Your release meeting
A teammate says,“Golden tests pass and canary health is green, so promote.” Prepare a one-page release/incident review using Maple 1–3: artifact contract, offline evaluation gap, per-version gates, label maturity, rollback prerequisites, owner roles and remaining evidence. No cloud account, production traffic or real model is changed.
Write before reading the reference
What did golden arithmetic establish? What did it not evaluate? Which windows are healthy, faulty or insufficient? Can restoring a model alias undo a data migration or prior automated action? Keep a rehearsal distinct from actual deployment approval.
Reference solution
The authored JSON scorer passes its declared compatibility/golden contract but has no trained-model quality evidence. The offline healthy window is eligible for human review, candidate_fault triggers a rollback advisory, delayed_labels holds and shared_fault triggers investigation. None is a real traffic action or an authorised production release decision.
An evidence table
Include exact artifact/configuration identity, input contract, evaluation provenance, window/denominator details and failure evidence. A green screenshot without those facts is insufficient.
scenario | observed teaching result | review implication
stable restored on r1 | score0.4 | local reproduction, not real rollback
healthy | eligible_for_human_review | not automatic approval
candidate fault | rollback_candidate | advisory only
missing mature labels | hold | quality unknown
shared fault | investigate | model revert may not fix causeRecovery checklist as an exercise
Name incident commander, model owner, service owner and data owner roles; preserve evaluated artifacts and evidence; isolate scope; restore only a known compatible release under an agreed process; verify service AND mature quality; document decisions and unresolved harms. Real credentials, permissions and irreversible actions require the actual operating team. The reference program performs none of them.
Schema and dependency counterexample
Candidate adds a required feature or changes meaning/units. The old artifact cannot necessarily serve current requests, and reverting weights alone does not roll back feature pipelines or data writes. Discuss compatible transitions, preserved contracts and tested restoration. Do not put this teaching JSON or any model fixture into the website database as a migration.
Honest portfolio statement
“Built a local versioned JSON scoring contract, strict validation and an offline monitoring-gate simulator, with golden checks and a recovery memo.” Do not claim a trained model, production MLOps automation, cloud deployment, drift detector, security certification or measured operational improvement.
Rubric and interview defense
Score 0–2 for evidence/identity, compatibility-vs-quality distinction, monitoring/label gates, recovery limitations and truthful communication. Answer: why a checksum needs a trusted reference; why pooled errors can hide canary failure; why labels absent means unknown; and what a model revert cannot undo.
Authoritative reference
MLflow contributors, Model Registry Workflows. Public references remain public. Supporter value is the authored offline workflow and reviewer material, not exclusive access to vendor documentation.
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.