eduKate Singapore

VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Next useful routeCanonical hubHELP

How Intelligence Works | Outcome Bias — Good Decisions, Bad Results and Fair Reviews

—

by

eduKate Asia
Original editorial illustration of a diverse research team comparing prototypes, experiment notes and feedback data in a public learning workshop.

Did you know? A careful decision can produce a bad result, and a careless decision can produce a good one. Outcome bias occurs when knowledge of the eventual result distorts how people judge the quality of the decision that produced it. To review a decision fairly, reconstruct what the decision-maker reasonably knew beforehand, compare available alternatives and consequences, and only then examine the result as evidence for improving future choices.

That does not mean ignoring outcomes. Results can reveal missing assumptions, unexpected mechanisms or harms that deserve repair. The crucial distinction is whether new evidence should update a model of the world or whether a single lucky or unlucky result is being smuggled backward into a judgement of the reasoning available at the time. Good intelligence holds both views: evaluate process against its ex-ante information and learn from later evidence without rewriting history.

This guide draws on original psychological research, worked expected-value and probability examples, school and workplace reviews, decision records, causal reasoning, limits and a changed-context exercise. Numerical cases are transparent inventions, not observations from the experiments or real eduKate students.

Contents

The central mistake: confusing decision quality with realised outcome

Imagine a teacher makes a well-supported choice to spend a week rebuilding a class’s missing algebra prerequisites. A few learners still perform poorly on a later test. Does the disappointing score prove the decision was poorly reasoned? Not on its own. The teacher’s choice should be evaluated using information available when it was made, feasible alternatives and expected costs, not solely the realised mark.

Now imagine another teacher guesses a topic to revise without diagnosis and the test happens to reward that choice. A good score is welcome but cannot by itself establish that guessing was a robust method. If the topic mix had been different, the same approach might have failed. The quality of an action and the eventual result are related but different objects.

Outcome bias gives excessive weight to what happened when evaluating the original thought process. It can produce harsh blame for responsible choices that faced real uncertainty, and unwarranted praise for unsafe choices that escaped consequences by luck. These distortions matter in medical review, finance, hiring, operations, education and organisational governance.

A sound evaluation asks two questions in order: was the action reasonable given the evidence, objectives, alternatives and constraints at the time; and what does the outcome teach about the model of the world and the next action? Collapsing those questions loses both fairness and learning. Separating them prevents hindsight from rewriting what was knowable.

Good decisions do not guarantee good results because reality contains variability, measurement noise, human behaviour and incomplete information. A protocol that scores every decision by one outcome can train people to avoid reasonable risks, conceal errors or selectively report lucky successes.

Original research: the outcome changes the judgement even when information does not

Jonathan Baron and John C. Hershey studied the effect in their 1988 paper Outcome Bias in Decision Evaluation, published in the Journal of Personality and Social Psychology. The original work involved five studies in which undergraduate participants evaluated decisions under uncertainty. Scenarios concerned medical choices and monetary gambles, with favourable and unfavourable outcomes.

The researchers explored evaluations of the decision-maker’s reasoning, competence and willingness to entrust future decisions. People often rated the decision process more positively when its outcome was favourable, despite having the same information that was available to the original decision-maker. This is evidence about judgements in those study settings, not proof that every person and all real decisions show exactly the same effect.

A particularly revealing feature of the research concerned comparisons between chosen and unchosen gambles. What happened to an unchosen option could influence how observers evaluated the original choice. The observer was using outcome information that had not been available to the person deciding, even while assessing that person’s prior judgement.

The paper’s registered DOI is 10.1037/0022-3514.54.4.569. These links identify the original peer-reviewed study rather than a simplified online slogan. The correct inference is that outcome knowledge can influence evaluation in studied circumstances; the research does not supply a universal equation predicting every evaluator.

