Cumulative project companion
Use this to organize existing course project work or independent study. It does not replace Canvas instructions, model-counting rules, assessment formats, deadlines or accommodations. Choose a real task in IT or another approved domain and explain its operational relevance. A classical model need not acquire an LLM merely because later modules teach one.
Milestone 1: propose a decision
Name the stakeholder, observation, available inputs, unknown outcome, source/rights, expected benefit and reasons not to automate. Define a simple baseline and an error-cost question. Submit a short proposal or retain it in your independent-study log. Seek approval through the actual course process.
Milestone 2: establish trustworthy evidence
Provide a data dictionary, provenance, missingness/duplicate policy, development split and labeled visual/table exploration. Explain entity/time boundaries and when each input becomes available. Preserve original source data and document exclusions. Keep final-test exploration out of model-directed decisions.
Milestone 3: compare and demonstrate
Compare substantive methods required by the adopting assignment plus an explicit simple baseline. Do not assume that a dummy counts toward any specified number of models. Record a predeclared selection rule, development experiments, frozen evaluation, error analysis and uncertainty. Demonstrate inference with an input/output contract and rejected invalid case. Compare deployment options, resource needs, privacy, ownership and rollback.
Milestone 4: communicate and defend
Organize the report around the problem, relevant literature, methods, results/discussion and justified recommendations/future work. Follow the actual submission/presentation requirements. Include readable code, commands, versions, data/model identities, numeric evidence and limitations. For teams, keep individual contribution evidence and explain one decision independently. Solo technical work does not establish teamwork attainment.
Annotated standard
Developing: “The model got a high score, so deploy it.” Missing population, baseline, split and decision consequences.
Meets: “The frozen method improved the declared validation criterion over persistence. Final-test errors include larger underpredictions in the late period. I recommend a monitored pilot with a named owner and fallback, conditional on validating data availability.” Attach the actual values, units and code; the wording alone earns no evidence.
Extends: Predeclare a robustness or calibration study, investigate a specific failure slice and explain the resource/error tradeoff. Use development data for choosing the response. Complexity and model count alone do not establish graduate-level reasoning.
The authentic phishing case provides a shorter bounded IT example. For the full sequence, use the separate optional completed capstone exemplar, its annotated report and advanced investigations. It combines grouped tuning, multi-seed stability, frozen evaluation, inference, monitoring rehearsal and a written decision. The 14 modules remain independent; the worked report is a model of reasoning and evidence, not a submission to copy or a recommendation to deploy a historical benchmark.