Week 9: The prior state

Use a fictional gardener's plant diagnosis and Fischhoff's hindsight-bias study to explain why a prediction recorded before checking the answer is different evidence from a claim about what you "already knew."

19 April 2027

Open the slides

Where Week 8 left us#

Week 8’s surprise question exposed a limit: the final answer cannot show what existed before the probe. This week adds a prior state to the ladder record. Pre-commitment means recording a prediction and reason before checking, then preserving them beside the revision. Our new fictional case changes the setting, while continuing the same question about evidence.

What can we learn from what someone thought before they checked?

This is a self-guided lesson. Both these notes and the slide deck contain the explanations you need; no presenter is assumed. Allow about 30 minutes, with another 15–20 minutes for the investigation. Use paper or a text document for your answers, and keep what you write — this week’s point depends on comparing an early answer with a later one. You need no gardening background. The people and incidents below are fictional.

What you will learn#

By the end, you should be able to explain what a prior state is and why recording one matters, describe why a person’s memory of “what I already suspected” is not reliable evidence once they know the answer, and explain what a recorded prior claim can and cannot establish about someone’s later, more accurate performance.

1. A plant, and a note written before looking anything up#

Marlow’s tomato plant has yellowing leaves that curl inward at the edges. Before consulting any reference, Marlow writes down a diagnosis and a confidence level: “Probably underwatering — the soil looks dry near the top and the leaves are curling like they’re trying to hold onto moisture. Confidence: medium.” Only after writing this does Marlow open a gardening reference.

Tomato leaves that yellow and curl can have more than one common cause: too little water, too much water (which can drown the roots and cause rot), or a nutrient shortfall. Each has a different fix, so the distinction matters in practice. For this lesson, though, the horticulture is secondary. What matters is what Marlow did before finding out which cause was correct.

2. Marlow revises#

The reference Marlow reads explains that overwatered tomato leaves often curl and yellow in a pattern that looks similar to underwatering at first glance, but the soil stays wet rather than dry, and the lower stem may show soft, dark patches from root rot. Marlow checks the soil again — it is wet two centimetres down, not dry — and finds soft dark patches near the base of the stem.

Marlow’s revised note: “Root rot from overwatering, not underwatering. The soil is wet, not dry — I misjudged that from the surface, which had dried out faster than the soil below it. The stem patches confirm root damage.” This explanation is more accurate than the first. But we now have two notes, written at two different times, for two different reasons — and that pairing is the interesting part.

3. Naming a prior state#

A prior state is a claim — a prediction, a diagnosis, an explanation — that someone commits to and records before they consult a source, receive feedback, or otherwise learn the answer. Marlow’s first note is a prior state. The revised note is not: it was written with the answer already partly visible. Recording a prior state fixes what someone thought at that specific point, before anything else could change it.

Why does this distinction matter enough to build a whole lesson around it? Because the alternative — simply asking someone afterward what they thought beforehand — turns out to be far less trustworthy than it sounds.

4. A study on exactly this problem#

Psychologist Baruch Fischhoff (1975) ran an experiment addressing this directly. Two groups of subjects read the same brief description of a real but unfamiliar historical event, with four possible outcomes listed. A “Before” group estimated how likely each outcome was, without knowing which had actually occurred. An “After” group was told which outcome had occurred, then asked to estimate — as if they did not know — how likely each outcome had seemed beforehand.

Across 24 such comparisons (varying the event, the outcome reported, and the subject group), the After group’s estimate of the reported outcome’s likelihood was higher than the Before group’s estimate every single time (sign test, p < .001). The average increase was 10.8 percentage points, and over 70% of individual After subjects rated the reported outcome as more likely than the average Before subject had. Simply learning what happened made people believe, in hindsight, that it had looked more likely all along — even though nothing about the original description had changed.

The detail that matters most for this course is Fischhoff’s further finding that subjects were largely unaware this had happened to them. In follow-up experiments, subjects who were told an outcome and then asked to reconstruct their own earlier judgment — or to guess what someone without outcome knowledge would say — still overestimated it, as though the hindsight-shifted belief felt like it had been there from the start. This is why asking someone afterward “did you already suspect it?” is weak evidence: the person answering is not lying, but their memory of the earlier belief has been quietly revised by knowing the outcome.

5. Back to Marlow#

Marlow’s case is fictional and far lower-stakes than Fischhoff’s historical judgments, but the same mechanism threatens it. If Marlow had not written the first note down, and we asked afterward “did you think it might be overwatering?”, Marlow’s sincere answer would carry exactly the risk the study describes — it might report a belief shaped by already knowing the answer, not the belief that existed before. Because Marlow wrote the note down before reading the reference, we have something a recalled version could never give us: an unedited record of the reasoning available at that point.

