Optional advanced investigations
The 14 core modules remain the main route. These investigations add depth after the relevant foundations; they do not add a mandatory Module 15 or change an adopting course's grading, prerequisites or model-count requirements.
Start with the separate completed project. It uses CPU/core dependencies, runs without earlier notebooks, and produces a completed report as well as machine-readable artifacts.
| Investigation | Prerequisite connection | Current form |
|---|---|---|
| Bounded hyperparameter tuning | Modules 5–6 | Executed grouped-CV search with every candidate/fold recorded |
| Training-seed and split stability | Modules 6–9 | Executed multi-seed study and optional separate split experiment |
| Uncertainty, calibration and error costs | Modules 6, 12–13 | Executed bootstrap/reliability/cost evidence with interpretation tasks |
| Complete project assessment | Project companion | Annotated exemplar and evidence crosswalk for independent transfer |
Further depth, without pretending it is already implemented
These are discussion/extension proposals, not verified additional workflows:
- Nested grouped CV: distinguish selection-reused search scores from an outer estimate of the entire tuning procedure. The exemplar currently uses inner training CV plus separate outer validation/final partitions.
- Ablation: predeclare feature/component removals on development data, then explain paired differences and computation costs. Avoid choosing removals from final-test scores.
- Distribution shift: design a prospective temporal or external validation study with permitted records. Source-order batches in the exemplar are not drift-over-time evidence.
- Tracking and recovery at larger scale: translate the local hash ledger into a team registry with access control, scheduling and retention requirements. No external tracking service is needed for the current project.
- LLM and agent robustness: build new permitted adversarial/control sets and compare intent errors with executor permissions. Module 14's finite controls do not establish general safety.
Choose an extension because it answers a research or operational question. A larger model, paid infrastructure or another metric does not automatically make the work more advanced. Every extension should name its question, boundary, expected evidence, compute budget and failure interpretation.
Bridge from the core route
Use the small stability experiment after Module 6, and the evidence activity alongside each module.
Chronological anomaly triage
The optional anomaly extension connects Modules 7 and 12 using existing core dependencies and pinned bike-demand data.