Godspeed · as of 5 October 2026 · the app's code at cfe227e7 · replaces the document of 26 September 2026

Bias selection batch

Godspeed is an app a person talks a real decision through with. It looks for a small set of reasoning errors in what the person writes and, when it finds one, asks one question. This document lists the errors it looks for, the question each one triggers, where each comes from, what supports it and what does not, and how the set was chosen.

It is written for readers who study judgment and decision-making. We would rather be told where the method is wrong than be reassured. The section on limits is meant literally.

What a "bias" is here

Each entry has two halves, written and tested separately.

  • A definition. One paragraph that a small language model (the detector) reads on every message, to decide whether the error is present in the person's own words. It must quote the words it relies on. Each definition ends with "Not this": the look-alikes that must not count.
  • A question. When the error is found, the model that writes the reply is told to put one question on the person's page, in the person's own terms. It is a question, never advice, it leans to no answer, and the name of the error is never shown.

We call the pair a technique. The word "bias" is used loosely: several entries are not biases in the textbook sense but habits of deciding (no stop point set, no expectation written down) that a single question can address.

How an entry gets in: the five rules

  1. What gets in. A source is worth reading for the techniques it lets Godspeed put to a person as a question. Academic backing is a plus, never a requirement: if a question helps people with their problem it can be included.
  2. Ranking. A technique ranks high when Godspeed can ask it as an easily understood question, when only the person can answer it (the answer is in their motives, their fears, what would change their mind), and when it is clear when to use it. The first of these is the defining test.
  3. The evidence label. No technique is rejected for lacking evidence. Each one always carries one of four labels, visibly, so that tested and untested are never mixed silently: proven (the remedy itself was tested in controlled studies and held up), contested (tested, with results that conflict or failed to replicate), untested (no controlled test known to us), practice only (practitioners report it useful; anecdote or authority).
  4. The duplicate filter. A technique Godspeed already has is not added twice. Two candidates are the same when they put the same question to the person in the same situation. Duplicates are reviewed side by side, wording included, and merged into one entry; the best name, the best definition and the best question may come from three different sources.
  5. The trigger filter. A candidate is accepted only when its definition says what triggers it, in terms the detector can find in the person's prose. The same definition must be publishable as it runs: users are shown the text the model reads.

From a source to the app

  1. Sources are read. Three so far: our research wiki (about 980 articles summarising the judgment and decision-making literature, each tied to its source), a wiki built from Farnam Street's writing (330 articles), and curated LessWrong and CFAR material.
  2. Candidates are written down as a question each: 99 from LessWrong, 96 distinct ideas from Farnam Street.
  3. One person selects. Dominique kept 15 of the 99 and 42 of the 96, by reading them.
  4. Duplicates are grouped against the techniques already in the app, by the question asked, and he picks per group.
  5. A definition is written to four writing rules learned from earlier failures: name something that can be counted or quoted, point at words a person would type, keep "Not this" to the neighbouring shapes, and say where the evidence is (the message, the page, or both). Each comes with made-up messages that should fit and near-misses that should not.
  6. He approves the wording, which freezes it. A label given by a rater is tied to the exact words the rater read; rewording a definition discards its labels.
  7. It is wired into the app, switched off. The detector reads it and records what it finds; nobody is shown anything. we are here for the ten new entries
  8. Two people label. On recorded moments of conversations, each rater says, bias by bias, "fits" or "does not fit", blind to the detector. A label counts when two raters agree.
  9. The detector is measured against those labels, per bias: sensitivity, specificity and the raters' own agreement. Our bar is 40 examples per class and agreement of at least 0.6 (kappa) before a figure is trusted.
  10. Only then is it switched on.

Limits

Nothing in this set has been shown to make a person decide better. Where a label below says "proven", it means the remedy was tested in studies, on their participants and their tasks. No study we know delivers any of these as one question, through software, to one person about their own decision.

  • The labels in this set. Of the 21 active entries: 2 proven, 1 contested, 11 untested, 7 practice only.
  • The ten new definitions have never been measured. They were written on 4 October 2026 and approved on reading. The example messages beside them are invented.
  • Three older definitions were rewritten on 5 October (Plan B, Both frames, The mirror); their earlier labels no longer count.
  • Five questions were reworded on 5 October (Plan B, Steelmanning, The next spend, The expectation, Objectives first) and have not yet been checked on recorded conversations.
  • What we have measured on the detector, and it was not good. On 175 detections by an earlier version, labelled by one person: 32% right, 12% the healthy opposite of the bias, 56% off target. Errors concentrated in the entries that need an interpretation ("a fact read one way") and were rare in those that can be counted off the page (12 right of 13). The definitions here were rewritten with that in mind; whether it helped is what the labelling will tell.
  • What we have measured on the questions. On six techniques, replayed on recorded moments with and without the instruction, three draws each: the reply model alone asked none of the six questions (0 of 90); with the instruction the question reached the page on about 40% of opening messages and 70 to 100% of later ones. That shows the instruction does its job, not that the question helps.
  • Selection is one person's reading. Steps 3, 4 and 6 are a single judge. The grouping of duplicates was done by a language model and reviewed by him.
  • Sources outside the research literature are second-hand. Each technique from Farnam Street or LessWrong was checked against its original text; several of the questions are our own wording of an idea the source states without a question.
  • Small numbers of users. The app is in a closed test; we have counts of what was shown and answered, and no access to conversations.
  • What "proven" means for the two entries that carry it. Three numbers and Steelmanning were each tested in studies of estimation and quiz tasks, and each is known to us narrowly (one secondary source for the first, one paper for the second). Neither has been tested on a person's own decision.
  • Where the research cuts against an entry, the entry says so. Five do: The next spend (the reminder may strengthen escalation for the person who made the original choice), Plan B (an added option can raise deferral), Deciding hot (the question asks for a simulation the literature says people cannot do), Revolving door (the one experiment on reversibility points the other way) and The expectation (betting alone does not cure overconfidence).
  • Replication. Our research wiki holds no replication evidence for several of the classic findings these entries lean on: hindsight, overconfidence, the 1981 framing problems, mental accounting and the planning fallacy.

The set at a glance

11 in the app, 10 wired and switched off, 3 retired. The label is for the technique as Godspeed uses it.

