../

Cognitive biases

The systematic ways human judgment departs from what the evidence supports: where the heuristics-and-biases idea came from, a catalog of the well-supported biases grouped by kind, which famous effects failed to replicate, how persuasion exploits these tendencies, and the techniques that actually reduce error. The probability side (base rates, regression, selection effects) is in probability & risk, decision processes are in decision-making, and the general toolkit is in general thinking.

System 1 and System 2

Daniel Kahneman's Thinking, Fast and Slow (2011) popularised a two-system description of thinking, using the labels Keith Stanovich and Richard West had proposed in 2000.

System 1System 2
speedfast, automaticslow, deliberate
efforteffortless, always oneffortful, limited capacity, easily depleted by distraction
controlinvoluntaryvoluntary
examplesreading a face, 2 + 2, driving an empty road, sensing hostility in a voice17 × 24, checking an argument, filling in a tax form, parking in a tight space
strengthspattern recognition, expertise, speedrules, logic, statistics, overriding a first impression
failure modesubstitutes an easier question; jumps to a coherent storylazy: endorses System 1's answer without checking

Attribute substitution is the core mechanism: faced with a hard question ("how likely is this startup to succeed?"), System 1 answers an easier one ("how impressive was the pitch?") without noticing the swap. WYSIATI ("what you see is all there is", Kahneman's acronym) is the tendency to build a confident story from whatever information is at hand, without asking what is missing.

When to trust intuition

Kahneman and Gary Klein, from opposing camps, agreed (American Psychologist, 2009) that intuition is trustworthy only when all three conditions hold:

ConditionPresent (trust it more)Absent (distrust it)
a regular environment with valid cueschess, firefighting, anaesthesiology, debugging a familiar systemstock picking, long-range political forecasting, early-stage startup picking
prolonged practicethousands of repetitionsa handful of cases
rapid, clear feedbackyou learn quickly whether you were rightoutcomes arrive years later, noisily

Confidence is not a signal of accuracy: people feel just as sure in low-validity environments.

The heuristics and biases program

Amos Tversky and Daniel Kahneman's 1974 Science paper, "Judgment under Uncertainty: Heuristics and Biases", argued that people judge probability using a few mental shortcuts that are usually useful but produce systematic, predictable errors.

HeuristicJudges probability by…Typical error
representativenesshow much A resembles the typical Bbase-rate neglect, conjunction fallacy, insensitivity to sample size, misconceptions of chance
availabilityhow easily examples come to mindoverweighting vivid, recent, reported events
anchoring and adjustmentstarting from a number and adjustingadjustments are too small, even from obviously irrelevant anchors

Prospect theory (Kahneman and Tversky, 1979) extended the program to choices: people evaluate outcomes as gains and losses relative to a reference point, weigh losses more than gains, and overweight small probabilities. Kahneman received the 2002 Nobel prize in economics for this work (Tversky had died in 1996).

Criticisms worth knowing

CritiqueArgumentWhat to take from it
ecological rationality (Gerd Gigerenzer)simple heuristics often perform well in real environments; many "biases" shrink when problems are posed as natural frequenciesheuristics are tools, not defects; the question is fit to the environment
task artifactssome classic results depend on wording (e.g. what "probable" means in the Linda problem)test the bias in your own setting before assuming it
replicationsome effects in the wider literature have not held up; see replication statusprefer effects replicated in large pre-registered studies
bias blind spotknowing about biases does little to stop you having themrely on process (checklists, outside view), not willpower

Judgment and decision biases

Each entry: what it is, what it looks like, and the countermeasure most likely to work. Effects here are ones with substantial replicated support; contested ones are flagged.