Some outcome knowledge can be genuinely informative for learning. The error is not remembering that a result happened; it is judging the earlier reasoning as if the earlier decision-maker possessed information only revealed later. Preserve that temporal boundary while still asking what the outcome adds to future understanding.

Why a visible result dominates a less visible reasoning process

A result is often simple and memorable: the project succeeded, the patient improved, the exam mark rose or the investment lost money. The quality of the decision process is harder to observe. It requires reconstructing uncertain probabilities, alternative actions, evidence available at the time and constraints that a later observer may not even know.

Once the ending is visible, supportive reasons for that ending may feel unusually compelling. A bad result makes previously reasonable risks seem obvious; a good result makes ignored dangers feel remote. This change in attention can occur without anyone intentionally falsifying a record. A vivid outcome is easier to narrate than an uncertain decision tree.

Organisations reinforce the distortion when promotions, bonuses and postmortems rely solely on final outcomes. If only successes are rewarded, people may learn to take unwarranted gambles, hide near misses or seek assignments with easy success conditions. If every unlucky failure is punished, people may stop reporting uncertainty and avoid responsible experiments.

The remedy is not to pretend that all choices are equally good. Quality standards still require accurate evidence, reasonable estimates, attention to harms, compliance with relevant constraints and a decision proportional to the stakes. Outcomes can test these assumptions after the event, but the tests must distinguish genuine model failures from ordinary variability.

Intelligence in organisations depends on a feedback loop that can recognise both a flawed decision with a lucky ending and a sound decision with a disappointing ending. A system that sees only winners and losers loses information about how to decide better next time.

Outcome bias is related to, but distinct from, hindsight bias

Hindsight bias concerns the tendency, after learning an outcome, to judge that it had been more predictable than it actually was. Outcome bias concerns the evaluation of a decision or decision-maker in light of its result. A person can show both: claiming a bad event was obvious, then concluding the earlier choice was foolish because it did not prevent that event.

An original forecast record can counter the hindsight component. It captures what someone expected before the answer was known. A contemporaneous decision rationale can counter outcome bias by showing the information set and trade-offs. The two records are complementary, not substitutes for careful analysis.

The verified Hindsight Bias guide treats rewritten memories and predictability as its central reader problem. This article addresses a different question: how to evaluate the quality of a choice when the result is already visible.

Someone may accurately remember that an adverse outcome had a ten-percent predicted probability and still condemn the original choice because that ten-percent event occurred. That is the distinct outcome-bias error even when the earlier probability was remembered faithfully. Conversely, somebody can misremember the forecast as fifty percent without changing their assessment of the decision-maker’s competence.

Keep the chronology explicit: forecast or information first, decision next, outcome later, retrospective evaluation last. Changes between these stages are legitimate topics for learning, but they must not be silently merged into one story.

Worked case: expected value before the outcome

Consider an invented decision between two projects under a risk-neutral scoring rule. Project A has an 80% chance of gaining forty points and a 20% chance of losing ten points. Project B produces a certain gain of twenty-four points. Under the stated utility assumptions, A’s expected points are 0.8×40+0.2×(−10)=32−2=30, compared with twenty-four for B.

If these probabilities are well supported and points capture all relevant consequences, choosing A is reasonable under an expected-point objective. The expected advantage is thirty minus twenty-four, or six points per decision. But a single A trial might produce the loss of ten points. A realised −10 does not retroactively change the arithmetic that was available when the choice was made.

Suppose A fails on its first trial. A critic who says ‘B would have earned twenty-four, so only an incompetent person chose A’ is substituting the result of one draw for the expected decision criterion. The comparison was between distributions, not between the eventual outcome of A and the certainty of B after the draw.

The example has limits that matter. If a loss of ten points would violate a safety threshold, a legal requirement or a hard cash-flow constraint, A might be unacceptable despite a higher mean. A risk-averse decision-maker might reasonably prefer B. Decision quality depends on the objective and constraints, not on expected monetary or numeric value alone.

