CU Boulder pairs AI research exploration with mentoring; outcomes remain prospective

Discuss stalled research experiments with research-development leaders, principal investigators and computing support staff. Ask which methods teams cannot yet test, what would make an experiment scientifically useful, and who can review its outputs. Offer a bounded pilot-design engagement covering one cohort and a small set of permitted workflows. The value hypothesis is stronger evidence for subsequent funding decisions. Do not promise proposal success or infer demand for a particular product from the stipend. Include mentoring and evaluation labor in the estimate, and distinguish a research-readiness service from a software purchase.

Begin with a reproducible workflow inventory, approved datasets and a baseline result. Map any copilot or coding agent to campus identity, storage and execution controls; compare local, cloud and hybrid options only where the workload requires them. Use isolated environments for generated code and retain model, prompt and dependency versions. A proof of value should test scientific correctness and researcher review effort together. No platform architecture is established by the program page, so obtain local requirements before recommending hardware or endpoints. Require reviewers to inspect failed as well as successful experiments.

Assign a research-development owner, disciplinary mentor and technical support contact before onboarding. Dependencies include eligible participants, permitted data, protected staff time and a usable evaluation environment. Establish a baseline, midpoint review and final evidence handoff, with accessible instructions and office hours for adoption. Proposed acceptance criteria are a replayable experiment package, documented human approval of conclusions, and a comparison of total effort with the prior workflow. These are proposed controls, not observed program results. Risks include mentor overload, uneven access and mistaking polished proposal prose for scientific progress.