Something shifts in the mid-afternoon. A decision that felt clear at 10am — the safe option, the one with the defensible rationale — starts to look unnecessarily cautious. By 4pm, people are overriding it. Not because new information arrived, but because the confidence attached to their original judgement has quietly eroded. The question worth sitting with is why certainty behaves like a consumable resource rather than a fixed state, and what that means for anyone designing systems where people make repeated judgement calls under uncertainty.

The Confidence Slider Is Not a Fixed Dial

Most people intuitively treat confidence as a readout: you assess a situation, and your certainty reflects how good the evidence is. But the evidence for that model is thin. What the research more often shows is that confidence is constructed — assembled from mood, recent outcomes, cognitive fatigue, and social context, then retrofitted to whatever decision is in front of you.

This matters because it means confidence is manipulable by factors that have nothing to do with decision quality. A well-rested person at 9am and the same person at 4pm may hold identical information and reach different conclusions, not because one is smarter, but because the internal weighting of "how sure am I?" has drifted.

Behavioural economists have documented this drift in various forms. The classic framing is Kahneman and Tversky's work on loss aversion — losses loom roughly twice as large as equivalent gains — but the more relevant mechanism here is judgement under uncertainty combined with what's sometimes called the "confidence heuristic." People use their own feeling of certainty as a shortcut for actual probability estimation. When that feeling is inflated or deflated by context, the shortcut misfires.

Why the Safe Pick Loses Its Grip

The safe pick has a particular psychological profile. It's usually the option with the lowest variance, the one that's hardest to criticise, and the one that requires the least justification. In the morning, that profile is attractive. It aligns with what Kahneman's System 2 would call deliberate, effortful reasoning — you weigh the options, you note that the safe pick has a better expected outcome, you commit.

By mid-afternoon, three things have typically happened.

Cognitive load has accumulated. Decision fatigue is real, though the popular version of it is oversimplified. What degrades isn't willpower in the folk sense; it's the capacity to hold competing considerations in working memory simultaneously. The safe pick often requires more active justification because it's the less exciting option. When that capacity drops, the path of least resistance shifts toward whatever feels more immediately compelling.

Recent outcomes have coloured the frame. If the last few decisions went well, confidence rises and the safe pick starts to feel like leaving money on the table. If they went badly, confidence falls and the safe pick starts to feel like the only remaining option — but the reason for choosing it has changed from "this is optimal" to "I can't afford another miss." Same choice, different psychology, different downstream behaviour.

The reference point has moved. Kahneman's work on reference points is instructive here. People don't evaluate outcomes in absolute terms; they evaluate them relative to a baseline that shifts with context. A safe pick that looked like a solid gain at 10am can look like a disappointing ceiling by 4pm if the surrounding conversation has drifted toward more ambitious alternatives.

The Variable-Ratio Problem

There's a reinforcement schedule that behavioural psychologists have studied since Skinner: variable-ratio reinforcement, where a behaviour is rewarded after an unpredictable number of repetitions. It produces the highest and most persistent response rates of any schedule. The reason is straightforward — unpredictability keeps the organism engaged because the next attempt might be the one that pays off.

In decision-making contexts outside the laboratory, this schedule shows up wherever outcomes are intermittent and uncertain. The dangerous part isn't the reward itself; it's how variable-ratio schedules interact with confidence. When a safe pick produces a modest, predictable return, it doesn't generate the same dopaminergic signal as an uncertain pick that occasionally pays big. Over a long session, the safe pick's predictability starts to feel like a lack of progress rather than a series of small wins.

This is where the 4pm override becomes legible. It's not that the person has decided the safe pick is wrong. It's that the feeling of the safe pick has changed — from prudent to stagnant — and the confidence slider has drifted just enough to make the alternative look reasonable.

What the Override Actually Costs

The interesting question isn't whether overrides happen. It's whether they're worse decisions or just different ones. The honest answer is that it depends on the decision domain.

In domains with genuine uncertainty and asymmetric payoffs — where the occasional bold call is worth many cautious ones — the 4pm override might be rational. The morning's caution was the anomaly, driven by a fresh cognitive state that overvalued certainty. The afternoon's willingness to take a calculated risk might be closer to the true expected value.

In domains where the safe pick really is optimal — where variance is pure downside — the override is a straightforward error. And the error is systematic: it clusters in the late afternoon, it's driven by internal state rather than external information, and it's invisible to the person making it because the confidence slider feels like a readout even when it isn't.

A Concrete Illustration

Consider a study design that's been replicated in various forms: participants are asked to choose between a guaranteed £50 and a 50% chance of £120. The expected value of the gamble is £60, so the rational choice is the gamble — but most people take the £50. Now run the same choice at different times of day. The proportion taking the gamble rises through the afternoon, and the stated confidence in the gamble also rises, even though the objective probabilities haven't changed at all.

What's happening is that the confidence slider is being recalibrated by something other than the evidence. The person isn't learning that the gamble is better; they're feeling more confident about it. And because confidence feels like information, the decision follows.

Designing for the Drift

If confidence drifts predictably, the practical response isn't to fight it — it's to design around it. A few principles follow from the research.

Separate the decision from the timing. If a choice is genuinely important, make it once, in a stable cognitive state, and then commit to it. Re-deciding at 4pm imports a new set of variables that have nothing to do with the original question.

Make the safe pick's rationale explicit. The safe pick loses its grip partly because its justification is implicit. Writing it down — "I am choosing this because X, Y, and Z" — gives you something to check against when the feeling of certainty starts to shift.

Track overrides, not just outcomes. An override that works out well is still worth examining if it was driven by state rather than information. The outcome doesn't validate the process.

Respect the variable-ratio trap. If you notice that predictable, modest returns are starting to feel unsatisfying relative to uncertain, larger ones, that's a signal worth attending to. It's not necessarily wrong, but it's worth knowing whether the shift is coming from a change in the underlying situation or a change in your own reinforcement history.

The 4pm override isn't a character flaw. It's a predictable feature of how confidence is constructed, and it shows up wherever people make repeated judgement calls under uncertainty — in trading floors, in clinical decision-making, in editorial meetings, in any domain where the safe pick and the bold pick are both live options. The useful move isn't to eliminate the drift; it's to notice it, name it, and decide in advance which decisions deserve to be insulated from it.