Every quantity is explicit: 80% and 20% are probabilities for one trial, forty and minus ten are outcome points, and thirty is the expectation under those assumptions. The case is fictional and demonstrates reasoning, not a recommendation to gamble with real obligations.

A good strategy can fail often enough to be visible

Suppose an operation has a genuinely calibrated 80% success probability on each independent attempt. That leaves a 20% failure probability per attempt. Over five independent attempts, the probability of all five succeeding is 0.8 to the fifth power, which equals 0.32768, or 32.768%. The chance of at least one failure is therefore 1−0.32768=0.67232, or 67.232%.

An observer might see one failure and conclude that the original forecast was deceptive. Yet failures are expected even under an 80% success model, and over five attempts at least one failure is more likely than none. This does not prove the model is calibrated; it shows why a single disappointing result cannot establish that it is false.

The independence assumption is important. If five attempts share equipment, weather or one poorly measured prerequisite, their results may be correlated. The numerical 67.232% answer then need not hold. A competent analyst declares dependence assumptions rather than using a familiar binomial formula on every sequence of events.

Across a large representative group of similarly forecast events, one can compare predicted probabilities with observed frequencies. A single success or failure is relatively weak evidence about a model’s long-run calibration; a systematic pattern of surprises is more informative. The denominator and how cases were selected must be visible.

Outcome bias becomes especially damaging in organisations that issue a blame finding after every expected failure. Real learning asks whether the forecasts were reasonable and whether the observed rates over an appropriate set match expectations. If not, investigate the model; if so, manage inevitable variation without punishing sound analysis.

A causal picture: choices, hidden conditions and outcomes

A simple causal diagram can help. Evidence available at decision time influences the choice. The choice affects the chance of different outcomes. Unobserved conditions and randomness also affect the result. A later reviewer sees the chosen action and result but may not see all the conditions that made the situation uncertain.

One lucky outcome is not proof of a strong decision process, because the causal path from randomness may have been favourable. One bad result is not proof of a weak decision process, because hidden conditions may have dominated. Reconstructing the decision should therefore include available alternatives and uncertainties rather than only the observed path.

However, outcome evidence can change estimates of underlying conditions. A failure caused by a previously unknown design flaw is useful evidence for later decisions. The fair response is to separate ‘the flaw was not reasonably detectable with then-available methods’ from ‘we now know the flaw exists and should change practice’. Both can be true.

Confounding complicates comparisons across teams. A team given difficult cases may have worse raw results despite stronger decisions than a team assigned easier work. Ranking those teams only by outcome can reward favourable case mix. Use task severity, opportunity, constraints and decision records where available to avoid false comparisons.

Decision review should avoid inventing counterfactual certainty. The unchosen alternative usually has no observed outcome for the same case. Statements such as ‘the other treatment definitely would have worked’ need evidence, not an imagined outcome known only after the fact.

The decision ledger: capture what was knowable before the result

A decision ledger is a brief record made before an outcome becomes known. It identifies the decision, available alternatives, what evidence was considered, forecast probabilities or ranges where appropriate, important constraints, expected benefits and harms, and the reason for choosing one course. It also records what observations would prompt reconsideration.

Make the record proportional to the stakes. A school lesson adjustment may need a concise note about the diagnosed prerequisite, assessment evidence and the next review date. A clinical or safety-critical action requires a more formal authorised record. The purpose is to preserve a fair snapshot of the decision context, not to bury every choice under paperwork.

The ledger should distinguish measured evidence from opinion. For example, ’14 of 20 students missed the same equation step’ is an observation with an explicit denominator. ‘This misconception appears to be the root cause’ is a hypothesis. ‘We will test with a changed-context question on Friday’ is a plan. Keeping these statements separate makes later evaluation more precise.

