Chapter 7. Judging Before and After the Answer

Five of WSTiger's records sit in this chapter: two set in different times around one subject, three about what a loss proves, what a profit excuses and what evidence may be counted. Some of the five ask how a decision should be judged once the answer to it is known. The other two have no answer yet: in one, a position is held after its pre-defined invalidation conditions have been met, and the loss is not confirmed while the holding continues; in the other, a validated strategy is losing in live trading while the predefined boundaries and the normality of the data have not yet been checked, so what would count as the answer is itself part of the question. The chapter moves from the psychology of judging to the design of the records and then to the question of what an outcome can prove. It reads the five through three bodies of work. The first describes how people judge probabilities when they have to judge at all. The second describes how knowledge of an outcome changes the judgment made afterwards. The third treats a long record of stated probabilities as the proper test of a forecaster. The three do not contradict each other; they answer different questions, and the chapter keeps them apart. Each record is a synthetic setting published for study, and no record here reports a real trade, a real firm or a real outcome. Where a reading is this study's own rather than something a record states, it is marked as this study's reading; where the records are silent, this chapter says so rather than filling the silence.

7.1Heuristics and loss

The note on Amos Tversky and Daniel Kahneman, "Judgment under Uncertainty: Heuristics and Biases" (1974), names three heuristics: representativeness, availability, and anchoring and adjustment. The note breaks down no single heuristic further, and this chapter does not add to it. What the note carries is the claim that attaches to all three together: these shortcuts can be relied on most of the time and still produce error that runs in a predictable direction. The 1974 work therefore treats the errors as systematic and predictable rather than random.

The note on "Prospect Theory: An Analysis of Decision under Risk" (1979), the work of Daniel Kahneman and Amos Tversky, gives this chapter one more claim it needs: probabilities are weighted, not taken as they are, so the weight a person acts on is not the probability as it is stated. It is a claim about a person judging at a moment. It says nothing about when the answer to a decision will arrive, who writes down the reason for it, or who waits and pays while the answer is still out.

That is the limit of the fit, and it is worth stating plainly. This study reads both works as the psychology that sits inside any record about judging decisions. The records in this chapter are not experiments, and the works are not records; they cannot be matched cell for cell. What the works give is a vocabulary for what the records worry about: that a person facing a loss will measure it against a reference point, and that a person judging probabilities will not act on the probabilities as stated.

One thing in this chapter's cases is legible through that vocabulary, and one is not. The legible one is the holding of an invalidated position, which returns in section 7.3: the record's question there asks whether a position may be held merely because the loss is not desired to be confirmed. This study reads that as a reference-point effect made into a decision about time, since the record links the holding of the position with not confirming the loss. The other thing — the time between a question and its answer — the works do not touch at all, and that is the subject the rest of this study returns to.

One more point about loss, drawn from the same 1979 work as it stands in the note. In the two trades a record might place side by side, one that broke a rule and made money and one that followed the rule and lost, the profitable breach is a gain and the planned loss is a loss. The records treat that asymmetry as a problem to be designed around rather than a fact to be corrected, and section 7.3 shows how they write it.

7.2The reason written before the outcome

The past record is set in 1956 as a synthetic setting, and the present record in 2026 as another. They belong to the same thread in WSTiger's atlas, which carries the same insight in both: that if the reason at the time of a decision is not recorded, memory and evaluation will rewrite it to fit the outcome once the outcome has occurred, so that a good decision and a good result can no longer be told apart.

Synthetic scene. In the past setting, a virtual trader sits at a desk where each decision to send a ship is made one voyage at a time. Before any ship leaves, the trader writes in a log the reason for the choice and the expected risks, and the log is what any later re-evaluation would have to work from. After a storm takes one ship, the business partner accuses the trader of having made a wrong judgment about the decision that lost it. What is waited for is the evaluation of the judgment and the reading of the page: the partner's eye is on the loss, the trader's hand is on the entry that was written earlier.

The three options name three different courts in which the same decision might be tried: by what happened, by what was known and written when it happened, or by what has happened in many such cases. The record's example judgment distinguishes the first from the other two, and does so by naming two kinds of information as separate entries:

The condition for revisiting this record is itself written in terms of the log: if the log shows a record of ignoring risk signals that were knowable at the time, the judgment is reviewed again. That is a condition about the reason, not about the storm. The record's ledger cells leave the decider unsettled, the horizon as unknown, the bearer of the waiting as unknown and the human place as human-decides (implied); the coders read the strict H2 cell as supports and the strict H5 cell as unclear.