Estimating and predicting

BiasDefinitionExampleDebias with
anchoringestimates are pulled toward an initial number, even an arbitrary oneTversky and Kahneman (1974): after a rigged wheel stopped at 10 or 65, median estimates of the share of African countries in the UN were 25% and 45%. Replicated strongly in Many Labs 1.make your own estimate before seeing anyone else's; generate several anchors; in negotiation, set the first number yourself
availabilityjudging frequency by how easily examples come to mindfearing plane crashes more than car journeys; overrating the risk of whatever was in the newslook up base rates; ask "what am I not hearing about?"
representativenessjudging probability by resemblance to a stereotype"he looks like a founder"start from the base rate, then adjust (Bayes)
conjunction fallacyjudging a specific combination more likely than one of its partsthe Linda problem (Tversky and Kahneman, 1983): 85% ranked "bank teller and active feminist" above "bank teller". Much weaker when posed as frequencies ("out of 100 people like Linda…").multiply the probabilities of each step; count, don't narrate
planning fallacyunderestimating time, cost and risk of your own plans while knowing that similar plans overranBuehler, Griffin and Ross (1994): students predicted their thesis would take 34 days on average; it took 56, and only about 30% finished by their own estimatereference class forecasting; compare with your past estimates
overconfidencebeing more sure than your accuracy justifies; "90%" intervals that contain the answer far less oftenmost founders rate their own odds far above the base ratecalibration training; ranges; keep score
optimism biasbelieving good things are more likely, bad things less likely, for you than for others (Weinstein, 1980)"most startups fail, but not this one"pre-mortem; outside view
hindsight biasafter the outcome, believing you knew it all along (Fischhoff, 1975)"the crash was obvious"write predictions down beforehand, with dates
outcome biasjudging a decision by its result rather than its quality at the time (Baron and Hershey, 1988)praising a reckless deploy that happened to workreview the decision with the information available then
neglect of probabilityresponding to whether a vivid risk exists, not how likely it is (Sunstein, 2002)paying heavily to remove a tiny, frightening risk while ignoring a larger dull onestate the probability and the size of the effect as numbers
scope insensitivityvaluing outcomes without regard to their sizeDesvousges et al. (1993): groups asked what they'd pay to save 2,000, 20,000 or 200,000 birds answered about $80, $78 and $88put quantities side by side; compute per-unit value

Choosing and valuing

BiasDefinitionExampleDebias with
framingthe same choice described differently gets different answersTversky and Kahneman (1981), "Asian disease": 72% chose the sure option when framed as lives saved; 78% chose the gamble when the same numbers were framed as deathsrestate the choice in the opposite frame; use absolute numbers
loss aversionlosses hurt more than equal gains please; typical estimates of the ratio are about 2 (Tversky and Kahneman, 1992: 2.25)refusing a 50/50 bet to win $110 or lose $100frame decisions as portfolios of many bets; ask about final states, not changes. Magnitude and generality are debated (Gal and Rucker, 2018)
endowment effectvaluing something more once you own itKahneman, Knetsch and Thaler (1990): mug owners asked roughly twice what buyers would payask "would I buy this today at this price?" (debated: some studies find it shrinks with procedure and experience)
status quo biaspreferring the current state (Samuelson and Zeckhauser, 1988)organ-donor consent is far higher in opt-out countries than opt-in ones (Johnson and Goldstein, 2003)design good defaults; periodically make "keep" an active choice
sunk cost fallacycontinuing because of what you've already spent (Arkes and Blumer, 1985)finishing a doomed project "because we've put a year into it"ask only about future costs and benefits; kill criteria set in advance
IKEA effectovervaluing what you built yourself (Norton, Mochon and Ariely, 2012)not-invented-here: preferring your in-house tool to a better libraryhave someone who didn't build it evaluate it
zero-risk biaspreferring to eliminate a small risk entirely over a larger reduction in a bigger one (Baron, Gowda and Kunreuther, 1993)fixing every low-severity bug in one module while a high-severity class of bugs stays opencompare expected harm removed, not whether a risk hits zero
hyperbolic discounting / present biasdiscounting the near future steeply and the far future gently, so preferences flip as a date approaches (Laibson, 1997)many people take $100 today over $110 tomorrow, but choose $110 in 31 days over $100 in 30 dayscommitment devices; automate the long-term choice (auto-save, scheduled refactor time)

Believing and defending

BiasDefinitionExampleDebias with
confirmation biasseeking, noticing and remembering evidence that supports what you already thinkWason's 2-4-6 task (1960): people test only triples that fit their rule, never ones that would break itlook for disconfirming evidence first; ask "what would change my mind?"
motivated reasoningreasoning toward the conclusion you want (Kunda, 1990)holding a rival's code to a higher bar than your owndecide the criteria before seeing whose result it is
belief perseverancebeliefs survive after their evidence is discredited (Ross, Lepper and Hubbard, 1975)still trusting a metric after learning it was mis-instrumentedexplain how the opposite could be true; "consider the opposite"
bias blind spotseeing biases in others more than in yourself"their estimate is anchored; mine is just right"process-level safeguards rather than self-assessment

