Decisioncraft tutorials
Each completed milestone includes reproducible commands, their purpose, alternatives, troubleshooting and evidence limits. Read the chapters in order: later measurements depend on the data and software boundaries established earlier.
| Milestone | Status | What you learn and why it comes here | Tutorial and evidence |
|---|---|---|---|
| 0 — Kickoff | Verified | Define routing labels, validate strict inputs and outputs, lock the CPU environment, and pin upstream source before executing it. | Kickoff tutorial, source inspection |
| M1 — Feasibility | Verified | Establish actual inference, differentiable classification, intended parameter updates and fresh-process reload before investing in a dataset or training study. | Feasibility tutorial, measured report, portable evidence |
| M2 — Data and benchmark | Verified synthetic study | Write consistent labels and group related scenarios before splitting; create an independent holdout that later model choices cannot contaminate. | Data tutorial, label policy, independent review |
| M3 — Baselines and adaptation | Verified selection study | Compare rules, MiniLM, unchanged Kai and adapted Kai under common inputs; use selection data to choose among measured alternatives. | Adaptation tutorial, measured results, independent review |
| M4 — Calibration and protocol | Verified; protocol frozen | Fit probabilities and routing thresholds on separate calibration data, then freeze the analysis before observing final results. | Calibration tutorial, measured report, fitted policies |
| M5 — Evaluation and integration | Verified | Measure unseen-case routing, uncertainty and resource costs, and demonstrate integration without shell execution or inflated claims. | Roadmap: M5; tutorial accompanies the final report and working demonstration. |
| M6 — Reproducible release | Planned | Package exact software/model identities, document limitations, and verify installed inference from freshly downloaded authorized releases. | Roadmap: M6; tutorial accompanies the verified release. |
Planned rows describe objectives, not completed work or runnable future commands. A chapter becomes verified only after its instructions agree with the implementation and saved evidence. The roadmap defines the acceptance gates; plan.md (private development record) identifies the current milestone. A passing milestone can expose a limitation or a negative result without requiring a claim of improved model quality.