Synthetic scene. In the present setting, a fictional desk reviews a trade that produced a large loss. The review report is written after the loss and reaches the conclusion that the signal was clear at the time. The pre-trade record that should show the basis for the judgment leaves almost no trace of it, and the counter-basis almost none at all. What is waited for is the reconstruction: whether the conclusion in the report survives when only the material from before the decision time is used.

Here the first option is the report itself, and the third moves the record forward in time, requiring that a record exist before the next decision rather than after its result. The example judgment names the phenomenon at issue directly, as clarity that appears only after a result, and then describes the test that would separate it from the uncertainty that was there at the time:

The record's condition for revisiting is again written in the future tense, and it names two things: whether the reconstruction confirms that the signal was clear at the time, and whether the obligation to record before decisions substantially harms decision speed. The second of those is a cost that nobody has measured in the record; the ledger's horizon cell for this case reads unsettled, and the bearer cell reads those nearest the stoppage (implied).

Two works name what these records are trying to prevent. One is Baruch Fischhoff's "Hindsight ≠ Foresight: The Effect of Outcome Knowledge on Judgment under Uncertainty" (1975), with the creeping determinism it names. The other is Jonathan Baron and John C. Hershey's "Outcome Bias in Decision Evaluation" (1988), where the rating given to a decision moves with the outcome that follows it. In the first work the error is about probability; in the second it is about the quality assigned to a decision. On the question of awareness the two notes differ: Fischhoff's reports that people do not realise the knowledge has changed their judgment, while the Baron and Hershey note says nothing about whether the people rating a decision can see the influence on them. Both works turn on a single outcome and the judgment made after it is known.

Read beside the two records, the works sharpen what is at stake in the past setting. A ship lost to a storm is a single outcome; a business partner's charge that the judgment was wrong is a single decision judged after it. This study reads Fischhoff's pattern as a risk inside the partner's charge: how likely the loss seemed at the time may already have been reshaped by knowing about it, and the partner may not see the reshaping. This study reads the Baron and Hershey pattern as a risk to any later rating of the decision: the rating may move with the outcome that follows it, so that a decision followed by loss is judged more harshly than the same decision followed by no loss. The record's design is an answer to both: it separates the information that was available from the information learned afterwards, and it asks for the judgment to be re-evaluated from the first kind alone.

The present setting changes the subject without changing the problem. There the loss is already on the books, and the question is not what the storm proved but what a post-mortem report may prove about the person who made the call. Here the ledger's decider cell is team-or-organisation (implied) and the bearer cell names those nearest the stoppage (implied). This study reads the two settings as one problem at two distances: in the past setting the reason exists on paper and the argument is about which paper to trust; in the present setting the reason barely exists and the argument is about who may be blamed without it.

What the records leave out is worth naming. Neither states how long a reconstruction takes, what the obligation to record in advance would cost in time, or by when any of these judgments must be settled. The two conditions that appear in the records are written as things that would trigger a review, not as dates or thresholds; only 5 of the 53 conditions across the whole book are observable in the sense of naming a number, a date, an event or a role, and Chapter 13 takes that count in full. The records also contain no outcome: the ship's fate and the trade's result are already inside the setting, and nothing is added afterwards.

7.3Results and rules

Three present records ask what a loss and a profit each prove, and when evidence may be counted. All three carry the same setting date, 2026, in the records; what differs between the reason and the answer is inside the record rather than outside it.

Synthetic scene. On one desk two trades sit in the same review. One broke the risk limit and produced a profit; the other was executed as planned and produced a loss. The review that follows reads like a table with two columns: what was done, and what happened. What is waited for is the ruling on whether a rule is worth keeping after a single loss, and whether a breach can be excused by the money it made.

The example judgment keeps the two kinds of assessment apart and refuses both easy repairs at once:

On the A5 axis — what would make the judgment change — the record's condition is stated as a deterioration in the net performance of rule-compliant trades across a sufficient follow-up sample, with an accompanying instruction that past assessments are not revised to fit individual results. The first half of that condition needs a sample; the second half forbids the practice the whole record is against. On the ledger's A6 axis the setting is present, and the horizon cell reads unknown: the record does not say how long the follow-up sample takes, and the decider cell reads unsettled, so the record does not say who keeps the two kinds of assessment apart over that time. This study reads the design of this record as an answer to a problem the works of section 7.2 name — the same decision rated differently once its result is known — carried out not in one person's head but in the structure of the assessment, where the result assessment and the process assessment sit in independent fields.

