Experiment

Claim boundary

This is a confirmatory artifact-efficacy experiment, not the bounded preflight and not a test of ROS as a whole. It must not begin until EX-HANDOFF-2026-6E42 passes and a final sample-size analysis is locked.

Research question

Does a task-targeted structured handoff improve correct continuation relative to a conventional free-form status note?

Method

Use matched completed repository tasks and the comparator defined by DF-BASELINE-2026-48C0. Randomize recipients to a parallel group for the initial effect estimate; stratify broad expertise and task difficulty. Ask each recipient to produce a continuation plan and execute one safe verification step. Independent reviewers score plans against concealed source records.

Acceptance criteria

Use the outcome model in RP-MEASURES-2026-71B4. Pre-register safe continuation success under an accessible time policy as primary; a critical-harm non-inferiority gate; and separate efficiency, calibration, coordination, accessibility, and lifecycle cost outcomes. Do not create post-hoc composite weights.

Apply JR-SAFEGUARDS-2026-7F14 to both conditions. Do not recruit participants until qualified ethics/human-subject review, data stewardship, fixture safety, and participatory accessibility gates are satisfied.

Preregistration draft

Population and unit

The unit of randomization and analysis is an eligible adult recipient performing one scored continuation task. Define the target developer population, recruitment frame, required repository skills, compensation, and broad expertise/familiarity strata before enrollment. Do not use employer performance ratings or require disability diagnosis.

Treatment and control

Use the treatment and conventional-status-note comparator in DF-BASELINE-2026-48C0, generated without cross-condition contamination. Preserve the exact prompts, templates, training, examples, authoring budget, and actual protocol fidelity for both conditions.

Assignment and concealment

Randomize within predeclared fixture-difficulty and expertise strata using a reproducible seed held by a person or service not enrolling/scoring participants. Conceal allocation until the workspace is assigned. Participants cannot be blinded to visible format; reviewers receive a normalized, pseudonymous response that conceals condition where feasible.

Primary estimand

Intention-to-treat risk difference in safe continuation success under the predeclared accessible time policy: structured bundle minus conventional note. All randomized participants with retained consent remain in the denominator. Noncompletion is failure. Withdrawal requiring deletion is not imputed; report its count and condition without retaining prohibited outcome data.

Harm estimand

Risk difference in critical unsafe action or high-confidence materially incorrect continuation. Any critical event triggers protocol review. A small pilot cannot establish rare-harm non-inferiority merely because no event is observed.

Secondary outcomes

Time among successful recipients, plan correctness dimensions, omissions, unsupported assumptions, verification result, confidence calibration, clarification usefulness, cognitive load/satisfaction, accessibility-critical failure, authoring/review/maintenance/interruption time, and rework.

Primary analysis

Estimate the stratified risk difference with a two-sided 95% confidence interval and randomization-based test preserving assignment strata. Report unadjusted condition counts. Use a prespecified task/author hierarchical model only as sensitivity analysis after qualified statistical review. Secondary outcomes are estimation-oriented; control the family or label them exploratory.

Missingness and exclusions

Define eligibility and fixture failure before assignment. Do not exclude protocol-adherent difficult cases, incorrect answers, long times, or noncompleters. Record post-assignment technical failure, consent withdrawal, accessibility failure, and protocol deviation separately. Run worst/best-case sensitivity analysis for missing retained outcomes where ethically permitted.

Sample-size constraint

The final sample size is unresolved because baseline success and task/author clustering are unknown. Under a simple independent two-group normal approximation, 80% power and two-sided 5% alpha require about 388 recipients per arm to distinguish 50% from 60% success, before clustering, attrition, or multiple safeguards. This scenario is a feasibility warning, not a final power analysis.

Do not relabel a small convenience sample as confirmatory. Use the preflight to estimate rubric and operational parameters, then obtain an independent statistical review and lock the smallest practically meaningful effect, baseline-rate range, clustering assumptions, attrition, and sample size before treatment data.

Stopping and changes

Pause for a critical harm, security/privacy incident, consent failure, accessibility-critical barrier, fixture leakage, allocation failure, or reviewer reliability failure. Do not stop for apparent benefit or futility without a predeclared sequential design. Date, reason, author, and impact of every post-registration amendment; label analyses affected by outcome-aware changes exploratory.

Falsification criteria

No practically meaningful improvement, increased unsafe confidence, or total authoring cost exceeding reconstruction savings.

Threats to validity

Template familiarity, task difficulty, reviewer subjectivity, artificial time pressure, author effects, visible condition, task representativeness, and reviewer inference of condition. A later crossover would add period, learning, and carryover threats.

Results

Not run.