Week 3: Reconstruction
Study a supplied explanation of a pin tumbler lock, close it, rebuild the mechanism from memory, then restore the source and read the annotated difference as evidence.
8 March 2027
What becomes visible when the answer is no longer in front of us?
This is a self-guided lesson. These notes and the slide deck each contain everything needed; no presenter is assumed. Allow about 40 minutes, plus 25 more for the Week 3 investigation. You need something to write with and nothing else: no lock, no tools, no paid book and no AI account. Priya and the written answers below are fictional teaching examples.
What you will learn#
By the end you should be able to distinguish consumption from reconstruction and recall from reconstruction; to run a source removal on yourself and annotate what came out using a four-part scheme; to read that annotation by the type of loss rather than its volume; and to state precisely what a strong reconstruction does and does not establish. You should also be able to say why an unwitnessed, self-administered removal limits the claim you can make from it.
1. What Week 1 settled, in a sentence#
Week 1 established the examiner — the person deciding what a performance shows — and separated what you observed from what you inferred, insisted that the conditions an answer was produced under belong in the record, and used Clark and Chalmers to argue that working with a reference does not automatically disqualify a performance. It also warned that taking a source away adds a demand on memory, so you can end up testing something other than what you intended. Week 2 found that a fluent explanation can still conceal missing parts and causes.
This week does not re-argue any of that. Removal has a cost, and it has to be chosen deliberately; that is settled. The question here is what the removal makes visible and how an examiner reads the result. If the source is gone and something is missing from what a person produces, that absence is evidence. Our subject is what kind of evidence it is.
2. Consumption and reconstruction#
Imagine Priya opening a short explanation of how a door lock works. It makes sense as she goes; each sentence follows from the last; by the end she would say she understands it. Call that consumption: taking in an explanation while it is in front of you, following it, and finding it coherent.
Consumption is not a lesser activity, and this lesson is not an argument against it. You cannot rebuild a structure you have never met, so consumption is where the material comes from.
Now Priya puts the explanation away and writes the mechanism out with nothing in front of her. Reconstruction, as this course uses the word, means rebuilding the structure of an explanation without the source in view: naming the parts, saying where each sits relative to the others, and giving the causal steps that make the thing work. It is not reproducing the wording. A reconstruction in completely different words, with the structure intact, is a success.
Recall and reconstruction are related but not interchangeable. Recall returns remembered material — a term, a sentence, a list — and succeeds when the material comes back. Reconstruction succeeds when the parts are present, placed correctly, and joined by steps that carry the mechanism. A reconstruction usually needs recall to supply its pieces, which is one reason a poor reconstruction is ambiguous evidence: it may indicate a missing structure or a missing word.
3. Why a pin tumbler lock#
A reconstruction can only be diagnosed if the thing reconstructed has nameable parts. Then a gap is a specific missing item rather than a vague sense that an answer was thin. A pin tumbler lock — the ordinary key lock on a front door — has eight nameable parts, one named boundary and three states, and needs no prior knowledge. You will not learn to open one; we are examining an explanation, not operating a mechanism.
One warning before the reference. Rozenblit and Keil used a cylinder lock as a test item and asked participants to explain step by step how one could pick one. As in Week 2, this was a diagnostic question, a probe for a specific gap; here it informed self-assessment: their interest was in what people discovered about their own knowledge, not in the technique. This lesson does not cover picking a lock and nothing here should be read as instruction in it.
4. The reference#
Read this once or twice, then close it. Everything after it assumes you are not looking at it.
An outer casing has a cylindrical hole in it. Inside that hole sits a rotating cylinder called the plug. To unlock the door, the plug has to turn. The front face of the plug carries a straight slot, the keyway, which the key enters. At the back of the plug is a cam: a projecting piece that, as the plug turns, draws back the bolt, the metal bar holding the door shut.
Five or six holes are drilled down through the plug. Each holds a key pin, cut to its own length and rounded underneath so a key can slide across it. Directly above each key pin, in a shaft in the casing, sits a driver pin, with a spring above it pressing down. A key pin and the driver pin above it make up one stack. The boundary where the plug meets the casing is the shear line: the surface along which the plug has to rotate.