Synthetic scene. A position is still open on a screen whose pre-entry hypothesis has been invalidated. The stop has not been executed, and the reason given for the delay is that the position might still be recovered to break-even. The trader's eye is on the price that would end the loss; the current confirmed market structure that the record lists as required evidence is not something this setting displays. What is waited for is the termination order and whether it can be executed, and meanwhile the loss stays unconfirmed on the study.

The record's example judgment draws the line the options do not: an invalidated position goes through the procedure set before the trade, and any extension of holding or movement of the stop is evaluated not as maintenance of the original trade but as a new risk decision. The ledger puts the decider here as individual (implied), the bearer of the waiting as the decider (implied), the reversal constraint as costly and the human place as ai-or-rule-decides (implied); the strict H4 cell is not-applicable and the lenient cell is against. On A5 the record's condition for revisiting names a new independent hypothesis with a validated risk budget, and the next check asks that the termination reason of the original judgment and the start of the new hypothesis be kept separate — a separation that can be written down, though nobody in the record is named to keep it. On A6 the setting is present; the record sets no expiry for the recovery expectation and names no date by which the position must be closed. The pre-defined invalidation conditions have already been met and a termination procedure stands ready, so the duration of the deferral is left unstated. This study reads that absence as the record's own form of loss aversion made structural: the reference point does not move because the position has not been closed, and the record's answer is to make the closing independent of the reference point by making the procedure pre-defined.

Synthetic scene. A strategy that has been validated loses in live trading. Screens show the boundary fixed before the trades began, the simulated fills that produced the validation, and the live fills that differ from them. A meeting has to decide whether what is happening is a period of loss or an undermined hypothesis. What is waited for is the predefined boundary being read against the live record, and the four kinds of profit-and-loss alert that the desk is told to divide into.

The example judgment separates two things the loss could mean — a loss, or an undermined hypothesis — and orders the account risk limits to be applied first, the predefined boundaries to be used in review, and the criteria not to be changed favourably after the results are seen. The record's condition for revisiting names confirmed risk-limit breaches and structural divergence as triggers for the stop and scaling-down procedures, and bars reverting to the original state before recovery and re-validation; the next check is a classification into risk, data, execution and hypothesis, which is a thing a reader could apply. The ledger's decider cell is unsettled, the horizon cell unsettled, the reversal cells unsettled, and the strict H2 cell reads against, since the record writes its triggers in terms a reader could observe.

Across the three, the design pattern is the same: the reason and the rule are fixed before the result is known, and the record asks what evidence would count against them afterwards. The cost of that design falls in the waiting. A rule-compliant sample takes time to accumulate; a break-even recovery expectation takes time that the account funds; a review against predefined boundaries takes staff time that no record quantifies. In none of the three does the record name a deadline. Of the 53 records in the study, 0 have a deadline in the record. And in none of the three does the record name a single person as decider; the decider cells read unsettled or individual. This study reads the pattern as a form-level answer to outcome bias: the decision is made twice, once on the rule and once on the result, and the record tries to keep the two from contaminating each other. What the form cannot do is tell whether the second decision arrives on time.

7.4The debate: is an outcome evidence about a decision?

Two positions face each other in this part of the literature, and the books in the debate do not obviously contradict each other, because they are about different objects.

The first position, from Fischhoff (1975) and Baron and Hershey (1988), concerns judgments made once an outcome is known. Fischhoff reports that outcome knowledge makes an outcome look more likely than it would have looked without it, and that people do not realise the knowledge changed their judgment. Baron and Hershey report that the same decision, made on the same information, is rated better when its outcome is good than when it is bad. Taken together, the two works say that knowledge of an outcome colours the judgment that follows it: it changes the witness.

The second position comes from the scoring tradition. Glenn W. Brier, in "Verification of Forecasts Expressed in Terms of Probability" (1950), proposes a score for probability forecasts: the mean squared difference between the forecast probabilities and what happened. The note adds the point that matters for the debate: a forecaster cannot improve that score by stating anything other than the probability believed. A score with that property is a test of stated beliefs against results. Philip E. Tetlock, in "Expert Political Judgment" (2005), reports a long study of expert forecasts on political and economic questions: they were poorly calibrated and often no better than simple rules; he also reports that those who drew on many ideas and revised readily did better than those committed to one big idea. The finding concerns a record of forecasts, not a single decision, and the verdict on an expert comes from the record as a whole.

