Decision Record

Context

The handoff experiment needs a credible control before data are observed. A no-documentation control would be a straw comparison, while treating a handoff artifact experiment as a test of all ROS components would exceed its estimand.

Decision

Compare the ROS structured handoff bundle with a conventional free-form status note prompted only for goal, current status, remaining work, blockers/risks, and checks run. Hold task information, tools, safeguards, accessibility, authoring opportunity, recipient instruction, and verification action constant.

Use a randomized parallel-group recipient design for the initial effect estimate. Do not expose a recipient to both formats in treatment data unless a later preregistration explicitly handles learning, period, and carryover effects.

Alternatives

• No handoff artifact: rejected as an artificially weak baseline.
• Existing repository artifacts only: retained as shared context, but not a credible substitute for a successor-directed note.
• Equal-word-count control: rejected because it hides a treatment cost and changes normal author behavior.
• Same participant in both conditions: deferred because structured-format exposure can contaminate later control behavior.
• Historical ROS-versus-pre-ROS project comparison: insufficient for the artifact effect and highly confounded for whole-system attribution.

Evidence

EV-BASELINE-2026-2B6D requires reproducible descriptions of intervention delivery and fidelity. EV-BASELINE-2026-A4F8 identifies learning, period, and carryover risks in software-engineering crossover designs.

Consequences

The experiment may support or weaken an artifact-level and bounded workflow-level claim. It cannot validate ROS portfolio value. A separate prospective ROS attribution study is required.

Reversibility

High before preregistration and data collection. Reconsider if formative real handoffs show the comparator is implausible or if a pilot establishes negligible carryover.

Follow-up validation

Independent review, formative plausibility check, task/author preflight, rubric reliability pilot, and final preregistration.