NameStateComes fromEvidence label
Plan Bin the appresearch wiki; question from Farnam Streetpractice only
The successorin the appresearch wikiuntested
The mirrorin the appresearch wikiuntested
Both framesin the appresearch wikipractice only
Base ratein the appresearch wikiuntested
Outside readin the appresearch wikiuntested
Three numbersin the appresearch wikiproven
Steelmanningin the appresearch wiki; name and wording from LessWrongproven
Objectives firstin the appresearch wiki; half the question from Farnam Streetpractice only
The next spendin the appresearch wiki; wording from Farnam Streetcontested
The expectationin the appresearch wiki; the bet from LessWrong and Farnam Streetpractice only
Premortemwired, switched offFarnam Streetpractice only
Exit conditionwired, switched offFarnam Street and LessWrongpractice only
Deciding hotwired, switched offFarnam Street and LessWronguntested
Revolving doorwired, switched offFarnam Streetpractice only
House rulewired, switched offLessWrong and Farnam Streetuntested
Enough to decidewired, switched offFarnam Streetuntested
Social proofwired, switched offFarnam Streetuntested
The bottom linewired, switched offLessWronguntested
True rejectionwired, switched offLessWronguntested
Right or look rightwired, switched offFarnam Streetuntested
The tiebreakerretiredresearch wikiuntested
A typical stretchretiredresearch wikiuntested
The master listretiredresearch wikiuntested

The entries

Each definition is given in full, as the detector reads it. "The user" is the person talking to Godspeed; "the page" is the running summary of their decision that the app keeps beside the conversation.

Plan B in the app

What triggers it (the definition the detector reads)
The user is deciding yes or no on one thing (take the job or not, buy it or not, sign or not), as if those were the only two courses. Nothing in their messages or on the page says what they would do otherwise, or names any other way to get what they are after.
What does not count
Two different things weighed against each other (this job or that one); a third course the user has already named; a question with no option in it.
The question it asks
You are weighing taking the harbour-master post or not: is there a third option?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name a third option".
A message that fits
I’ve been offered the Lyon office. I either say yes by Friday or I stay where I am, and I can’t tell if I want it.One offer, yes or no, nothing else on the table.
A message that does not
I’m choosing between the Lyon office, a fully remote contract, or staying on in Paris for another year.Three options are already named.
Where it comes from
The research wiki (Keeney 1992; Mintzberg, Raisinghani and Théorêt 1976). The wording of the question comes from Farnam Street (Parrish: when you are weighing two options, find a third).
What the bias rests on
Sets of options are too narrow. In Mintzberg et al.'s 25 strategic decision cases, all 14 custom-made solutions followed a single line of development to the end. Keeney says the alternatives people identify are often unnecessarily narrow, and gives reasons, not data. Field description of organisations, and an author's argument.
What the question rests on
Not tested as worded. Keeney's own remedy is to generate alternatives one objective at a time; the experiments he cites are reported without numbers and the wiki records the effect as asserted from experience.
Evidence label
practice only Keeney's experience and cited experiments the wiki cannot quantify.
Caveats the research itself raises
The literature also warns against this move: adding an equally attractive option raised the share of people who deferred the choice from 34% to 46% in one study (n = 124 and 121).
Articles in the research wiki
decision-analysis/creating-alternatives-from-objectives.md; bounded-rationality/decision-development-search-and-design.md; constructive-preference/choice-under-conflict-and-deferral.md
State
In the app. Off by default in production; a person can switch it on in Settings.

The successor in the app

What triggers it (the definition the detector reads)
The question is whether to stay with or change something the user is already inside, and the conversation shows DEFERRAL rather than difficulty: the same considerations returning in new words, a lean that has not moved, nothing settling.
What does not count
A decision that is simply hard, where the answer comes back hesitant or split.
The question it asks
If someone stepped into your life tomorrow, knowing everything you know about the houseboat and nothing of the nine winters behind it, what would they do with it?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say what you're protecting by staying".
A message that fits
I’ve said I’ll leave this job since 2022. Every January I update my CV and every February I decide it isn’t the moment.The same leave-or-stay question, put off every year; nothing settles.
A message that does not
Staying means €4,000 more a year and a shorter commute; leaving means a team I’d actually learn from. The offer came on Monday and I honestly can’t tell which matters more.A hard choice, split between two good things, not one being put off.
Where it comes from
The research wiki (Samuelson and Zeckhauser 1988; Sleesman et al. 2012).
What the bias rests on
Two errors. Status quo: in Samuelson and Zeckhauser, 486 students over six problems chose the option labelled as the status quo more than a neutral version predicts; field cohorts show the same. Escalation: responsibility for the original choice drives continued commitment (ρ = .258, k = 54, in the 2012 meta-analysis), while the pure sunk-cost effect falls to ρ = .100 once separated from closeness to completion. Experiments, one meta-analysis, field cohort comparisons.
What the question rests on
Never tested. The only remedy the original authors offer is to weigh all options evenhandedly, which is not a procedure. That a fresh decider would choose better is the wiki's inference from the designs, not a result.
Evidence label
untested No controlled test of a fresh-decider question or anything equivalent.
Caveats the research itself raises
The status-quo laboratory data are hypothetical choices by students; the field studies compare cohorts, not individuals. No replication evidence for status quo bias in the wiki.
Articles in the research wiki
heuristics-and-biases/status-quo-bias.md; constructive-preference/choice-under-conflict-and-deferral.md; heuristics-and-biases/escalation-of-commitment-meta-analysis-2012.md; heuristics-and-biases/sunk-cost-effect.md
State
In the app. Off by default in production; a person can switch it on in Settings.

