Lesson 7 of 7 · 9 min
Governance, cost and responsible deployment
Documentation, fairness, privacy and the cost controls that keep an ML platform sustainable.
Lineage is the core requirement
For any prediction, you should be able to name the model version, the code commit, the training data version, the feature values used, and who approved the deployment. Regulated industries require it; everyone else needs it the first time a prediction is disputed.
This is why the registry, the tracked runs and the request logging in earlier lessons matter together. Each alone is insufficient — lineage is the chain.
What a model card should state
- Intended use, and explicitly out-of-scope uses.
- Training data: source, period covered, known gaps.
- Performance overall and by relevant slice.
- Known limitations and failure modes.
- Ethical considerations and fairness assessment.
- Owner, review date, and contact for issues.
Fairness is measured, not assumed
Removing a protected attribute does not remove bias — correlated proxies such as postcode carry it. The only way to know is to measure outcomes per group.
The definitions conflict mathematically: demographic parity, equal opportunity and calibration cannot generally hold simultaneously when base rates differ. So the choice of fairness criterion is a documented decision about which harm matters most in that context, not something to be solved technically.
Cost levers
| Area | Lever |
|---|---|
| Training | Spot or preemptible instances with checkpointing |
| Training | Stop unpromising trials early during tuning |
| Inference | Quantisation and distillation for smaller, faster models |
| Inference | Batch requests, and cache repeated inputs |
| Inference | Scale to zero for intermittent endpoints |
| Storage | Lifecycle old artifacts and intermediate datasets |
| Attribution | Tag everything by model and team so cost is visible |
The judgement that closes the loop
A marginally better model that doubles inference cost, triples latency and cannot be explained is usually the wrong choice. The deployable model is the one whose accuracy, cost, latency, interpretability and operational burden are jointly acceptable.
That trade-off is the actual job. Everything else in this track exists to let you make it with evidence rather than intuition.
Worth remembering
- Lineage — data, code, model, prediction — is what makes an incident answerable.
- Fairness must be measured per slice; aggregate metrics conceal it.
- Inference cost usually exceeds training cost over a model's life.
Test it now
Governance and cost trade-off cases appear in the hard bank.