Social biases

BiasDefinitionExampleDebias with
social proofdoing what others are doing, especially when unsureAsch's line studies (1950s): about a third of answers conformed to a unanimous wrong majorityask what you would think if you hadn't seen what others did; collect opinions independently
authority biasoverweighting the view of a perceived authorityMilgram (1963): 26 of 40 participants went to the maximum "450-volt" shock when instructed (later archival work shows much variation across conditions)judge the argument, not the title; invite dissent explicitly
halo effectone good trait colors judgments of unrelated traits (Thorndike, 1920)assuming a charismatic founder must also be a good operatorscore attributes separately and independently
in-group biasfavoring your own group, even one formed arbitrarily (Tajfel et al., 1971, minimal groups)backend team discounting a frontend team's estimatesmixed teams; blind review; rotate people
fundamental attribution errorexplaining others' behavior by character, not circumstances (Ross, 1977; Jones and Harris, 1967)"that engineer is careless" when the deploy process made mistakes easyask what situation would make a reasonable person do this; blameless post-mortems
false consensusoverestimating how many people share your views (Ross, Greene and House, 1977)"everyone wants dark mode"survey or test; you are not the user
curse of knowledgebeing unable to imagine not knowing what you knowNewton (1990): tappers predicted listeners would name about half of the tapped songs; listeners named 3 of 120test explanations on newcomers; write for a reader who lacks context
spotlight effectoverestimating how much others notice you (Gilovich, Medvec and Savitsky, 2000)students wearing an embarrassing T-shirt far overestimated how many people noticed itremember everyone else is busy with their own spotlight
bystander effectfeeling less responsible to act when others are present (Darley and Latané, 1968)an incident channel full of people watching and no one owning itname an owner; "who is doing what by when?" Note: real-world CCTV data (Philpot et al., 2020) found someone intervened in about 90% of public conflicts, and the Kitty Genovese story that inspired the research was exaggerated
groupthinkcohesive groups suppress dissent to preserve harmony (Janis, 1972)a leadership team where no one challenges the founder's planred team; leader speaks last; anonymous input. Janis's full model has had mixed empirical support
Dunning–Kruger effect (contested)Kruger and Dunning (1999): the least skilled overestimate their performance mostbeginners rating their code highlysee below; calibrate everyone with objective feedback

The Dunning–Kruger critique

The famous chart (low scorers hugely overestimate; high scorers slightly underestimate) is largely produced by two general effects: regression to the mean (self-estimates correlate imperfectly with scores, so extreme scorers' estimates sit closer to the middle) and the better-than-average effect (most people rate themselves above average). Nuhfer and colleagues (2016) reproduced the pattern with random data, and Gignac and Zajenkowski (Intelligence, 2020) found that, with valid methods, the relation between ability and self-assessment is close to linear, calling the effect "mostly a statistical artifact". People are imperfectly calibrated at every skill level; the popular version ("the incompetent are uniquely confident") is not well supported.

Memory biases

BiasDefinitionExampleDebias with
peak–end ruleexperiences are remembered by their most intense moment and their end, not their totalKahneman et al. (1993): most participants preferred to repeat a longer cold-water trial that ended slightly warmer; Redelmeier and Kahneman (1996) found the same with colonoscopiesdesign endings deliberately (onboarding, offboarding, incident close-out); measure during, not only after
duration neglectthe length of an experience barely affects its remembered valuethe same studiestrack totals, not impressions
rosy retrospectionremembering events as better than they felt at the time (Mitchell et al., 1997)"the early days were fun"; nostalgia for the old stackkeep contemporaneous notes and metrics
misinformation effectlater information alters memory of an event (Loftus and Palmer, 1974)asking how fast cars were going when they "smashed" rather than "hit" raised speed estimates and false reports of broken glasstake statements early; ask open, non-leading questions in incident reviews and interviews
serial position effectthe first and last items in a list are remembered best (primacy and recency; Murdock, 1962)the middle candidate in a day of interviews blursscore each candidate immediately; put key points first and last
hindsight and consistencymemory of past beliefs drifts toward current beliefs"I always thought this would work"decision journals

Munger's psychology of human misjudgment

Charlie Munger gave the talk "The Psychology of Human Misjudgment" at Harvard in June 1995, listing 24 "standard causes of human misjudgment". An expanded version with 25 "tendencies" appears in Poor Charlie's Almanack (2005). His emphasis differs from the academic literature: incentives first, and combinations of tendencies (his lollapalooza effect) that produce extreme outcomes.