With no key, the springs push each stack down, so every driver pin is forced part-way out of the casing and into the plug. Pins cross the boundary, and the plug cannot turn. With the correct key, its cut edge lifts each key pin by a different, exact amount, so every key-pin/driver-pin junction comes level with the shear line. Nothing crosses the boundary, the plug turns, and the cam draws the bolt back. With a wrong key, the cuts lift the stacks by the wrong amounts: lifted too little, a driver pin still crosses into the plug; lifted too far, a key pin is pushed up into the casing. One wrong stack is enough to hold the plug.
Technical grounding: Wikipedia’s article on the pin tumbler lock. That article names the boundary the shear point; “shear line,” used here, appears in the same article and means the same surface. The article grounds the mechanism and is not evidence for any claim about learning or cognition.
5. Pause and reconstruct · 8 minutes#
Stop reading here and close this page, or cover section 4. Write out how a pin tumbler lock works, from memory, in your own words. Name the parts; say where each sits relative to the others; explain why the plug cannot turn with no key, why it turns with the correct key, and what a wrong key does. Permitted: your memory and something to write with, nothing else. Your output is one paragraph or a labelled list, which you keep. The purpose is to produce the artefact the rest of this lesson examines.
The honest problem with that instruction#
You can scroll back up. Nothing prevents it, and nobody is watching you. A source removal you administer to yourself is not something another person can check.
Follow the instruction anyway, for a practical reason rather than a moral one: your reconstruction is the raw material for the next step, and a copied version gives that step nothing to work on. You would lose the lesson rather than fail a test. And if you did look back, write that down. “Reconstructed with one look at the reference” is a usable record; an unrecorded look is what ruins it.
Treat that unverifiability as material to study, not a flaw to paper over. Examining yourself means holding both roles at once: you are the person producing the performance and the examiner judging it, and you cannot independently confirm your own account of the conditions. That does not make self-examination worthless. It makes the condition — self-administered, unwitnessed — part of the record, exactly as Week 1 required. State it, and claim less than you otherwise would.
6. Restore the source and annotate · 5 minutes#
Read section 4 beside your own version and mark every difference on your reconstruction with one of four letters.
| Letter | What it marks |
|---|---|
| P | A part omitted: a named component missing altogether. |
| R | A relation omitted: parts named, but not placed with respect to one another. |
| C | A causal step omitted: a “because” the mechanism needs. |
| X | A scenario omitted: a case the reference covered, such as the wrong key. |
Your output is the annotated reconstruction with a count for each letter. The scheme exists so that “something was lost” becomes a specific, arguable finding.
A fuller reconstruction, and its annotation
A written model answer, not a record of anybody’s work:
“An outer casing holds a rotating plug. The key goes into the keyway at the front of the plug, and a cam at the back pulls the bolt. Six holes run down the plug, each holding a key pin cut to its own length. Above each key pin, in the casing, sits a driver pin with a spring pressing down. Plug and casing meet at the shear line. With no key, the springs push the stacks down, driver pins cross into the plug, and it jams. A correct key raises every junction level with the shear line, so nothing crosses it, the plug turns, and the cam draws the bolt. A wrong key lifts a stack too little, leaving a driver pin in the plug, or too far, pushing a key pin up into the casing.”
This is strong. Every part appears, the stacking relation is right, and all three states are explained with the boundary doing the causal work. It still loses something: C — the reference said the correct key lifts each pin by a different, exact amount, because the pins differ in length. This version says a correct key raises every junction level with the shear line without saying that differing cut depths on the key are what achieve it.
One C against an otherwise complete structure is a small and specific finding, which is exactly what a scheme is for.
Priya’s reconstruction, and its annotation
“A door lock has a barrel that turns when you put the right key in. Inside there are pins of different heights. The key pushes the pins up to the right heights, and then the barrel can turn and the door opens. A wrong key does not push them to the right heights, so it will not turn.”
Notice first that nothing in it is false. Every sentence is consistent with the reference.
P — four parts are gone: the springs; the driver pins, since Priya has one set of pins where the reference had two; the cam; and the bolt.
R — nothing is placed. There is a barrel, and pins are “inside,” but there is no casing and nothing about driver pins sitting above key pins.