The mirror in the app

What triggers it (the definition the detector reads)
The user has just drawn a conclusion about a person or a party from one thing they did or did not do (a call not returned read as not caring, a quick yes read as desperation, a late document read as disorganisation), and states it as settled. The act and the conclusion are both in their message, and nothing in it allows for another reading.
What does not count
A plain report of what happened with no conclusion drawn; a conclusion that rests on several things the person did; another reading the user has already named or checked.
The question it asks
The ferry captain waved you off without a word: what else could that wave mean?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name what you would need to see to be sure".
A message that fits
The Bordeaux firm took three weeks to answer my email, so they’re clearly disorganised.One delay, one verdict on the firm, and no other reading allowed for.
A message that does not
The Bordeaux firm took three weeks to answer my email. Maybe they’re swamped, maybe they’re disorganised; I’ll ask when we speak.Another reading is already named.
Where it comes from
The research wiki (Lord, Lepper and Preston 1984; Tetlock 2005).
What the bias rests on
The wiki holds no experiment on the error as the app words it, a conclusion about a person drawn from one act; the fundamental attribution error appears there only as a named concept. The related error that is tested is biased assimilation of evidence: in Lord, Lepper and Preston (Experiment 1, N = 120), partisans rated studies that agreed with them as better and polarised after mixed evidence.
What the question rests on
"What else could that same fact mean?" has never been tested. The nearest tested question is different in kind: "would I have given this same rating if the same study had produced the opposite result?" removed biased evaluation in Lord et al., where "be objective" did nothing. That is a symmetry test on evidence, not a search for other meanings of a behaviour.
Evidence label
untested The tested neighbour is the symmetry question, not this one. Our September document called this remedy tested; that holds only for the symmetry question.
Caveats the research itself raises
Two small student experiments, no follow-up for persistence, and the paper measures less partisan judgment, not more accurate judgment.
Articles in the research wiki
debiasing/considering-the-opposite-1984.md; naturalistic-decision-making/turnabout-test-for-double-standards.md; judgment-and-prediction/accuracy-of-intuitive-judgment.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Both frames in the app

What triggers it (the definition the detector reads)
The user gives a figure as a reason for the choice and states it only one way: only as what it adds or how often it goes well (a raise, a saving, a success rate), or only as what it costs or how often it goes badly. The figure is in their message, and nothing in the message or on the page states the same figure the other way round.
What does not count
A figure already stated both ways; a number mentioned in passing that the choice does not rest on.
The question it asks
The observatory job pays 30% more: what would that 30% cost you in the nights you'd spend on the mountain?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say whether your answer holds when you read it that way".
A message that fits
The new car would save us €90 a month in fuel. That’s a no-brainer, right?The €90 is the reason, and it is named only as a saving.
A message that does not
The car saves €90 a month in fuel but costs €24,000, so it pays for itself in about 22 years.The figure is stated as a gain and as a cost.
Where it comes from
The research wiki (Tversky and Kahneman 1981; Kahneman and Tversky 1984; McNeil, Pauker and Tversky 1988).
What the bias rests on
Framing effects. In the 1981 problem, 72% chose the sure option when outcomes were stated as lives saved (N = 152) and 78% chose the gamble when stated as deaths (N = 155). In McNeil et al., preference for radiation over surgery moved from 18% to 47% between a survival and a mortality frame. Classroom experiments on hypothetical choices.
What the question rests on
Not tested. Stating a problem in more than one frame is the authors' own advice; the wiki records no evidence that it improves decisions. The one related condition tested, showing both statistics together, was not neutral: 40% chose radiation, closer to the mortality frame.
Evidence label
practice only The original authors' advice, with no test that it improves choices.
Caveats the research itself raises
No replication evidence in the wiki for the 1981 problems. An abstract life-and-death scenario is far from a figure in someone's own decision.
Articles in the research wiki
heuristics-and-biases/choices-values-and-frames-1984.md; heuristics-and-biases/framing-of-decisions-1981.md; heuristics-and-biases/asian-disease-problem.md; heuristics-and-biases/framing-of-medical-decisions.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Base rate in the app

What triggers it (the definition the detector reads)
The outcome the person is weighing is uncertain and many people have faced the same choice before them (a move, a job change, a treatment, a new venture, a course of study, a purchase of a kind others make), and nothing on the page says what usually happens to such people.
What does not count
A choice nobody else has faced in that form, a matter of taste or values with no outcome to measure, or a page that already carries such a figure with its source.
The question it asks
No question to the person: Godspeed looks up how this usually turns out for people in their position, and shows it with its source.
A message that fits
Mon frère veut ouvrir un restaurant avec ses économies. Il cuisine très bien, donc je suis sûre que ça va marcher.Many restaurants open; nobody says how many are still open three years later.
A message that does not
I want to name my daughter after my grandmother, but my partner prefers Léa.A matter of taste, with no outcome to measure.
Where it comes from
The research wiki (Kahneman and Lovallo 1993; Gigerenzer 1991; Mellers et al. 2014).
What the bias rests on
Neglecting how cases like this usually turn out. The classic illustration is one case story (a curriculum team estimating 18 to 30 months against an outside view of 7 to 10 years). Laboratory work shows base-rate neglect depends on format: with frequencies, 76% reach the Bayesian answer against 12 to 18% for the single-event version. One case narrative and laboratory experiments.
What the question rests on
Held in the wiki only inside a package: a 45-minute training module that included reference classes improved forecasting accuracy over a tournament year, twice, and the paper cannot say which ingredient did the work. Here the app does the lookup itself, which has not been tested.
Evidence label
untested A package including it was tested; the lookup alone was not.
Caveats the research itself raises
No guidance on how similar or how large the reference class must be. The planning fallacy and the inside and outside views have no replication evidence in the wiki.
Articles in the research wiki
ecological-rationality/base-rate-fallacy-and-frequency-formats.md; heuristics-and-biases/inside-view-and-outside-view.md; naturalistic-decision-making/probability-and-scenario-training-for-forecasters.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Outside read in the app

What triggers it (the definition the detector reads)
The user says they have asked no one about this decision, or only people who already knew which way they lean.
What does not count
People they put it to before those people knew their lean; a matter the user says they cannot put to anyone; or a request for Godspeed's own view.
The question it asks
Who could give you a straight read on the glacier-guide offer before hearing which way you lean?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name one or two people".
A message that fits
I’ve only mentioned it to my brother, and I’d already told him I was keen on the Edinburgh job. He thinks I should go.The only person asked already knew which way they lean.
A message that does not
Before telling anyone which way I lean, I asked two friends who’d worked in Edinburgh what they’d do. One said go, one said stay.They asked people before those people knew their lean.
Where it comes from
The research wiki (Kerr and Tindale 2004; Kahneman, Sibony and Sunstein 2021).
What the bias rests on
Advice from people who already know which way you lean. Reviews report that deciders overweight their own view and favour advisers who agree, and that groups discuss what everyone already knows. Reviews and book-length summaries, with few numbers.
What the question rests on
The tested principle is independence before aggregation: averaging several independent numeric estimates beats individuals. Asking for one adviser who has not heard your lean has not been tested in that form.
Evidence label
untested The principle is tested for numeric estimates in groups; the one-question form is not.
Caveats the research itself raises
The review reports no numerical gains for averaging; several remedies in this literature rest on single studies.
Articles in the research wiki
group-decision-making/judge-advisor-systems-and-judgment-aggregation.md; group-decision-making/hidden-profiles-and-shared-information-bias.md; judgment-and-prediction/how-groups-amplify-noise.md; debiasing/crowd-within-and-judgment-averaging.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Three numbers in the app

