Three inputs measurably move how well you decide across a working day. How long you have been awake. Where the hour sits relative to your own circadian phase. How long you have been inside the current block without a reset. The count of decisions you have already made does not appear in the measurements at all.
So build the day around the three that measure. Log them for ten working days, cap the blocks, and put the reset where your own curve dips rather than where a calendar template says lunch.
The ego depletion reckoning is settled ground and this piece treats it that way. Even the meta-analysis written in the theory’s defence concludes that whether the effect shows up depends on which depleting task the lab used.10 That is a finding about laboratory procedure. It is not a finding about your Tuesday.
The theory collapsed. The calendar built on it did not. Founders still front-load hard calls to 08:00 on faith, still batch investor conversations into a single morning, and still explain a bad 16:00 conversation with a word that has no mechanism behind it. This piece is the replacement schedule.
Why these three models
This runs the Wire Model: score the features of the decision, route to a small ensemble, then force the ensemble to produce dated actions. The scores that mattered:
- Continuous inputs with readable derivatives (0.9). Hours awake, clock hour and minutes-on-task are quantities you can log with a phone. What can be measured one variable at a time asks for a model that varies one variable at a time.
- Many agents writing the same calendar (0.85). Customers, investors, a co-founder, a bank and a distributor each run their own rules. The schedule you designed is not the schedule you get.
- High outcome variance from causes outside you (0.8). Whether a call goes well depends heavily on the state of the person on the other end. That is noise from your side of the table.
- Strategic conflict (0.25). Low. Nobody is misrepresenting anything here.
- Historical-analog density (0.7). Clinics, air traffic control, call centres and parole sessions have had their within-day decision curves measured at scale.
- Regime-break risk (0.5). Asynchronous work and AI triage are changing who controls the queue, which is the assumption the ensemble leans on hardest.
That routes to comparative statics (which input moves the outcome, and in which direction), agent-based (what shape emerges when many local rules collide in one calendar), and the luck-skill continuum (how much of a single day’s result is a draw). Three outcome types: equilibrium, complex, random. Their errors point in genuinely different directions, which is the only reason to run three.
Two things are folded rather than shipped as cards. The behavioral layer lives inside the luck-skill card, because on this topic the bias in question is the sampling error: reading one bad afternoon as a pattern. A separate behavioral card would restate the same lever. Governance folds into the first GEER move, since the only governance this problem needs is one written rule with one owner.
The framework: an input panel, a queue and a draw
1. The inputs: move one, read the sign
Comparative statics is the plainest tool in the kit. Hold everything constant, move one input, record which way the outcome goes and by how much. Applied to a founder’s day it produces a short and slightly uncomfortable list.
Hours awake. Human sleep and wake regulation is described as the interaction of a sleep-dependent homeostatic Process S and a circadian Process C.1 Process S is pressure that accumulates from the moment you wake. It is the closest thing in the literature to the fuel gauge people imagined willpower to be, driven by time awake rather than by effort spent. Dawson and Reid put moderate fatigue on the same scale as alcohol and found that moderate levels of fatigue produce higher levels of impairment than the proscribed level of alcohol intoxication.5
Circadian phase, which is yours and not the office’s. Process C is a clock with an individual setting. In a study using the Munich ChronoType Questionnaire to split 38 participants into early and late types, then testing at 14:00, 20:00 and 08:00, the group difference on the psychomotor vigilance task came from the 08:00 session, where late types performed significantly worse than early types (p = 0.0058). Significant diurnal variation appeared for the late types and not for the early ones, on both the vigilance task and the Stroop.2 The 08:00 hard-decision block is a policy that works for one chronotype and quietly taxes the other.
And you cannot assume which one you are. Pooling twelve years of American Time Use Survey diaries, 53,689 records, chronotype came out near-normally distributed overall and within every age group.3 The middle is crowded and the tails are real. Yours is an empirical question with a ten-day answer.
Minutes inside the block. Mackworth showed performance falling off after the first half hour on task, and also showed that a salient interruption, a half-hour break, or simply giving people feedback on how they were doing could reduce or eliminate the decrement.4 Seventy-five years of subsequent work has argued about the mechanism and not about the shape.
Now the real-world curve. A retrospective study of 33 primary care practices looked at cancer screening orders against appointment time. Among 19,254 patients eligible for breast cancer screening, order rates ran 63.7% at 08:00, fell to 48.7% by 11:00, rose back to 56.2% at noon, then fell again to 47.8% at 17:00.9 The trend is downward. The shape is a decline that resets when the session breaks and then declines again, which is what a block-length model predicts and what a depleting-tank model does not.