Placed side by side, the two notes show something a single final answer cannot. The final answer — “root rot from overwatering” — is correct and well-supported, but on its own it gives us no way to tell whether Marlow reasoned toward it, was told it outright, or copied it from somewhere. The prior note supplies exactly that missing piece: independent reasoning from visible cues, wrong about the cause but not unreasonable given what was visible at the time. A wrong first answer, honestly recorded, is still evidence — of what someone could infer from limited information, at a specific moment, before help arrived.

6. Pause and write: a second case, before you look anything up#

Allow four minutes. Marlow has a second plant: basil, growing on a sunny windowsill, wilting badly by mid-afternoon during a heatwave. Before reading further, write down what you think is wrong, why, and how confident you are (low / medium / high). Do not revise this note once written — you will compare it against a later answer.

7. Worked response: revising the basil diagnosis#

Suppose the basil looks wilted at 3pm but has fully recovered by the next morning, two days running. That pattern matches heat stress rather than underwatering: a genuinely dry plant would not fully recover overnight without being watered. If your prior note guessed “underwatering,” that guess was reasonable — the two look similar at first — but the overnight-recovery pattern is what distinguishes them. If your prior note guessed “heat stress,” you had that reasoning available before checking; record that plainly rather than downplaying it.

Whichever way your prior guess landed, you now have a recorded prediction and a recorded revision, written at genuinely different times with genuinely different information available. That pairing, not just the final correct-sounding answer, is the useful record. One instance does not establish that you reliably reason well under time pressure or with a real plant in front of you — it is a starting point, not a pattern.

8. A trap to watch for#

Once you have read the worked response, the heat-stress explanation will start to feel obvious. Resist rewriting your memory of your own prior note. If it said “underwatering,” that is what you thought at the time — leave it exactly as written, the way Fischhoff’s subjects could not stop their own memories from doing invisibly. This is not about admitting fault. It is about keeping the record honest enough to be useful later, including to you.

9. Calibration needs a prior claim to exist at all#

Calibration describes how well someone’s stated confidence matches their actual accuracy — for example, whether the things a person calls “high confidence” turn out right more often than the things they call “low confidence.” You cannot measure calibration from a single case, and you cannot measure it at all from an answer given after the outcome is known, because by then confidence and accuracy have collapsed into the same moment. A recorded prior claim with a stated confidence level is the minimum unit that calibration would need — repeated across many cases — to say anything at all. This lesson gives you one such unit, not a calibration score.

10. Keep a short evidence record#

Write four short entries about the basil case:

This follows the same observation / interpretation / limit habit introduced in Week 1, applied here to a claim made before evidence rather than a claim made after it.

11. What a prior state does not settle#

A recorded prior state is real progress: it rules out the specific distortion Fischhoff’s subjects showed, because the claim exists in writing from before the answer was known. But it does not tell us how the revision happened. Did Marlow reason it out alone from the reference? Did a friend point out the stem patches? Did an AI system suggest root rot directly, and Marlow simply confirmed it? The prior-state record separates “before” from “after.” It cannot yet tell us whose reasoning produced the “after.”

Apply the lesson yourself#

Use the Week 9 investigation to record your own prior claim, confidence and revision, and a short evidence record. Work alone or with a partner. Before you finish, explain these distinctions in your own words:

• A prior state versus a recalled claim about what you “already thought.”
• Why a wrong first answer, honestly recorded, is still evidence.
• What a recorded prior claim can, and cannot, tell us about a later, more accurate answer.

If you cannot explain one yet, revisit that section and add an example.

Reading with a purpose#

Empirical finding: Fischhoff, B. (1975). “Hindsight ≠ Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertainty.” Journal of Experimental Psychology: Human Perception and Performance, 1(3), 288–299. Read the paper. Ask: why were Fischhoff’s subjects largely unaware that outcome knowledge had changed their own judgment, even when asked directly to reconstruct it? This reading deepens the lesson; you can complete the activities without it.

What this leaves unresolved#

A prior state, honestly recorded, separates what existed before an answer was known from what came after. It rules out the specific hindsight distortion this lecture describes. It does not tell us how much of a person’s later, better answer came from their own reasoning versus assistance available during the task — another person, a search, or an AI system working alongside them. Next week takes that question on directly, using a collaborator present during the task rather than only a reference checked afterward.