After the outcome, append rather than overwrite. Record what actually happened, what was surprising, whether assumptions were contradicted and which process changes follow. The earlier forecast should remain intact. A review that silently edits the original entry into a hindsight-perfect explanation destroys the very evidence needed to diagnose bias.

Protect privacy and sensitive operational information. A useful decision record does not require unnecessary personal details, and access should reflect the roles authorised to inspect it. The best ledger is short enough to be used but precise enough to distinguish reasonable uncertainty from ignored evidence.

What aggregate forecasting can reveal that one outcome cannot

Suppose an invented analyst makes one hundred forecasts, each of which assigns an 80% probability to a defined event. Seventy-eight events occur and twenty-two do not. The observed frequency is 78/100=78%, which is two percentage points below the announced 80%. That difference alone may be compatible with ordinary sampling variability and does not prove deception.

The Brier score for a binary forecast can be calculated as (p−y)², with y equal to one when the event occurs and zero otherwise. If p=0.8 and the event occurs, the loss is (0.8−1)²=0.04. If the event does not occur, it is (0.8−0)²=0.64. The average over the hundred invented events is (78×0.04+22×0.64)/100=(3.12+14.08)/100=0.172.

This score judges forecast quality over a dataset rather than labelling an unlucky single outcome as an incompetent decision. A lower average score is better under that metric. But the score still depends on the event definition, representativeness of cases, dependence, and whether the predicted probability corresponds to the right time horizon.

Good calibration does not imply an action was ethically or legally acceptable. Even accurately estimated high probability of success cannot justify violating a non-negotiable safety constraint. Forecasting is one input to decision quality, not a replacement for judgement about permitted actions and human consequences.

Calibration analysis works best with multiple cases and clear denominators. A team that reports only winning predictions or discards difficult cases cannot claim reliable evidence. Preserve failed forecasts and near misses so the decision model can be improved.

Medical and safety reviews must separate blame from improvement

Medical decisions often involve uncertain diagnoses, incomplete information and treatment risks. The eventual patient’s outcome matters profoundly, but it does not by itself reveal whether the clinician’s reasoning was within professional standards given the earlier evidence. A safe review needs clinical context, alternatives, documentation and relevant professional requirements.

A good result should not shield an unsafe process. If a clinician ignores an important contraindication and the patient happens to recover, the decision may still warrant correction. Equally, an adverse result after careful informed consent and reasonable evidence does not automatically establish negligence. These judgments require qualified review, not a chatbot’s unsupported diagnosis.

Incident investigations work better when they look for system improvements rather than treating every bad outcome as proof of one individual’s incompetence. Equipment design, staffing, communication, information handoffs and safety barriers can influence what was reasonable and what could have prevented harm.

This does not eliminate accountability. If a decision-maker ignored a known high-probability hazard or concealed essential evidence, the quality of the decision may have been low before the bad outcome arrived. Proper review can identify that defect using pre-outcome records rather than reverse-engineering blame from the ending.

An AI agent should not assert that a medical decision was safe or negligent based only on a reported outcome. It can help structure evidence and questions for qualified professionals, preserve chronology and surface missing information without replacing authorised clinical assessment.

Education: judge the intervention and the learning evidence separately

A tutor chooses a retrieval-practice routine because students can solve examples immediately after teaching but forget procedures the following week. The decision might be reasonable based on diagnostic evidence. If the next class performs poorly on a particularly difficult test, that result alone does not demonstrate the retrieval routine was mistaken.

Measure whether the intervention was implemented as intended, whether prerequisites were addressed and how students performed on independent comparable questions. Distinguish immediate confidence from delayed retention. A good test mark after a lucky topic match is not sufficient proof that a long-term teaching approach is effective.

Suppose fifteen of twenty fictional learners correctly answer a delayed transfer item after a new routine. The observed share is 15/20=75%. Without a valid baseline or comparison, that figure alone does not establish that the routine caused the result. A strong parent or teacher report explains the assessment conditions rather than claiming guaranteed grade improvements.