One input behaved surprisingly, and it stops the obvious over-correction. In Chennai, researchers measured sleep with actigraphy among low-income adults and found 5.5 hours a night despite eight hours in bed. A three-week treatment raised night sleep by 27 minutes and produced no detectable effects on cognition, productivity, decision-making or well-being. Short afternoon naps at the workplace moved an overall index by 0.12 standard deviations, with gains in productivity, cognition and psychological well-being, offset by less time worked.6 Marginal night sleep did nothing measurable. A mid-day reset did.
Comparative statics, the input panel
Assumes: the inputs move roughly independently and the response is stable enough to read from a short log.
Fits because: readable derivatives scored 0.9, and every input here is loggable with a phone clock.
Breaks when: the inputs interact hard. Sleep debt shifts your circadian phase, which changes what block length costs you, so the partial derivatives stop being separable.
Counteracts: the belief that the fix is a rule about how many decisions to make.
May reinforce: false precision. A ten-day log on one person is a small sample with an obvious flatterer holding the pen.
2. The queue: your calendar is written by other people
Comparative statics assumes you can set the input. Mostly you cannot, because the day is an emergent object. Each agent around you follows a local rule, and the shape you experience is what those rules produce when they collide.
The observational evidence is blunt. Shadowing 24 information workers, Mark, González and Harris found people spent an average of 11 minutes 4 seconds in a working sphere before switching or being interrupted, and 57% of working spheres were interrupted. The finding that matters most for scheduling is the perverse one: the longer someone stayed in a sphere, the more likely it was to be interrupted, and the longer the interruption lasted.7 Depth attracts traffic.
And the cost does not appear where you would look for it. In a controlled experiment, interrupted tasks were completed in less time with no difference in quality. People compensated by working faster, and paid for it in stress, frustration, time pressure and effort.8 The output looks fine, so the damage lands on the operator and shows up three hours later as a bad call you then blame on the hour.
The local rules that write an African founder’s day are specific and knowable. Customers place WhatsApp orders after their own shops close, which is 20:00 to 22:00. Agents call to reconcile float when their till is down. Distributors compress everything into the last three days of the month. A 10:00 San Francisco investor slot lands at 20:00 in Kampala, on top of the order queue.
None of that is a willpower problem. It is a queueing problem, and queues move when you change the arrival rules.
Agent-based, the emergent calendar
Assumes: other actors follow stable local rules you can observe and partly reprice.
Fits because: many agents writing one calendar scored 0.85, and the interruption data shows the shape is produced, not chosen.
Breaks when: one agent dominates. A single anchor customer or a term-sheet counterparty sets your day unilaterally and no arrival rule survives contact with them.
Counteracts: the assumption that a protected block is protected because you named it.
May reinforce: defensive scheduling. Push the queue too far and the orders go to whoever answers at 21:00.
3. The draw: what a single bad afternoon can tell you
Now the part founders get wrong most often, which is inference from one observation.
Return to the clinic study. Detecting a within-day trend in ordering behaviour, with an adjusted odds ratio of 0.94 per step and a confidence interval of 0.93 to 0.96, required 19,254 patients across 33 practices over two years.9 The effect is real and well measured. It is also small enough that thousands of observations were needed to separate it from noise.
You have one afternoon and one call. The counterparty’s state, the quality of your preparation, the strength of the deal and plain conversational luck all sit in that outcome alongside the hour. Where a single result is dominated by draw, one result licenses no conclusion. Attributing that call to the time of day is a sample of one.
The behavioral layer sits here too. The bias worth naming is that a vivid, recent, self-relevant failure gets a causal story attached within minutes, and the story then rewrites the calendar. Founders reorganise entire weeks off one bad Thursday.
The escape is dull and it works. Timestamp the decision, record the counterparty, hours awake and minutes into the block, then record the outcome when it resolves. After forty entries you have a denominator. Before forty, you have an anecdote wearing a theory.
Luck-skill continuum, the variance lens
Assumes: outcomes carry a large draw component and you are willing to wait for a sample before acting on a pattern.
Fits because: outside-cause variance scored 0.8, and the best real-world measurement of the effect needed roughly nineteen thousand observations.
Breaks when: the effect is genuinely large for you personally. Someone with a strong late chronotype forced into 07:00 board calls does not need forty data points to know.
Counteracts: rebuilding a schedule from one memorable failure.
May reinforce: paralysis. “It might be noise” is available as an excuse for every uncomfortable signal.
GEER: the levers, cheapest first
One. Write the rule and name the owner. One line in the operating doc: irreversible decisions require a logged timestamp and a named second reader. This is the governance layer, folded here because a separate governance model produces the same single lever. Cost: ten minutes.
Two. Timestamp for ten working days. Wake time, decision time, minutes into the block, counterparty, outcome. A note on your phone is enough. It is the only way to learn your own phase.