Cause (1995 list)MeaningEngineering / startup form
underrecognition of incentive powerpeople underestimate how strongly incentives drive behaviorthe metric you reward is the one you get
incentive-caused biaspeople (including professionals) sincerely come to believe what benefits themthe vendor's consultant recommends the vendor; the team that owns a service argues it can't be deprecated
consistency and commitmentonce committed, people defend the positionpublic roadmap promises that outlive their rationale
reciprocationreturning favors, even unrequested ones"free" enterprise pilots that create obligation
social prooffollowing the crowdadopting a framework because everyone else did
contrastjudging by comparison to what came just beforea $50k tool looks cheap after a $500k quote
authoritydeference to ranknobody questions the CTO's architecture
deprival super-reactionoverreacting to loss or near-lossfighting to keep a feature a handful of users use
envy / jealousycomparing to peerschasing competitors' features
liking / disliking distortionfavoring people and ideas you like; dismissing those you don'ta friend's startup gets an easier diligence
say-something syndrometalking to seem usefulmeetings that end with more opinions than decisions

His best-known illustration: Federal Express could not get its night shift to move packages on time until it stopped paying by the hour and paid by the shift, at which point the work got done faster. Incentive-caused bias is not dishonesty: people genuinely believe the conclusion that pays them. Discount advice by the adviser's incentives, including your own.

Cialdini's principles of influence

Robert Cialdini's Influence: The Psychology of Persuasion (1984) set out six principles from experiments and years spent inside sales, fundraising and advertising organizations. In Pre-Suasion (2016) he added a seventh, unity. Each is a legitimate way to help people decide, and each is a lever for manipulation.

PrincipleHow it worksLegitimate useWatch for (defense)
reciprocitywe feel obliged to return what we've receivedgive genuinely useful content, help or trials firstsmall unsolicited "gifts" before a big ask; judge the offer on its merits
commitment and consistencyafter a small commitment we act to stay consistent with itask users to state a goal during onboardingfoot-in-the-door escalation; ask "knowing what I know now, would I say yes?"
social proofwe follow what similar others do, especially under uncertaintyreal testimonials, usage numbers, logos (with permission)fake reviews, inflated counts; ask whether those people are like you
authoritywe defer to credible expertsshow real credentials and expertiseborrowed or irrelevant authority (uniforms, titles); check the expertise is in this domain
likingwe say yes to people we like: similarity, compliments, cooperationbuild rapport honestly; find real common groundcharm that substitutes for substance; separate the person from the deal
scarcitythings seem more valuable when rare or disappearingreal deadlines and limited capacityfake countdowns and "only 2 left"; ask why you want it, not how soon
unity (2016)shared identity: "one of us" (family, place, tribe, alma mater)genuine community among usersidentity appeals used to shut down scrutiny; ask whether you'd agree if an outsider proposed it

Replication status

Psychology's replication crisis matters here because many popular "biases" came from the same small-sample, flexible-analysis era. The big systematic replication efforts:

ProjectWhat it didResult
Open Science Collaboration (Science, 2015)replicated 100 studies from three leading psychology journals97% of originals were significant; 36% of replications were, with effect sizes about half the originals
Many Labs 1 (Klein et al., 2014)13 effects across 36 samples, about 6,300 participants10 replicated consistently (including anchoring, sunk costs and gain/loss framing); imagined contact weak; flag priming and currency priming did not replicate
Many Labs 2 (Klein et al., 2018)28 effects, about 15,000 participants in 36 countries14 (50%) replicated at a strict threshold; median effect size fell from d=0.60d = 0.60 to 0.150.15

Effects that failed or are contested

These should not be cited as established biases.