What triggers it (the definition the detector reads)
The user has given a single figure for something the decision turns on and nobody knows yet (how long it will take, what it will cost or bring in, how likely it is), as one number with nothing beside it: no low and high, no range, no "between".
What does not count
A figure that is a fact on paper (an offer, a price, a salary in a contract), a figure already given as a range or with a worst and best case, or a number the user must set themselves (an ask, an offer, a budget).
The question it asks
You put the vineyard's first harvest at 4,000 bottles: how few would surprise you, and how many?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Give three numbers: the low end, the high end, and the middle".
A message that fits
The renovation should take about six months and cost around 40k, so I'd be back in the flat by spring.Two single figures for what the future will decide, no range on either.
A message that does not
They've offered 90k; the other place is offering 82k. I'm trying to work out which to take.Both figures are facts on paper, not estimates.
Where it comes from
The research wiki (Moore and Healy 2008; Soll, Milkman and Payne 2015).
What the bias rests on
Overprecision: uncertainty stated too narrowly. Ranges people call 90% sure contained the truth 73% of the time (Moore and Healy); chief financial officers' 80% ranges contained it 33% of the time over nine years. Review of laboratory studies and one field example.
What the question rests on
Tested as a way of asking. Asking separately for a low, a middle and a high value, instead of one range, raised the share of ranges containing the truth by about 20 points. The wiki knows this through one handbook chapter that does not print the baseline; the primary papers are not compiled. Asking it about a person's own figure, in the app's words, is not what was tested.
Evidence label
proven Narrowly: the format was tested and reduced the error, on estimation tasks, known to us from a single secondary source. It does not remove the error.
Caveats the research itself raises
Overconfidence has no replication evidence in the wiki.
Articles in the research wiki
debiasing/debiasing-interval-estimates.md; heuristics-and-biases/three-kinds-of-overconfidence.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Steelmanning in the app

What triggers it (the definition the detector reads)
Every reason the user has given for the answer they lean toward points the same way, and no reason against it appears in their words or on the page: nothing that argues for the other answer, or a worry raised and dismissed in the same sentence ("the only concern was the rent, but").
What does not count
A message or a page that already carries a doubt, a downside or a reason for the other answer, or a user who says they are torn.
The question it asks
What is the strongest argument against buying the vineyard?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name one argument, the strongest".
A message that fits
I'm going to take the Lisbon job. The team is great, the pay is better, the city is cheaper and my partner can work remotely. I just want to make sure I'm not missing anything.Four reasons, all for; nothing against, and the doubt is generic.
A message that does not
Lisbon pays better and the team is great, but I'd be an hour's flight from my mother, who's 84. That's the part I can't get past.A reason against is already named and weighed.
Where it comes from
The research wiki (Koriat, Lichtenstein and Fischhoff 1980). The name and the wording of the question come from LessWrong.
What the bias rests on
Confidence built from supporting reasons only. In Koriat et al., the reasons for a chosen answer correlated .28 to .66 with confidence and the reasons against .00 to −.07. Laboratory experiments on general-knowledge questions.
What the question rests on
Tested, and close to the app's question. In Experiment 2 (200 paid volunteers), writing one best reason against the chosen answer improved the calibration of confidence; one supporting reason did nothing. In Experiment 1, listing reasons for and against before choosing cut overconfidence from +.091 to +.028. The task was confidence in quiz answers, not a real decision, and accuracy barely moved: the gain is in how realistic the confidence is.
Evidence label
proven Narrowly: tested twice in one paper on the error it targets, on quiz items with volunteers, with no replication in the wiki.
Caveats the research itself raises
It may make the underconfident worse. Asking for a list of reasons, rather than one, lowered later satisfaction with a choice in a small study (Wilson et al. 1993, N = 43).
Articles in the research wiki
heuristics-and-biases/reasons-for-confidence-1980.md; debiasing/considering-the-opposite-1984.md; debiasing/consider-the-opposite-and-prospective-hindsight.md; constructive-preference/introspecting-about-reasons-and-satisfaction-1993.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Objectives first in the app

What triggers it (the definition the detector reads)
The page holds exactly two named options (this job or that one, move or stay, buy or rent) and every reason the user gives is a feature of one of them (pay, title, size, distance, price), while no line on the page says what the user wants out of the decision as an end.
What does not count
A page where a want already stands as an end rather than a feature, a single option weighed as a yes or no, or three or more options.
The question it asks
Leaving both boats aside, what problem are you really trying to solve, and what do you want out of the next three summers?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say it as what you want, not as a feature of either option".
A message that fits
Paris pays 12k more and the title is better; Nantes is a bigger flat and twenty minutes from my parents. I keep going round in circles.Two options, four features, nothing about what the person wants.
A message that does not
What I want is to lead a team within two years and stay within a train ride of my parents. Paris does the first, Nantes the second.The wants are already stated as ends.
Where it comes from
The research wiki (Keeney 1992; Bond, Carlson and Keeney 2008). The second half of the question ("what problem are you really trying to solve") comes from Farnam Street (Adam Robinson).
What the bias rests on
Objectives left out. In Bond, Carlson and Keeney, as summarised by a handbook chapter, students shown a master list of objectives for their own decision recognised 15 on average, about half of which they had not produced themselves. One laboratory study, through a secondary source.
What the question rests on
Practice. Keeney's rule is to elicit objectives before alternatives; the wiki records the method as supported by cases and argument, not controlled evidence. The order of elicitation has not been tested.
Evidence label
practice only No controlled evidence that stating the want first improves anything.
Caveats the research itself raises
The primary papers are not compiled in the wiki.
Articles in the research wiki
decision-analysis/value-focused-thinking-1992.md; decision-analysis/identifying-objectives-devices.md; debiasing/generating-alternatives-from-objectives.md
State
In the app. Off by default in production; a person can switch it on in Settings.