Three. Cap the block, not the decision count. Fifty minutes, then a reset that leaves the chair. The vigilance evidence says the reset restores the curve, and the clinic data shows that restoration in real professional output at the noon boundary.
Four. Take the afternoon reset seriously. Twenty minutes, after the mid-day meal, before the block containing anything irreversible. It is the only move here with a randomised field estimate behind it.
Five. Reprice the arrivals. Publish a decision window. Tell agents and distributors which two hours carry approvals. Route the 21:00 WhatsApp order queue to a person or a bot with a written threshold, so it stops arriving as a judgement call.
Six. Reassign the 08:00 slot on evidence. After ten days, move the hardest recurring decision to your measured peak. If the log says you peak at 15:00, the morning ritual was costing you the best hour of your week.
RADAR: the dated portfolio
DO NOW (T+0 to T+3). Start the log today. Set the fifty-minute block cap in the calendar app tonight. Publish one decision window to your team and your three largest counterparties by T+3. All three are reversible within an hour and dominate across every scenario in this piece, including the scenario where the time-of-day effect turns out to be tiny for you.
HEDGE (T+7 to T+14). Cheap tail insurance against the case where your curve is steeper than average. Add a standing second reader on any irreversible call taken outside your two best measured hours. Insert a twenty-minute reset before any conversation you cannot have twice. Take the Munich ChronoType Questionnaire and treat the result as a prior, not a verdict.
DEFER AND TRIGGER (T+21 to T+28). Moving the company’s whole meeting rhythm, or hiring to absorb the arrival queue, is expensive and socially irreversible. Pre-commit the observable instead. Trigger: if the log shows that more than a third of your irreversible decisions land outside your measured top two hours, or that interruption arrivals exceed six per protected block, then restructure the rhythm at T+28. If neither fires, the block cap was enough and you have saved a payroll line.
The investor portfolio. DO NOW: ask a founder to describe their approval thresholds and who holds them. A founder who personally approves every discount and every refund is running an arrival queue that will not survive scale, and one question exposes it. HEDGE: stop scheduling first meetings at times that are 20:00 or later for the founder, because you are then measuring their vigilance and recording it as their conviction. DEFER AND TRIGGER: before taking a board seat, ask for the operating doc’s decision rules. If they still do not exist a round later, that is a governance finding with a date attached.
CHAIN
The reference class, chosen by structure. The comparison group is not “busy executives”. It is professionals making repeated same-type judgements inside a scheduled session whose order is set by somebody else. Primary care clinics, parole sessions, air traffic shifts, contact centres, and a founder’s sales-and-approvals day all share that skeleton. What they share is a queue plus a session boundary, which is precisely what the models above act on.
The base rate. In the best-measured member of that class, the within-day decline in a specific decision is real, statistically strong, small per step, and non-monotonic across the session boundary.9 The honest prior for your day is therefore a modest downward drift punctuated by a genuine rebound at the break, not a linear slide into evening incompetence.
Present-state modifiers. Three push you off the base rate. A founder is a small-n operator, so personal variation matters more than a population mean. A founder’s queue is less controlled than a clinician’s, since the clinic at least has a fixed appointment grid. And chronic short sleep is more common in founders than in the studied professionals, which lowers the whole curve rather than tilting it.
Subtracting the counterfactual. Some of the improvement you will see is the log itself. Mackworth’s result includes the point that giving people feedback on their performance reduced the decrement.4 Measurement is an intervention. Credit the scheduling change only with what survives after week three, once the novelty of writing things down has worn off.
Matrix-break flag. The agent-based card assumes a human sits at the front of the queue. That assumption is going. When an approval bot with a written threshold handles the 21:00 order flow, the arrival rate against your attention falls close to zero and the remaining decisions are all high-stakes and novel. The base rate above comes from sessions of repetitive judgements, and it stops applying to a day made entirely of exceptions. Watch for the moment your log stops containing routine entries. When it does, the block-length lever gets more important and the queue lever gets less.
What this ensemble cannot see
Start with the largest gap. Every input on the panel is yours, and roughly half of what determines a call belongs to the person you are calling. Their sleep, their hour, their week. None of the three models sees the other side of the table.
Second, the timescale is wrong for the biggest risk. All of this operates within a day. Sustained short sleep across a quarter, the thing most likely to be genuinely damaging a founder, moves the level of the whole curve rather than its shape, and a within-day analysis is structurally blind to a level shift.
Third, comparative statics wants separable inputs and this system refuses to be separable. Sleep debt shifts phase. Phase shift changes what a block costs.
Fourth, none of it addresses whether the decision should have been on your desk. A perfectly scheduled founder approving refunds at their circadian peak is still the wrong architecture.
One action survives all four. Open your calendar now and find the next irreversible decision on it. Write down the hour it is scheduled for, the hour you woke that day, and who chose that slot. If the answer to the third question is not you, move it before the week ends. That single move is defensible whatever the size of the effect turns out to be, because you were going to make the decision anyway and the only variable you gave away for free was when.
