next: Maple 2.
Try the activity before reading its reference. Keep your sources, commands and assumptions with your work. Companion activities are unsaved self-practice and do not award a certificate or change lesson progress.