A fair review can conclude that the original decision was well reasoned while the realised progress was insufficient, requiring adaptation. The two assessments are not contradictory. The tutor can preserve the earlier rationale and then alter practice dosage or misconception repair based on new evidence.

Education-first judgement prioritises learning that transfers, not the appearance of uninterrupted success. See the verified learning-transfer overview for the distinction between performing one familiar item and using a method in life.

Project reviews: a late delivery is evidence, not an automatic verdict

Projects have stochastic durations and dependencies. A team may choose a reasonable plan with an explicit chance of delay; it can still deliver late. Another team may promise an unrealistic deadline, skip safety checks and happen to finish on time. Ranking the second team as more competent solely by the outcome rewards the wrong behaviour.

An ex-ante review examines the information available when the plan was approved: reference classes, uncertainty ranges, capacity, critical dependencies and contingency. An ex-post review measures actual delivery, unexpected bottlenecks and what would improve planning. Both are needed for accountable operations.

The published Planning Fallacy guide covers the separate mechanism of underestimating work through inside views and weak reference-class adjustment. Outcome bias is about evaluating the plan after its result, not merely forecasting too optimistically beforehand.

If one project exceeds a deadline, verify whether the forecasted risk had actually been considered. If delays repeatedly exceed reasonable forecast distributions, that is stronger evidence of a planning-model defect. Avoid announcing that the difference was obvious in hindsight when no one could reasonably predict its specific cause.

A constructive retrospective keeps both a process judgement and an outcome-impact account: why the original plan was chosen, what happened, who was affected and what changes will prevent avoidable harms. Removing either half produces a misleading story.

Near misses reveal why lucky outcomes should not certify unsafe choices

An unsafe act sometimes produces no visible injury. If observers evaluate the act only by its lucky result, the team may repeat it until a serious event occurs. This is outcome bias in the flattering direction: success conceals a weak process. Organisations need channels for reporting near misses and conditions that could have caused harm.

For example, a worker bypasses a required inspection and a machine runs normally. The absence of an immediate failure does not prove that skipping the inspection was prudent. The quality of the action can be judged from the known inspection policy, hazard and available safe options at the time.

The site’s verified Near-Miss Learning guide explains how reporting bias can hide weak signals. This article adds why the favourable realised outcome should not contaminate evaluation of the original safety decision.

Conversely, responsible risk controls may not prevent every bad event. A review should assess whether safeguards were appropriate and maintained, as well as whether additional evidence shows they need improvement. Treating every negative result as proof of prior incompetence can make employees conceal incidents.

The best learning culture rewards accurate reporting, responsible process and justified improvements while still tracking realised harm. It neither excuses misconduct through luck nor blames uncertainty merely because the ending was bad.

Reward structures can amplify outcome bias across an organisation

If managers reward only high outcomes, workers may optimise for visible wins while hiding risk exposure. If employees are judged only by negative events, they may avoid innovation, escalate every small choice unnecessarily or seek low-risk assignments irrespective of value. The incentive system then alters the decisions it supposedly measures.

Consider a fictional team allowed to choose between a safe return of ten units and a risky action with an expected return of twelve but a substantial chance of loss. A policy that rewards only realised high returns might encourage unjustified gambles, especially if losses are borne by customers. A policy that punishes any loss might force the team into the safe action even when responsible experimentation is permitted.

Separate adherence to a reasonable decision standard from achieved outcomes. That does not mean paying everyone regardless of performance. It means tracking process evidence, long-run results, risk-adjusted performance, safety and whether decisions stayed within authorised constraints. No single metric captures all of these.

Reviewers should be cautious when one team receives easy cases and another receives hard cases. Raw outcomes may reflect case mix more than decision quality. Contextual adjustment and transparent assignment records can help, although such adjustments must not become excuses to hide avoidable underperformance.