Sources and notes
- Borbély, A. “The two-process model of sleep regulation: Beginnings and outlook.” Journal of Sleep Research 31(4), 2022, e13598. Source of the statement that human sleep regulation is described by the interaction of a sleep-wake-dependent homeostatic Process S and a circadian Process C. Open-access full text at PubMed Central.
- Facer-Childs, E. R., Campos, B. M., Middleton, B., Skene, D. J., and Bagshaw, A. P. “Circadian phenotype impacts the brain’s resting-state functional connectivity, attentional performance, and subjective sleepiness.” Sleep 42, 2019, zsz033. Early (n = 16) and Late (n = 22) chronotypes by Munich ChronoType Questionnaire, tested at 14:00, 20:00 and 08:00. Group effect on the psychomotor vigilance task located at the 08:00 session, late types worse than early types, p = 0.0058; significant diurnal variation on PVT and Stroop for late types and not for early types. Open-access full text at PubMed Central.
- Fischer, D., Lombardi, D. A., Marucci-Wellman, H., and Roenneberg, T. “Chronotypes in the US: Influence of age and sex.” PLOS ONE 12(6), 2017, e0178782. Twelve years of pooled American Time Use Survey diary data, n = 53,689, chronotype computed from mid-point of sleep on free days. Near-normal distribution overall and within each age group; later through adolescence with a peak in lateness at about 19 years, earlier thereafter. Open-access full text at PubMed Central.
- Klein, R. M., and Feltmate, B. B. T. “The vigilance decrement: its first 75 years.” Frontiers in Cognition 4, 2025, 1632885. Cited here for two points reported from Mackworth (1948, 1961): the decrement in performance with visual and auditory targets was robust after the first half hour on task, and a salient interruption, a half-hour break, or the provision of performance feedback could reduce or eliminate it. Open-access full text at PubMed Central.
- Dawson, D., and Reid, K. “Fatigue, alcohol and performance impairment.” Nature 388, 1997, p. 235. DOI 10.1038/40775. The publisher page is paywalled below the abstract; the claim cited here appears in the accessible abstract text, namely that moderate levels of fatigue produce higher levels of impairment than the proscribed level of alcohol intoxication. No specific figure is taken from this source. Publisher page.
- Bessone, P., Rao, G., Schilbach, F., Schofield, H., and Toma, M. “The Economic Consequences of Increasing Sleep Among the Urban Poor.” NBER Working Paper 26746, 2021. Actigraphy among low-income adults in Chennai: 5.5 hours of sleep per night despite eight hours in bed. A three-week treatment increased sleep duration by 27 minutes per night and had no detectable effects on cognition, productivity, decision-making or well-being. Short afternoon naps at the workplace improved an overall index of outcomes by 0.12 standard deviations, with a decrease in work time. Working paper page and full PDF.
- Mark, G., González, V. M., and Harris, J. “No Task Left Behind? Examining the Nature of Fragmented Work.” CHI 2005, Portland, Oregon, 321-330. Observation of 24 information workers. Average time in central and peripheral working spheres was 11 minutes 4 seconds (sd = 18 minutes 9 seconds) before switching or being interrupted; 57% of working spheres were interrupted; the longer someone spent in a working sphere, the more likely it was to be interrupted and the longer the interruption. Author copy at UC Irvine.
- Mark, G., Gudith, D., and Klocke, U. “The Cost of Interrupted Work: More Speed and Stress.” CHI 2008, Florence, Italy. Interrupted tasks were completed in less time with no difference in quality; participants compensated by working faster and reported more stress, higher frustration, time pressure and effort. Author copy at UC Irvine.
- Hsiang, E. Y., Mehta, S. J., Small, D. S., Rareshide, C. A. L., Snider, C. K., Day, S. C., and Patel, M. S. “Association of Primary Care Clinic Appointment Time With Clinician Ordering and Patient Completion of Breast and Colorectal Cancer Screening.” JAMA Network Open 2(5), 2019, e193403. 33 primary care practices in Pennsylvania and New Jersey, September 2014 to August 2016. Among 19,254 patients eligible for breast cancer screening, order rates were 63.7% at 08:00, 48.7% at 11:00, 56.2% at noon and 47.8% at 17:00; adjusted odds ratio for the overall trend 0.94, 95% CI 0.93 to 0.96, P less than .001. Open-access full text at PubMed Central.
- Dang, J. “An updated meta-analysis of the ego depletion effect.” Psychological Research 82, 2018, 645-651. Cited only for the finding that effect size depends on the depleting task used, with an attention video appearing ineffective and an emotion video the most effective. Open-access full text at PubMed Central.