Assistance decisions: benefit, review cost and uncertainty
Use 2–3 hours within P4's existing 9–13-hour AI risk allocation for this comparison. It replaces the generic no-AI/bounded-AI comparison worksheet, while retaining normal/ambiguous/harmful cases, actual policy enforcement, disable path, owner and trigger. Semester 08 remains 510–725 hours; no new gate is added.
Define the decision before timing
Choose a bounded engineering-assistance task such as explaining a patch, proposing tests or drafting an incident summary. State the user, permitted inputs, outputs, consequences of error and forbidden actions. Keep synthetic data in the exercise. Decide which outputs require review and what review must establish. Reuse the Semester 07 proposal-review packet; do not rebuild it or label its fixtures actual model observations.
Define success as a correct, inspected, usable result under the same quality target. Fast drafting alone may move work into review or repair. Count setup, prompt preparation, generation wait, inspection, rejected attempts, repair and recheck separately. Record both active effort and elapsed time. Never silently drop failed tasks from the denominator or treat an unreviewed output as completed work.
A worked decision, not a productivity claim
Fictional manual task times are 20, 25 and 30 minutes, with all three meeting the same correctness target: total 75, mean 25. Assisted drafting takes 5 minutes each, but review/repair takes 10, 20 and 40: total 85, mean about 28.3. Reporting “drafting is five times faster” hides the decision-relevant total. These numbers are a teaching example, not model benchmark results. A fourth failed task would still consume effort and must remain in the record.
A small local comparison can inform your task choice but cannot establish general causation. Prior practice, task difficulty, model variation and evaluator bias may differ between conditions. Preselect matched tasks where possible, rotate order, preserve rejected work and use a reviewer unaware of the claimed speed advantage. Do not present this small exercise as a statistically powered study. If all tasks are already familiar, record learning/recall confounding.
Required comparison and safeguards
Before trying an assistant, predict one likely benefit, one failure and the cheapest test distinguishing acceptable from unacceptable output. If no model is available, use the authored proposals as a tabletop comparison and report real inspection time only; generation time and model performance stay unavailable. No paid API or account is required, and an evidence-supported no-AI decision can pass.
Keep the original P4 executable boundary test: an unsafe proposal must not trigger the business action, a legitimate proposal must still work, and review/disable controls must remain effective. A policy paragraph is not enforcement. For an engineering assistant, distinguish suggestion from merge/deploy permission and verify the actual local enforcement boundary you claim. Do not grant production credentials to make the exercise realistic.
Submit a short decision: supported task scope, observed completion and review effort, rejected alternatives, quality failures, residual uncertainty, owner and reconsideration trigger. For example, “use for drafting test ideas; require independent contract tests and review before adoption; reopen when repair effort exceeds the declared threshold.” Choose a threshold for the task before evaluating results; no universal percentage is prescribed.
P4 calibration and transfer
All original P4 criteria still need level 2. Its comparison subcheck now requires a consistent completion denominator, full effort accounting, clear fixture-versus-model provenance and a justified use/no-use decision. Invented model observations, omitted failed attempts or missing enforcement evidence require repair. A no-AI choice without comparing the stated alternatives is incomplete.
At the existing delayed defense, increase review time or introduce a harmful output and revise the decision without rewriting the original record. A different reviewer may reach a different adoption decision if assumptions are explicit; neither may waive mandatory safety boundaries.
Use NIST AI RMF with 1.0 as the named teaching version. Its official page notes revision activity; record the version consulted rather than claiming timeless framework completeness. Source checked 2026-10-23. This is risk reasoning, not certification or legal compliance.