4. Deployment and Governance

Operate an analytic decision system with monitoring, accountability and value evidence

4. Deployment and Governance

A notebook ends when the result is produced. A decision system begins when the result can affect people, money or operations repeatedly.

datafeaturesmodelpolicyactionoutcomenew data.\text{data} \rightarrow \text{features} \rightarrow \text{model} \rightarrow \text{policy} \rightarrow \text{action} \rightarrow \text{outcome} \rightarrow \text{new data}.

Every arrow is a possible failure point. Monitoring only model accuracy is therefore insufficient.

Learning outcomes

After this module, you should be able to:

  1. specify a data product and decision-policy interface;
  2. monitor service, data, model, decision and outcome behaviour separately;
  3. design shadow tests, staged rollout, rollback and incident review;
  4. audit fairness, privacy, security, human review and contestability;
  5. distinguish technical adoption from causally demonstrated business value.

The operating question

LayerQuestionHarborMart evidence
serviceDid the system respond correctly and on time?latency, failed requests, stale features
dataDo inputs still mean what the contract says?schema, missingness, timestamp and range checks
modelDo scores still rank and calibrate?later labelled outcomes by period and group
policyWere scores converted into the intended actions?threshold, capacity rule and override log
outcomeDid the action improve the objective without unacceptable harm?experiment or credible comparison

The owner and response rule must be written beside every alert. An unattended dashboard is not a control.

Chapter route

  1. Data Products and Monitoring — contracts, drift, feedback and rollback.
  2. Governance, Fairness and Privacy — accountability across the lifecycle.
  3. Adoption and Business Value — test whether the deployed policy changes behaviour and value.

Prepare, work, follow up

  • Prepare (45 min): draw the complete path from an order event to a capacity action. Mark every timestamp and owner.
  • Workshop (120 min): run a simulated incident review: stale capacity data makes the late-delivery score understate risk during a storm.
  • Follow up (60 min): submit a one-page control plan containing five monitors, alert thresholds, owners, rollback criteria and one value test.

Postgraduate extension: treat the policy as an adaptive intervention. Explain how selective labels, feedback loops and changing treatment effects invalidate a static evaluation.

Next: Data Products and Monitoring

Copyright © 2026