A systematic audit and dose–response meta-analysis of work-break length under a direct-evidence rule

Background. Precise work-break prescriptions — most visibly the Pomodoro Technique's 25-minute work block with a 5-minute break, and the widely repeated claim that the most productive workers work 52 minutes and break for 17 — circulate as if they were measurements. Whether any specific work-stretch length, short-break length, or long-break length is supported by experiments that directly measure that setting has not, to our knowledge, been audited under an explicit rule that forbids borrowing evidence across tasks, populations, conditions, and outcomes. Methods. We pre-specified a protocol (internal, timestamped 31 October 2026; not entered in any registry) with logged search strings, searched only free and open-access sources (no paywalls, no logins), and screened in two passes against experimental inclusion criteria. The search logs are deposited as supplementary material (papers/supplementary/SEARCH_LOG.md and papers/supplementary/SWEEP2_LOG.md). Every candidate study was then classified as DIRECT (matching task, population, condition, and outcome for the setting in question), INDIRECT (background only), or EXCLUDED, with access-excluded studies (identified but not legally obtainable open access) declared as a limitation. Dose–response fits used a local pipeline (REML meta-regression; three-level variance-components models where arm-level data exist) that was required to pass a validation gate before any new fit: it reproduced Pelland et al. (2026)'s published weekly-volume hypertrophy curve from the authors' own OSF data to within 0.04 percentage points across 0–48 sets/week. Fitted points, where data permitted any curve at all, come from unimodality-constrained two-piece isotonic fits; the headline presentation is the fitted-point presentation — monotone (PCHIP) interpolation through the fitted points, whose apex is the top fitted point itself — with natural cubic spline interpolation of the same points reported only as a clearly labelled sensitivity presentation. Results — work stretch. No direct evidence supports any number. The direct pool (Biwer et al. 2023; Smits et al. 2025; Göksu et al. 2026) contains no experiment varying work-block length over three or more levels with performance outcomes; the only true length contrast (12 vs 24 min in Biwer et al.) reported no per-arm performance, and system-vs-self-regulation contrasts were approximately null (Smits et al.: g ≈ −0.05 for Pomodoro and −0.17 for Flowtime on task completion; Göksu et al.: g = +0.24, non-significant, for a 12-min system). A prior ~21-minute figure rested on vigilance-curve inference excluded by the direct-evidence rule. Results — short break (≤10 min). Within the closest-matching laboratory attention-task system — Bennett (2015), in which breaks were taken between two 10-minute Attention Network Test periods, a laboratory reaction-time task rather than office or study work, with length effects estimated as marginals pooled across three break activities — the headline result for executive-control reaction-time benefit is the top fitted point at 1 min (fitted points 0 / 56.4 / 44.3 / 18.6 ms at 0 / 1 / 5 / 9 min); the peak is not robustly localized and is rated very low certainty. Fatigue likewise has its top fitted point at 1 min, while the recovery outcome (vigor) rose monotonically to 9 min. The result rests on that single dissertation experiment. Johanson et al. (2019) measures a different system — rests between practice sessions, next-day retention, in gaming — and is reported separately: its retention point estimates (4.90 / 6.24 / 6.24 / 7.46 s at ~0 / 2 / 5 / 10 min, valley-flattened values) rise to the 10-min edge, which neither confirms nor contradicts the on-system result, and its pairwise retention differences did not reach significance. Albulescu et al. (2022) reported a cross-system meta-regression slope of +0.07 per minute (p =.006, no standard error printed). Every tested length from 1 to 10 min had a higher point estimate than the no-break/control condition on its study's primary outcome. Results — long break, awake (>10 min). Insufficient direct evidence. The awake pool is Deshpande et al. (2025; one 20-min break, no control group, non-significant session changes) and Sievertsen et al. (2016; 20–30 min vs none, +0.017 SD [95% CI 0.012–0.022], in schoolchildren) — effectively one dose level, so no curve exists. Nap evidence (Hilditch et al. 2016; Kubo et al. 2007; Rosekind et al. 1994; Konya et al. 2025), obtained under sleep deprivation or night operations, is reported separately and is never a basis for the daytime setting; it is directional only, with figure-only outcomes not digitized. Conclusion. Under a direct-evidence rule, the honest outputs are: work stretch — no direct evidence; short break — an on-system headline at the top fitted point of 1 min, with the peak not robustly localized, resting on a single experiment (very low certainty), with every length in the tested 1–10 min band showing a higher point estimate than the no-break/control condition (for Johanson et al. 2019, pairwise retention differences did not reach significance); long break (awake) — insufficient direct evidence. The specific numbers in popular prescriptions are, on current direct evidence, judgment calls rather than evidence-based settings.