An intelligent institution creates an environment in which people can disclose uncertain forecasts without being punished for ordinary variation, but are still accountable for ignored evidence, unjustified risk and failure to learn from repeated surprises.

A practical review procedure that resists outcome bias

Step one is to freeze the ex-ante record: define the original goal, realistic alternatives, information available, forecasts, constraints and the identity of the decision owner. Do this before showing evaluators the outcome where feasible. A blinded process review can reduce contamination by the ending.

Step two is to score process quality against explicit criteria, such as correct use of evidence, recognition of uncertainty, attention to foreseeable harm and respect for the relevant decision authority. Avoid vague labels like ‘smart choice’ when the reasoning can be assessed through concrete questions.

Step three is to unblind the outcome and examine it honestly. Ask which assumptions were supported, which were contradicted, whether unforeseen causes emerged and which actions should change. Distinguish a surprising but plausible draw from a systematically inaccurate model.

Step four is to document both conclusions without allowing one to erase the other. For example: the decision was reasonable given the diagnostic evidence; the result was disappointing; subsequent evaluation identified a new prerequisite gap; the next plan will address it. Such a statement is not indecisive. It is more precise than a single praise or blame label.

Step five is to monitor repeated decisions and improvement. One episode may have limited statistical power. A longer series, relevant comparison group or independently validated process measure can strengthen inference. Record limitations rather than pretending that a perfect verdict is always possible.

When an outcome legitimately changes how we judge the underlying model

Outcomes are not irrelevant. Suppose a safety check was chosen on the assumption that a component fails no more than once in ten thousand operations. New repeated failure evidence may reveal that the assumed rate was wrong. Future decisions should use the revised estimate. Investigators may also discover that earlier evidence had been available but wrongly ignored, changing the judgement of the original process.

This requires careful chronology. Was the crucial warning knowable at decision time? Did the decision-maker have access to it? Was it reasonable to interpret it differently? A newly discovered physical defect can justify engineering changes without showing that the earlier person was negligent.

A single outcome may also be highly informative when it is almost impossible under the prior model. Yet drawing a conclusion still requires the likelihood of the evidence under competing explanations, selection of the case, and relevant measurement validity. ‘Bad result’ is not itself a statistical test.

Distinguish descriptive evaluation from accountability. A decision might be ex-ante reasonable and nonetheless create a moral obligation to repair harm, compensate affected parties or improve controls. Fair process assessment does not erase those responsibilities.

The intelligent stance is neither ‘ignore outcomes’ nor ‘outcomes are everything’. It is ‘preserve what was knowable, learn what the outcome actually reveals, and update policy without pretending the past had access to the future’.

Changed-context exercise: two exam-revision strategies with one unlucky result

Consider a fictional school team choosing a revision strategy. Strategy A uses diagnostic misconception repair and, based on appropriate prior evidence, is expected to help 80 of every 100 comparable learners pass a defined transfer check. Strategy B is familiar drilling and is expected to help 72 of every 100. These are invented expectations, not measured eduKate outcomes.

The team chooses A, but in one small class only six of ten learners pass. The observed result is 6/10=60%, below both expectations. Does that result alone show that choosing A was irrational? No. Ten learners provide a noisy sample, and further questions concern case mix, implementation, measurement and whether the original 80% estimate was appropriate.

Suppose another class using B happens to have nine of ten pass. Comparing 6/10 with 9/10 is not by itself a causal evaluation if the groups differ or assignment was uncontrolled. The second result may be a favourable fluctuation, or it may reveal that the original expectations were incomplete. Better evidence requires suitable comparison and repeated observations.

Now change one assumption: the school possessed reliable evidence before the choice that strategy A was unsuitable for these particular learners, and the team ignored it. The original process would then be weaker even if A happened to produce a good result. The decisive distinction is newly stated pre-outcome information, not simply which class scored higher.