EffectOriginal claimWhat happenedStatus
ego depletionself-control draws on a limited resource that exercise of willpower depletes (Baumeister et al., 1998)23-lab registered replication (Hagger et al., 2016, N = 2,141): d=0.04d = 0.04, CI including zero. 36-lab test (Vohs et al., 2021, N = 3,531): confirmatory d=0.06d = 0.06, not significantnot supported in its classic form
power posingbriefly holding expansive poses raises testosterone, lowers cortisol and increases risk-taking (Carney, Cuddy and Yap, 2010)Ranehill et al. (2015), with a larger sample, found no effect on hormones or risk-taking; lead author Dana Carney said in 2016 she did not believe the effects were realhormonal and behavioral claims failed; small effects on self-reported feelings still debated
behavioral (social) primingsubtle cues change behavior: e.g. elderly-related words make people walk slower (Bargh, Chen and Burrows, 1996)Doyen et al. (2012) failed to replicate the walking effect and suggested experimenter expectations; Many Labs 1 found no flag or money priming effects. Kahneman, who featured priming in Thinking, Fast and Slow, later acknowledged relying too much on underpowered studiesmuch of it not supported. Semantic priming (a word speeds recognition of related words) is a different, robust effect
nudges (as a class)small changes in choice architecture reliably change behaviorMaier et al. (PNAS, 2022): after adjusting for publication bias, no evidence remained for nudging overall. DellaVigna and Linos (Econometrica, 2022): academic papers averaged 8.7 percentage points; nudge-unit trials at scale averaged 1.4. The "sign at the top" honesty study (Shu et al., PNAS, 2012) was retracted in 2021 over fabricated dataeffects usually far smaller than published; defaults remain one of the better-supported tools (Jachimowicz et al., 2019 meta-analysis)
stereotype threatreminding people of a negative stereotype about their group lowers their test performance (Steele and Aronson, 1995)Flore and Wicherts (2015) found signs of publication bias; a large registered study in Dutch high schools (Flore, Mulder and Wicherts, 2018, N = 2,064) found no effect; other meta-analyses do find effectsmixed; probably smaller and more context-dependent than claimed
facial feedbackholding a pen in the teeth (forcing a smile) makes cartoons funnier (Strack et al., 1988)17-lab registered replication (Wagenmakers et al., 2016) found no effect; the Many Smiles Collaboration (2022, N = 3,878) found that mimicking or deliberately posing a smile did raise reported happiness, while evidence for the unobtrusive pen method was inconclusiveposed expressions have small effects; the classic pen result is not established
Dunning–Krugersee abovelargely a statistical artifactcontested
loss aversionlosses loom about twice as large as gainsfound in many choice tasks; critics (Gal and Rucker, 2018) argue it is weaker and less general than claimedreal, magnitude debated

Rules of thumb for judging a claimed bias: prefer effects that replicated in large, pre-registered, multi-lab studies; distrust small samples with big effects, especially "one weird trick" behavior changes; check whether the effect survives in the field, not only in a lab; and look for a replication before you build a product or policy on it.

Debiasing techniques

Knowing a bias rarely removes it. What works is changing the process: who judges, when, with what information and in what order.

TechniqueHowCountersSource / evidence
consider the oppositebefore deciding, list reasons your conclusion could be wrongconfirmation bias, overconfidence, anchoringLord, Lepper and Preston (1984) found it more effective than instructions to "be unbiased"
pre-mortem"It is a year from now and this failed. Write down why." Each person writes independently, then shareoptimism, planning fallacy, groupthinkGary Klein, Harvard Business Review (2007); builds on "prospective hindsight" research (Mitchell, Russo and Pennington, 1989)
reference class forecastingpick a class of similar past projects, take its distribution of outcomes, adjust modestly for specificsplanning fallacy, optimism, anchoringKahneman and Lovallo (1993); Lovallo and Kahneman, "Delusions of success" (HBR, 2003); Bent Flyvbjerg (2006) describes its first uses in UK transport planning
outside view before inside viewask "how do projects like this usually go?" before "how will ours go?"same as aboveKahneman's curriculum-writing team estimated 2 years; similar teams had taken 7–10 or never finished (Thinking, Fast and Slow, ch. 23)
checklistsshort, explicit lists of critical steps, used at defined pause pointsmemory lapses, overconfidence under pressureWHO surgical safety checklist: deaths fell from 1.5% to 0.8% and complications from 11% to 7% across eight hospitals (Haynes et al., NEJM, 2009); Atul Gawande, The Checklist Manifesto (2009)
blind evaluationhide names, genders, schools, authorship; score before discussionhalo, in-group, authority, anchoringGoldin and Rouse (2000) credited blind orchestra auditions with part of the rise in women hired (the statistical strength of that finding has since been questioned); blind code review; structured interviews
independent judgments, then aggregateeveryone estimates privately, then compare and averageanchoring, social proof, groupthink, noiseKahneman, Sibony and Sunstein, Noise (2021); wisdom-of-crowds averaging
red teams / devil's advocatea person or team whose job is to attack the plangroupthink, confirmation biassecurity practice; rotate the role so it isn't dismissed as theatre
decision journalsrecord the decision, options, expected outcome with probabilities, and your state of mind; review laterhindsight bias, outcome bias, overconfidencemakes calibration measurable (see forecasting)
cooling-off periodsdelay irreversible or emotional decisions (a night, a week)hot-state decisions, scarcity pressure, present biasthe logic behind statutory cooling-off rights for doorstep and distance sales (e.g. 14 days in the EU)
kill criteriawrite down in advance what result will make you stopsunk cost, escalation of commitmentset with the plan, not when the numbers come in
leader speaks lastthe most senior person gives their view after everyone elseauthority, social proof, groupthinkstandard advice in decision-making practice
incentive auditask who benefits from each recommendation, including youincentive-caused biasMunger put incentives at the top of his list