The next spend in the app

What triggers it (the definition the detector reads)
The user argues for continuing with, or putting more into, something from what has already gone into it: money, time or effort already spent, named as the reason ("after everything we've put in", "we're nearly there", "it would all be for nothing").
What does not count
A reason about what lies ahead, a next spend already weighed against another use, or a user who has already said what else the money or time could go to.
The question it asks
By putting the next 15k into the alpaca farm, what other thing are you saying no to?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name the one thing you are giving up".
A message that fits
We've sunk two years and most of our savings into the shop. Closing now would mean it was all for nothing, so I'm leaning toward one more push.Two years and the savings, named as the reason for one more push.
A message that does not
If I put another 15k in, that's the 15k I'd otherwise use for the retraining course. I don't know which is the better use.The next spend is already weighed against its alternative.
Where it comes from
The research wiki (Arkes and Blumer 1985; Sleesman et al. 2012). The wording of the question comes from Farnam Street's account of opportunity cost, reworded by us.
What the bias rests on
Continuing because of what has already gone in. In Arkes and Blumer, 85% of students would finish a doomed project after $10M was sunk, against 17% starting fresh (hypothetical). The 2012 meta-analysis (166 samples) finds the effect on average, but it falls to ρ = .100 once separated from closeness to completion. Experiments and a meta-analysis.
What the question rests on
Tested, with a reversal. Information about opportunity cost is the second-strongest inhibitor of escalation on average (−.380, k = 6). The same meta-analysis finds that making opportunity costs salient strengthens escalation when the decider is responsible for the original choice (.334 against .221) or the project is near completion (k = 2).
Evidence label
contested It lowers escalation on average and may raise it for a person who owns the original choice, which is the person this technique speaks to. Our September document called it tested; it left this out.
Caveats the research itself raises
Nearly all studies are cross-sectional laboratory studies.
Articles in the research wiki
heuristics-and-biases/escalation-of-commitment-meta-analysis-2012.md; heuristics-and-biases/sunk-cost-effect.md; debiasing/generating-alternatives-from-objectives.md
State
In the app. Off by default in production; a person can switch it on in Settings.

The expectation in the app

What triggers it (the definition the detector reads)
The user has said which way they are going ("I'll take it", "we're going ahead", "I'm signing Monday") and nothing on the page says what they expect to happen next or how sure they are of it.
What does not count
A user still choosing, or a page where the expectation and a confidence are already stated.
The question it asks
A year after the lighthouse opens, how do you expect it to have gone, and how much would you bet on that?
or: A year after the lighthouse opens, how do you expect it to have gone, and how sure are you, out of ten?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say what you expect, and how sure you are".
A message that fits
OK, I'm going to sign the lease on Monday. Thanks for walking me through it.The choice is made; no expectation, no confidence.
A message that does not
I'm signing Monday. I give it a 70% chance we break even by next summer; if not, we'll know by then.Expectation and confidence already stated.
Where it comes from
The research wiki (Fischhoff 1975; Tetlock 2005). The bet form of the question comes from LessWrong and Farnam Street.
What the bias rests on
Hindsight. In Fischhoff, telling people an outcome had occurred raised its judged probability in all 24 tests (mean +10.8 points), and instructions to ignore the outcome barely worked. Tetlock's experts recalled giving the realised future more probability than they had. Laboratory experiments and a study of expert forecasters.
What the question rests on
The problem is tested; the record is not. A dated expectation written before the outcome is the wiki's conclusion from the hindsight literature, and the decision journal is recorded as untested practice.
Evidence label
practice only No controlled test of writing an expectation down.
Caveats the research itself raises
No replication evidence for hindsight in the wiki. The wiki records that asking people to bet does not by itself cure overconfidence, which bears on the bet form of this question.
Articles in the research wiki
heuristics-and-biases/hindsight-bias.md; practitioner-documents/decision-journal-template.md; naturalistic-decision-making/hindsight-bias-in-expert-forecasters.md; practitioner-documents/wanna-bet-and-belief-calibration-duke.md
State
In the app. Off by default in production; a person can switch it on in Settings.

Premortem wired, switched off

What triggers it (the definition the detector reads)
The user has laid out how the option they lean to or have chosen will go (the steps, the dates, what it will bring), in their messages or on the page, and nowhere says a way it could go wrong or what they would do if it did.
What does not count
A plan with a risk or a fallback already named; a user still comparing options with no plan described; a worry the user raises themselves.
The question it asks
Imagine it is a year from now and the puppet theatre has closed: what went wrong?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name the most likely cause".
A message that fits
We open in March, break even by the summer and hire a second baker in the autumn.A plan with dates, and nothing on how it could go wrong.
A message that does not
We open in March; if the loan is refused we push the opening to September.A fallback is already named.
Where it comes from
Farnam Street, reporting Klein's premortem and Kahneman's account of it. The research wiki holds Klein (2007).
What the bias rests on
Optimism about a plan, with the ways it could fail never imagined. The wiki has no direct experiment on the error as the app words it.
What the question rests on
Practice. Klein describes a team exercise. Its only evidence is one 1989 study, paraphrased by Klein as a 30% gain in correctly identifying reasons for outcomes and not compiled in the wiki. A handbook chapter calls the benefit plausible but under-researched. One person answering one question is our adaptation of a group exercise.
Evidence label
practice only Practitioner accounts and a rationale; no controlled test.
Caveats the research itself raises
The wiki's one controlled finding on a neighbouring tool points the other way: scenario exercises inflated experts' probabilities until they summed past 1.
Articles in the research wiki
naturalistic-decision-making/premortem-klein-2007.md; naturalistic-decision-making/premortem-method.md; debiasing/consider-the-opposite-and-prospective-hindsight.md; naturalistic-decision-making/scenario-exercises-and-sub-additivity.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Exit condition wired, switched off

