# A07 instruments: timetable dividend study

Version 1.0 · prepared 15 October 2026 · prospective, unvalidated local instruments

## Voluntary next-day diary

A research coordinator assigns two baseline and three endline teaching weekdays, balancing Monday–Friday over the cohort and avoiding holidays. The participant reconstructs the previous 04:00–04:00 day. Each episode has a start, end, primary activity and optional secondary responsibility. All primary durations sum to 1,440 minutes. Approximate episodes are accepted and uncertainty marked; unknown episodes are not converted to zero work. Fifteen minutes of paid research time is offered for completion; refusal has no employment consequence.

Fields: pseudonymous teacher_id; school_id; scheduled_day; date; episode_start; duration_minutes; primary_category; job_related; inside_agreed_paid_timetable; task_family; AI_assistance; checking_or_repair; uninterrupted_block; uncertainty_flag. No student names, message text, websites, browser histories, keystrokes or screenshots are requested. AI assistance is self-report, never inferred from telemetry.

Primary categories are teaching/contact, preparation/feedback, administration, professional collaboration, travel, personal care, family care, leisure and other/unknown. Job-related episodes carry one mutually exclusive task family: original production, checking/repair, coordination, training, or other. Nested tags do not add time to the 24-hour total. A paid-timetable map agreed before baseline identifies time outside the scheduled paid timetable; this is an operational research category, not a legal determination of unpaid wages.

Primary daily measure: sum of job-related primary episode minutes outside that map, 0–1,440. Secondary measures: total job-related minutes; checking/repair minutes; uninterrupted preparation blocks of at least 25 minutes; perceived control (single 0–10 exploratory item); total research diary completion burden. Do not interpret a zero as nonresponse. Participants can mark an entire day missing and give an optional categorical reason.

## Allocation charter

Before assignment, list the administrative workflow, existing tasks to cease or consolidate, paid timetable, quality owner, non-AI alternative, approved data boundary and appeal contact. Treatment schools protect one 60-minute block weekly for teacher-selected learning design and prohibit compensating output quotas. Record date, planned block, delivered minutes, cancelled reason and what existing obligation was removed. The charter does not promise an empirically proven gain.

Both conditions receive the same AI access, two-hour training, 30-minute weekly support clinic and independent outcome assessment. Control schools retain ordinary discretion over any efficiency gain. Record naturally occurring protection in the control condition as contamination, without penalizing teachers.

## Blinded quality task and rubric

Use two parallel, pilot-calibrated task packets with synthetic, nonidentifying school scenarios. Each requires a curriculum communication, assessment-feedback plan or administrative explanation under explicit constraints. Counterbalance packets at baseline/endline. Teachers may use their assigned common AI support, and retain responsibility for the submitted response. Two independent human raters, blinded to school, assignment and occasion, score each artifact.

Five dimensions, each 0–4: factual fidelity, pedagogical appropriateness, accessibility and inclusion, actionable reasoning, and verification of claims. Anchors: 0 absent/materially harmful; 1 major unresolved defects; 2 usable only with substantive correction; 3 defensible with minor correction; 4 consistently defensible and independently checkable. The total is 0–20. This rubric is newly proposed and not a validated achievement test. Pilot agreement, task difficulty and ceiling effects must be evaluated before the main allocation. Primary quality score is the mean of two raters' total scores; third-rater adjudication is a separate safety process and does not selectively overwrite the primary mean.

Safety flags independent of total: invented source, disclosure of private data, discriminatory instruction, materially misleading safeguarding guidance. Escalate credible risk promptly to an independent safety lead. Synthetic tasks do not authorize AI decisions about actual safeguarding, disability eligibility, discipline or grades.

## Fidelity interview

At weeks 4 and 12, ask a purposive, consented subsample in both arms: What work disappeared? What checking appeared? Who received additional tasks? Was protected time interrupted? Could you decline a tool? Were output expectations changed? Report themes with negative cases and protected identities. Supervisors receive school-level summaries only, with a minimum disclosure threshold of ten respondents and suppression for identifiable combinations.