The transfer answer is to retain the original forecast, examine what was actually known, check the class-level denominators and design a fair comparison for future teaching. An honest report can say the outcome was bad and learning requires adaptation without rewriting a reasonable past choice into certain incompetence.

Limits of the concept and the danger of calling every criticism biased

Outcome bias is not a blanket defence for poor decisions. If a person knowingly violates a safety rule, conceals evidence or chooses a dominated alternative without justification, process quality can be low before any result arrives. The final outcome may increase urgency or reveal damage, but it is not necessary to identify the earlier defect.

The original Baron and Hershey experiments concerned particular participants and decision scenarios. Their findings justify taking the bias seriously; they do not prove an identical effect size across cultures, ages, professions and every evaluation task. High-stakes institutional judgement requires domain-specific evidence and qualified oversight.

Expected value is not the only defensible decision rule. Variance, tail risk, fairness, legal constraints, resource limits and individual rights may be decisive. A mathematically higher expected point total can still be an inappropriate action. Avoid using a simplified numerical example to override relevant obligations.

Some outcomes are consequences of processes that should have been recognised beforehand. Reviewers should examine that possibility rather than automatically attributing every failure to chance. A fair method can identify both unwarranted blame and warranted corrective action.

This guide explains a cognitive mechanism for contemporary human and organisational intelligence. It does not diagnose particular individuals or certify real decisions from fragments of information.

Frequently asked questions

What is outcome bias? It is allowing knowledge of a result to influence judgement of a decision’s quality beyond what the decision-maker reasonably knew when choosing.

Can a good decision have a bad result? Yes. Uncertain outcomes are distributions, and a responsible decision may still produce an adverse event.

Can a bad decision have a good result? Yes. Luck can conceal ignored hazards or weak reasoning.

Does outcome bias mean results should be ignored? No. Results matter for welfare, accountability and learning; the point is to separate retrospective evidence from ex-ante reasonableness.

How is it different from hindsight bias? Hindsight bias changes perceived predictability, whereas outcome bias changes evaluation of the decision or decision-maker.

Does expected value settle every question? No. Legal, ethical and risk constraints can make a higher-mean option unsuitable.

What is a decision ledger? A record made before the result, including evidence, alternatives, assumptions, predictions and reasons, preserved for fair review.

Why assess several decisions rather than one? Repeated cases help distinguish systematic model problems from ordinary variability, though dependence and selection still matter.

Is every negative decision review biased? No. Bad reasoning can be demonstrated from pre-outcome evidence, regardless of the eventual result.

What should schools measure? Independent learning, implementation, relevant comparison and diagnostic information, not only a lucky exam mark or subjective confidence.

How should a postmortem read? It should separately assess the original process, the realised consequences and what new evidence changes future practice.

Reader pathways and primary sources

If you are a student, calculate the expected points for projects A and B, then explain why A’s possible loss of ten points does not change the original 30-versus-24 expected-point comparison. Next compute the chance of at least one failure across five independent 80%-success trials.

If you are a teacher or tutor, write down the evidence and expected learning mechanism before changing instruction. Evaluate transfer and retention afterward without turning one test into a universal verdict. Preserve what was known before new marks arrived.

If you are a leader or reviewer, use a process-first retrospective: reconstruct the information set, judge the choice against explicit standards, then assess outcomes, model updates and obligations. Reward accurate risk disclosure and near-miss reporting alongside responsible performance.

Original peer-reviewed research: Baron and Hershey (1988), Outcome Bias in Decision Evaluation, Journal of Personality and Social Psychology 54(4), 569–579; registered identifier DOI 10.1037/0022-3514.54.4.569. The source concerns studied evaluations of medical and monetary decisions, not the fictional examples in this article.

Continue with the verified Hindsight Bias, Planning Fallacy, Near-Miss Learning and How Intelligence Works hub. Each has its own distinct question.

All numerical scenarios are invented for first-principles teaching; no classroom results, medical outcomes or live organisational audit has been claimed.