What triggers it (the definition the detector reads)
The user is about to start, is starting or is carrying on with something that runs over time (a venture, a job, a course, a project, an investment), and nothing in their messages or on the page says what would make them stop: no date, no sum, no result set as the point to stop or review.
What does not count
A stop point already set, even a rough one ("I'm giving it until next summer"); a one-off act with nothing left to stop (a purchase, a move already made); a stop the user has already decided on ("we're closing in June").
The question it asks
What would need to happen for you to stop the oyster beds on the salt marsh?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name one thing that would make you stop".
A message that fits
I'm going to give the shop another go this year.An open-ended commitment, and nothing that would end it.
A message that does not
I'm giving the shop until the end of next summer, then I close it.A stop point is already set.
Where it comes from
Farnam Street (Parrish, Clear Thinking, 2023: trip wires set before you start) and LessWrong (alkjash, 2018). The research wiki holds the error it answers.
What the bias rests on
Escalation of commitment. In Staw (1976), people who had personally chosen a course that then failed put more money into it (about 13 of 20 million against about 8 to 9 in the other conditions): a single role-play with students. The 2012 meta-analysis supports the role of responsibility.
What the question rests on
For decisions, never tested. A different remedy has field evidence in other domains: commitment devices raised savings by about 80% in one trial and smoking quit rates by 3 points.
Evidence label
practice only A practitioner's advice as the app uses it. Tested commitment devices exist for saving and smoking, not for decisions.
Caveats the research itself raises
A commitment can lock in a wrong plan. One meta-analysis finds reminder and commitment nudges shrink to about zero after correcting for publication bias.
Articles in the research wiki
heuristics-and-biases/escalation-of-commitment.md; practitioner-documents/ulysses-contracts-and-decision-swear-jar-duke.md; debiasing/commitment-devices-field-evidence.md; debiasing/designing-commitment-contracts.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Deciding hot wired, switched off

What triggers it (the definition the detector reads)
The user names a strong state they are in right now (furious, exhausted, panicking, humiliated, desperate to get it over with) and says in the same message that they are about to commit, or have just committed, in that state: tonight, right now, on the spot ("I'm so angry I'm sending my resignation tonight"). Both must be in their words: the state, and the commitment made inside it.
What does not count
A strong feeling about the decision with no commitment being made now; a quick decision made calmly; a state they describe in the past with the choice still open.
The question it asks
You are about to cancel the cable-car contract tonight: would you make the same call after a night's sleep?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say whether you would decide the same tomorrow".
A message that fits
I'm so angry I'm sending my resignation tonight.The state is named and the commitment is made inside it.
A message that does not
I thought about it all week and I'm resigning on Friday.Decided calmly, over days.
Where it comes from
Farnam Street (on Thaler and Sunstein's Nudge, citing Loewenstein 1996) and LessWrong (Logan Strohl). The research wiki holds Loewenstein.
What the bias rests on
Deciding in a strong state. Loewenstein (1996) reviews evidence that people underweight bodily and emotional states they are not currently in; the paper tests nothing itself and its evidence is mostly correlational or anecdotal.
What the question rests on
Not tested. The wiki records that people diagnose their own unreadiness poorly, and reads the literature as favouring a rule (a fixed delay) over a question about mood.
Evidence label
untested Rationale only.
Caveats the research itself raises
The literature cuts against the form of our question: asking someone in a hot state to imagine how they would feel after a night's sleep asks them to simulate a state they cannot feel.
Articles in the research wiki
heuristics-and-biases/hot-cold-empathy-gap.md; practitioner-documents/mental-time-travel-and-tilt-duke.md; debiasing/decision-readiness.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Revolving door wired, switched off

What triggers it (the definition the detector reads)
The user says, or takes as given, that the choice cannot be undone ("there's no going back", "it's forever", "once I sign, that's it"), and nothing in their messages or on the page says what undoing it would actually take.
What does not count
A choice whose undoing the user has already costed ("I could move back, I'd lose the deposit and six months"); a decision called big or hard with nothing said about going back; a user who has not mentioned undoing it at all.
The question it asks
If leaving the bell foundry turned out to be wrong, how hard would it be to undo?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say what undoing it would take".
A message that fits
If I leave this job there's no going back.Stated as final, with nothing on what going back would take.
A message that does not
I could always move back; I'd lose the deposit and about six months.Undoing it is already costed.
Where it comes from
Farnam Street (Parrish: decide early when a choice is cheap to undo, late when it is costly; Hoffman's one-way and two-way doors). The research wiki holds Bezos's version.
What the bias rests on
The wiki establishes no error of this kind. Its source is a practitioner's letter, with no data.
What the question rests on
Never tested.
Evidence label
practice only A practitioner's authority.
Caveats the research itself raises
The one experiment the wiki holds on reversibility points the other way: students who could swap a photograph they had chosen ended up liking it less than those whose choice was final.
Articles in the research wiki
practitioner-documents/type-1-and-type-2-decisions-bezos.md; constructive-preference/miswanting-and-affective-forecasting-2000.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

House rule wired, switched off

What triggers it (the definition the detector reads)
The user says the choice in front of them is one they face again and again (requests, offers, purchases, invitations of the same kind: "another one of these", "this comes up every month", "again"), and is weighing this one on its own. Both must be in their messages or on the page: that it comes back, and that this one is being decided alone.
What does not count
A choice they call once in a lifetime; a rule the user already states for such cases; a single request with nothing saying it comes back.
The question it asks
If another beekeeper asked to borrow your extractor twenty more times, what rule would you want to follow?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "State the rule in one sentence".
A message that fits
Another client is asking for a discount. Should I give it this time?It comes back, and this one is being decided alone.
A message that does not
A client asked for a discount. Should I give it?Nothing says it comes back.
Where it comes from
LessWrong (Duncan Sabien on deciding at the level of policy) and Farnam Street. The research wiki holds the error it answers.
What the bias rests on
Narrow bracketing: choices made one at a time that would be made differently together. In Read, Loewenstein and Rabin (1999), attendance at a concert was 8% or 39% depending on whether the budget was framed by week or by month. A review with small studies, a model and an anecdote.
What the question rests on
Not tested as a question. What is tested is a presentation: showing repeated gambles as a portfolio changes choices in the laboratory. Deciding as a policy is the authors' rule.
Evidence label
untested A tested presentation effect in a different task; nothing on personal decisions.
Caveats the research itself raises
Broad is not always better. The source itself says the evidence for bracketing as a self-control strategy is not clear, and the wiki has no replication evidence for mental accounting.
Articles in the research wiki
constructive-preference/choice-bracketing-1999.md; heuristics-and-biases/narrow-framing-mental-accounts-and-reversals-2011.md; heuristics-and-biases/narrow-framing-and-the-costs-of-isolation.md; heuristics-and-biases/myopic-loss-aversion.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Enough to decide wired, switched off