The step-by-step formats for a pre-mortem, red teams and decision journals are on the decision-making sheet. Whatever the technique, three rules carry most of the benefit: write before you talk (independent judgments), start from the outside view, and decide the criteria before you see the candidates.

Biases most dangerous to founders and engineers

BiasHow it shows upCountermeasure
planning fallacy in estimates"two weeks" for a migration that took two months last timecompare with your own history; estimate in ranges; multiply by your observed overrun ratio; break work down
sunk cost on featureskeeping a feature because of the effort put in, not its usagekill criteria set at launch; review features by usage and maintenance cost
confirmation bias in customer interviewsasking "would you use this?" and hearing yesask about past behavior and money spent, not future intentions (Rob Fitzpatrick, The Mom Test, 2013); see validation
survivorship bias in startup advicecopying the habits of famous winnersfind the base rate and the failures who did the same; weigh advice from people who see many outcomes (Y Combinator)
overconfidence / optimism in fundraising and runwayassuming the next round will close on timeplan runway for a round that takes twice as long; default-alive math
IKEA effect / not-invented-herebuilding in-house what a library does bettercost the build including maintenance; let someone else evaluate
false consensus / curse of knowledge in product design"users will obviously understand this"usability tests with real newcomers; you are not the user
anchoring in pricing and negotiationfirst number mentioned sets the rangeresearch the range first; make the first offer when you know the market
authority bias / groupthink in architecturenobody questions the lead's designwritten design docs with a named reviewer tasked to find flaws
availability in prioritizationthe loudest customer's request jumps the queueweigh requests by revenue, frequency and strategic fit, not volume
outcome bias in post-mortemsblaming the person when the process failed, or praising luckblameless reviews; judge the decision on what was known then
status quo bias in tooling"it works" for a stack that costs hours a weekput the switching and staying costs side by side, annually
incentive-caused bias in advicethe vendor, the agency, the investor with a thesisask what the adviser gains; get a view from someone without the incentive
escalation of commitment in hiringkeeping a bad hire because you made the decisiondecide as if hiring them afresh today

Debiasing checklist

BEFORE A DECISION THAT MATTERS
 
Frame
[ ] Have I stated the decision in both frames (gain/loss)?
[ ] What is the reference class and its base rate?
[ ] Did I make my estimate before seeing anyone else's?
 
Evidence
[ ] What evidence would change my mind? Have I looked for it?
[ ] Am I judging by the story or by the numbers?
[ ] Is my sample only the survivors / the loudest / the recent?
 
People
[ ] Who benefits from each option, including me?
[ ] Did everyone give their view independently before discussion?
[ ] Has someone been asked to argue the other side?
[ ] Am I deferring to rank, charm or consensus?
 
Commitment
[ ] Would I start this today, ignoring what's already spent?
[ ] What are the kill criteria, and when do we check them?
[ ] Is this reversible? If not, have I slept on it?
 
Record
[ ] Written: options, choice, reasons, probability, date.
[ ] Pre-mortem done: top three failure causes addressed.
[ ] Review date set to compare outcome with expectation.

References