The two positions can therefore be read as answering different questions. Fischhoff asks how likely an outcome looks once it is known; Baron and Hershey ask how a person rates one decision after one outcome. Brier and Tetlock ask how a person rates a forecaster after many. The experiments show that both judgments are coloured; this study reads the scoring view as accepting that a single outcome is weak evidence and treating accumulated outcomes as the test that a single outcome is not. This study reports both readings as their notes give them and does not merge them. Where they do meet is the practical question in these records: what should be done with the outcome when it arrives, and what should exist before it arrives.

A third position comes from John Kay and Mervyn King, "Radical Uncertainty: Decision-Making Beyond the Numbers" (2020). Their note says that many decisions face uncertainty that probabilities cannot capture, and recommends asking what is going on, reasoning with narratives, and choosing strategies that are robust and resilient to what cannot be foreseen rather than optimising on invented numbers. This study reads their answer to "is an outcome evidence?" as neither a yes nor a no: a single outcome is not a test, and neither is a single story, and the decision worth defending is the one that survives a range of outcomes it was not sized to predict.

Where these records bear on the debate is specific. All five keep the reason before the outcome as a matter of form: the past record writes the reason in a log before each voyage; the present post-mortem record offers the third option of mandating a record before the decision rather than after it; the underperformance record refuses to change criteria after seeing results; the invalidated-position record requires that the existing judgment record be not retroactively amended, and that the end of one hypothesis be kept separate from the start of another; the rule-and-result record requires that the result assessment and the process assessment be stored in independent fields. Read together, the five are a design answer to Fischhoff's and Baron and Hershey's problem: fix what was believed before the result exists, so that later judgment has something to be measured against that the outcome did not rewrite.

The records cannot settle the debate, and the reason is structural. They are synthetic settings published for study, and none of them contains an outcome to score: no follow-up sample, no set of forecasts to check, no series of similar decisions whose results could be counted. Brier's score cannot be computed on a record; Tetlock's method cannot be applied to a record; Fischhoff's effect cannot be measured on a record, because there is no experiment in which one group knows the outcome and another does not. What a record can do is show how a studio writes the moment before an outcome: what is fixed, what is left open, and what the form says would make the judgement change. This study reads that as evidence about record design, not about whether outcome evidence is or is not sound. Two further limits belong here. The first two positions in the debate concern people judging or forecasting under test conditions; several of these records place judgment with a team or an automated rule, and the ledger's human-role cell reads no (implied) 38 times and unsettled 11 times across the study. And the record form itself guarantees a reason and a revisit condition, because the form asks for them, so their presence in these five cases cannot be counted as a finding that real desks record reasons in advance; what can be counted is how the reason is positioned relative to the outcome, and Chapter 13 takes the observability of the revisit conditions in full.

7.5Hands on

The five records of this chapter are a small sample of the comparison the ledger makes across all 53. The table below gathers the axis rows from "The ledger's numbers", with what each cell counts. Each row is a count with its label attached, and the later chapters work one axis at a time from these rows.

The options across the study number 159. Both coders agree that 44 of them reduce the immediate loss. What those options give up, as both coders read it, is resilience 18 times, long-term-stability 17, ethical-coherence 13 and evolvability 7. That is the shape of the trade this chapter has been circling: the records repeatedly offer an option that cuts the loss now, and the records repeatedly write down what that option gives up afterwards.

Two of the eight pre-registered hypotheses are visible in this chapter without being settled here. H2, that few decisions write down in advance what would reverse them and those that do write it in words no one could observe, does not read one way across this chapter's cases: all five carry a revisit condition, but the per-case H2 cells recorded above include against as well as supports. The whole-book count on observability is 5 of 53 (9%); the full treatment of conditions and of H2 is Chapter 13's. H8, that the options which cut the immediate loss most often give up evolvability and resilience among the four higher goals, has a verdict from "The ledger's numbers" of supports in both readings, 24 to 13 in each, and its treatment belongs to Chapter 14. The first axis, who decides, is not settled here at all: the decider cells in these five cases are partly unsettled and otherwise split between team-or-organisation and individual, which is the shape the whole-book row takes at larger scale. Chapter 8 takes that axis apart, case by case, beginning with the records that place authority between a person, a team and an automated agent.

Hands on: the first axis, who decides (Chapter 8).