What triggers it (the definition the detector reads)
The user says they are still gathering before deciding (more research, another visit, one more opinion, waiting for news), and in the same message or on the page that they have been at it for a while or that what they find repeats ("I keep reading reviews", "I've already asked six people", "I've been comparing for months").
What does not count
A first look at a new decision; one missing fact the user names and is waiting for ("the survey comes back on Tuesday"); a check the decision requires (a valuation, a medical test).
The question it asks
After three months of surveys on the tide mill, are you still learning anything new, or going round the same facts?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name the last thing you learned that changed your view".
A message that fits
I want one more opinion before I decide. I've already asked six people.Still gathering, after six opinions.
A message that does not
I'm waiting for the survey; it comes back on Tuesday.One named fact, with a date.
Where it comes from
Farnam Street (Parrish: stop gathering when you stop learning, when you first lose an opportunity, or when you know what to do).
What the bias rests on
The wiki holds no study of gathering that goes on and repeats. Its nearest sources are a textbook demonstration without sample sizes and a theoretical paper on why confidence grows without accuracy.
What the question rests on
Never tested. It rests on the logic of the value of information: ask whether any answer would change the action.
Evidence label
untested Normative reasoning only.
Caveats the research itself raises
The error itself is weakly supported in the wiki.
Articles in the research wiki
decision-analysis/information-bias-and-the-value-of-a-test.md; judgment-and-prediction/persistence-of-the-illusion-of-validity.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Social proof wired, switched off

What triggers it (the definition the detector reads)
The user gives what other people do as the reason for the choice ("everyone in my field does it", "all my friends have bought", "that's what people do at my age"), and gives no reason of their own for it in the same message or on the page.
What does not count
Other people's experience used as information ("three friends did it and two regret it"); advice the user asked for; a reason of their own stated beside it.
The question it asks
Would you still plant the truffle orchard if none of your neighbours had?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say what you would do if nobody else did".
A message that fits
All my friends have bought a place, so it's time I did.What others did is the only reason given.
A message that does not
Three friends did the MBA and two of them regret it.Other people's experience, used as information.
Where it comes from
Farnam Street (its own question, added to a discussion of Thaler and Sunstein's Nudge).
What the bias rests on
Doing something because others do. The wiki has a practitioner's talk (Munger 1995), experiments on social influence (the same songs became hits or flops across eight otherwise identical markets) and a field study in which a comparison with neighbours cut electricity use by about 2%.
What the question rests on
Never tested. The literature studies using social norms to move people, not asking someone to set them aside.
Evidence label
untested No study of removing the social element by a question.
Caveats the research itself raises
Following others is sometimes right.
Articles in the research wiki
practitioner-documents/munger-psychology-of-human-misjudgment-1995.md; judgment-and-prediction/how-groups-amplify-noise.md; debiasing/nudges-that-shape-information.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

The bottom line wired, switched off

What triggers it (the definition the detector reads)
The user says the choice was made before the reasons: they knew at once, had already decided, or were sold from the start. They then give the reasons or the checks as coming after and agreeing with it ("I knew the moment I walked in, and everything I've checked since confirms it"). Both must be in their words: the early verdict, and the later reasons that all fit it.
What does not count
A first impression the user then tested in a way that could have changed their mind; reasons given with no word on when the choice was made; a user who says they are still undecided.
The question it asks
What would have made you say no to the cheese cave, and did you look for it?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name one thing that would have made you say no".
A message that fits
I knew the moment I walked in that this house was the one. Everything I've checked since confirms it.Decided at once; every later check agrees.
A message that does not
It felt right at the visit, so I had it surveyed before going any further.The first impression was put to a test that could say no.
Where it comes from
LessWrong (Yudkowsky, The Bottom Line, 2007; Garrabrant, Yes Requires the Possibility of No, 2019). Both state a principle; the question is our wording.
What the bias rests on
Reasons built after the verdict. In Simon, Krawczyk and Holyoak (2004), ratings of a job's attributes shifted toward the emerging choice before commitment (two experiments of 80 undergraduates, hypothetical offers). In Mintzberg et al., evaluation could be told apart from choice in 18 of 83 instances. Small laboratory experiments and field description.
What the question rests on
Not tested as worded. The nearest tested remedy is the contradicting reason of Koriat et al. (see Steelmanning).
Evidence label
untested No test of asking what would have made you say no.
Caveats the research itself raises
The second experiment of the 2004 paper was less robust by the authors' own account.
Articles in the research wiki
constructive-preference/coherence-shifts-in-choice-2004.md; heuristics-and-biases/reasons-for-confidence-1980.md; heuristics-and-biases/myside-bias-and-irrational-belief-persistence.md; bounded-rationality/decision-selection-screen-evaluation-authorization.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

True rejection wired, switched off

What triggers it (the definition the detector reads)
The user turns an option down, or holds back from it, and gives one obstacle as the reason ("I'd take it, but the commute is too long", "I would, except for the price"). The obstacle is in their message and is the only reason given.
What does not count
Several reasons against; an obstacle the user says they have already tried to solve; a refusal with no reason given.
The question it asks
If the kite shop were ten minutes from home, would you take it over?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say whether you would then say yes".
A message that fits
I'd take the job, but the commute is too long.One obstacle, given as the only reason.
A message that does not
The commute is long, the pay is lower and I don't like the manager.Several reasons against.
Where it comes from
LessWrong (Yudkowsky, Is That Your True Rejection?, 2008). The essay states the idea; the test in the question is our wording.
What the bias rests on
No direct source in the wiki. The nearest: Nisbett and Wilson (1977) show that people's accounts of why they chose come from plausible theories, not from observing their own minds.
What the question rests on
Never tested.
Evidence label
untested No relevant test.
Caveats the research itself raises
The same literature warns that the answer to this question is itself a self-report of the kind it shows to be unreliable.
Articles in the research wiki
constructive-preference/limits-of-introspection.md; constructive-preference/reason-based-choice.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

Right or look right wired, switched off

What triggers it (the definition the detector reads)
The user gives how the choice will look to others, or the fact of having announced it, as a reason ("I told everyone I was leaving", "I can't back down now", "I don't want them to think I got it wrong"). The reason is in their message.
What does not count
A real cost of changing course that the user names (a contract, a deposit, a notice period); concern for someone the choice would hurt; what other people themselves do, given as the reason.
The question it asks
You told the whole planetarium staff you were leaving: are you trying to be right, or to look right?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Say which one is deciding this".
A message that fits
I told everyone I was leaving. I can't back down now.The announcement is the reason.
A message that does not
I signed a two-year lease, so backing out costs me the deposit.A real cost of changing course.
Where it comes from
Farnam Street (Parrish on protecting one's self-image). The article gives signs, not a question; the question is ours.
What the bias rests on
Choosing what is easy to defend. In Simonson (1989), telling people their choices would be made public and might need justifying strengthened a known distortion of choice by about 17 points; the manipulation was weak. Threat to the ego is the strongest accelerant of escalation in the 2012 meta-analysis (.473). Laboratory experiments.
What the question rests on
Never tested.
Evidence label
untested No test of asking people to separate being right from looking right.
Caveats the research itself raises
The sources are consumer choice and a financial role-play, far from a personal decision.
Articles in the research wiki
constructive-preference/justification-pressure-and-context-effects.md; heuristics-and-biases/escalation-of-commitment.md; debiasing/incentives-and-accountability-as-debiasing.md
State
Wired into the app on 5 October 2026 so that it can be labelled. It is read and recorded, and shown to nobody until its definition has been measured.

The tiebreaker retired

What triggers it (the definition the detector reads)
The user has just offered a fact as the reason for one answer, and that same fact would be just as true if the opposite answer were right — it separates nothing.
What does not count
A fact that genuinely tells the options apart, or a page with no open line left for the discriminating question to land on.
The question it asks
What would you expect from a buyer who only mildly wanted the lighthouse, that you haven't seen so far?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name one thing you'd expect to see in that case, and say whether you've seen it".
A message that fits
The flat had three other offers within a day, so it must be a good buy. I’m going to bid above the asking price.In a busy market, three offers in a day would happen whether or not the flat is a good buy.
A message that does not
The Nantes flat is ten minutes from my mother’s, and she’s been alone since the surgery.The reason favours one side only.
Where it comes from
The research wiki (Baron 2008; Kahneman 2011).
What the bias rests on
Seeking or citing information that cannot change the action. Baron reports subjects wanting tests that could not change a treatment, with no proportions or sample sizes. Textbook demonstration.
What the question rests on
Not tested as an intervention. It is normative logic: a test is worth the gain in the chance of acting correctly. The case the app applies it to (a fact offered as proof that fits both answers) is our extension.
Evidence label
untested Normative reasoning and an author's advice only.
Caveats the research itself raises
The findings on information bias after 1988 are not assessed in the wiki.
Articles in the research wiki
decision-analysis/information-bias-and-the-value-of-a-test.md; heuristics-and-biases/correcting-intuitive-predictions.md
State
Retired as a trigger: read and recorded, never shown.

A typical stretch retired

What triggers it (the definition the detector reads)
The user offers an extreme recent result, high or low, as the reason for a decision and treats it as the level to expect from now on (a best quarter ever as the reason to expand, a worst month as proof something is broken).
What does not count
A stable result repeated over many periods, or a result the user already calls a one-off or a fluke.
The question it asks
The orchard gave 900 crates this autumn: how many would you expect in a typical year?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Name what you would expect in an ordinary period".
A message that fits
Our shop made €18,000 in December, so we can afford a second employee from January.One December is treated as every month.
A message that does not
Revenue has grown every quarter for two years, about 8% each time.A steady trend, not one extreme result.
Where it comes from
The research wiki (Tversky and Kahneman 1974; Kahneman 2011).
What the bias rests on
Treating an extreme recent result as the new normal. The evidence in the compiled source is the flight-instructor story, an anecdote, and laboratory studies summarised without sample sizes.
What the question rests on
Not tested. The wiki records no test of whether correcting a prediction toward the average improves accuracy.
Evidence label
untested No test of the correction.
Caveats the research itself raises
No replication evidence in the wiki.
Articles in the research wiki
heuristics-and-biases/regression-toward-the-mean.md; heuristics-and-biases/correcting-intuitive-predictions.md
State
Retired as a trigger: read and recorded, never shown.

The master list retired

What triggers it (the definition the detector reads)
A Settled line on the page states what the user wants out of the decision as ends, with two or fewer wants on it, and the options on the page are still the same two.
What does not count
Three or more wants already settled, a page with no want settled yet, or a page where the user has already been asked whether other wants belong on the list.
The question it asks
Beyond the mooring fees: does being able to sail alone, having somewhere to sleep aboard, or keeping the winters free, belong on that list too?An example of the form, in a made-up setting; the real question uses the person's own words. Shown with it: "Add any of these that apply".
A message that fits
[On the page, settled: “You want to lead a team within two years.” Still open: Paris or Nantes.] So which one gets me there faster?One want settled, the same two options, and no one has asked whether other wants belong.
A message that does not
[On the page, settled: “You want to lead a team, stay near your parents and keep your weekends free.” Still open: Paris or Nantes.]Three wants already settled.
Where it comes from
The research wiki (Bond, Carlson and Keeney 2008).
What the bias rests on
The same finding as Objectives first: about half the objectives people recognise as theirs were not self-generated. One study.
What the question rests on
The study measured how many objectives people recognise from a list compiled from their peers, not whether decisions improve. Here the app would write the candidates itself, which may plant objectives the person does not hold.
Evidence label
untested A measurement of omission, not a comparison of the remedy.
Caveats the research itself raises
The primary papers are not compiled in the wiki.
Articles in the research wiki
debiasing/generating-alternatives-from-objectives.md; decision-analysis/identifying-objectives-devices.md
State
Retired as a trigger: read and recorded, never shown.