C — the largest loss. “Right heights, so it turns” is a rule, not a mechanism. The shear line is gone, so nothing explains why correct heights permit rotation.
X — the wrong key appears, but only as under-lifting. Over-lifting is absent.
7. Reading the annotated difference as evidence#
Count the letters, then weigh them. Priya’s four P items and her one C are not equivalent findings.
A missing part name is often a vocabulary gap. Priya may hold the structure and lack the phrase “driver pin”; a follow-up in which she describes the same arrangement in her own words would tell us. A missing causal step is a different matter. Without the shear line there is no account of why the plug turns, so “right heights, so it turns” could have been produced by a reader who never took in the boundary at all. A lost noun and a lost “because” are not the same loss. Priya can say what happens; she cannot yet say what makes it happen.
A missing X is the least conclusive of the four, because the reader may have judged the wrong-key case not worth writing down rather than not having it. Where that matters, ask for the scenario directly instead of inferring from its absence.
8. Take a strong reconstruction seriously#
Allow three minutes. Suppose a reader produces something as complete as the model answer, with no letters to mark. Before reading on, write two things that result would establish and two things it would not; “I might be wrong” is not one of them, because it names no missing evidence.
It establishes that at that moment, with the reference out of view, this reader produced a structure containing those parts, those relations, those causal steps and the wrong-key condition. That is a real performance, and it is considerably more than handing the reference’s own sentences back. It is also more inspectable than a rating of confidence: there is a written artefact, and we can point at the part of it we are discussing.
It does not establish four things:
- Not that the structure was built rather than memorised. A reader who learned the reference by heart yesterday hands in the same page.
- Not that the reference was actually closed, because nobody witnessed it.
- Not that the same thing would happen for another mechanism, or for one not studied minutes earlier.
- Not that the reader can use the structure on a question the reference did not answer — say, what a lock with one broken spring would do.
That is the week’s limitation in full: reconstruction is stronger evidence, and it can still be memorised, rehearsed or assisted, and in a self-guided setting nobody can confirm the source was closed.
9. Reconstruction is an instrument, not a virtue#
Reconstruction is stronger evidence than a fluent read-through for one stated reason: the source is not supplying the structure. That is the whole claim. It does not make reconstruction morally better than reading, and it does not make consumption a bad habit. Priya could not have reconstructed anything she had never read. Consumption is the scaffolding; reconstruction is a measuring instrument built on top of it.
This matters for the course as much as for the reader. A lesson that treats reconstruction as self-improvement has changed the subject from examining understanding to recommending a study habit, and the recommendation would also be false, because the instrument depends on the consumption it is supposed to replace.
10. What the sources support, and what they do not#
Rozenblit and Keil asked sixteen graduate students from various departments at Yale to rate how well they understood each item on a list of 48 devices and phenomena, using a 7-point scale on which 1 was the lowest level of understanding and 7 the highest. Each participant then wrote a detailed, step-by-step causal explanation of four test items — the set including a cylinder lock also included a helicopter, a quartz watch and a sewing machine — and re-rated themselves. Ratings dropped. A repeated-measures ANOVA across their five rating points gave F(4, 56) = 16.195, p < .001, and the planned contrasts between the first rating and each of the next three were significant at p < .002. Ratings rose again at the final point, after participants had read an expert description, so they were not simply losing confidence in everything.
Two things follow for us, and one does not. Being asked to produce the explanation, rather than being handed one, is what moved the ratings; and the authors’ own reading of the unchanged rating after the expert text is that the diagnostic questions were not unfair, since participants still judged their lower estimates accurate once they saw what an expert thought was essential. What does not follow is any claim about what those participants could produce. Their measurement is, as Week 2 established, a self-rating: a number a participant gives for their own understanding. A drop in it shows that participants revised their opinion of themselves.
Our annotated reconstruction is a step better, because the evidence is a page we can inspect rather than a number. That improvement is our adaptation of their design, not their finding. Running it once on yourself, with no control and no comparison, replicates nothing.
The annotation step also exists because people differ widely in noticing that they have not understood something. Chi, analysing students explaining physics worked examples aloud, reports that poor learners acknowledged a comprehension failure about 15% of the time while good learners did so about 46% of the time, and that once a failure was detected it was usually followed by an attempt to explain. Her students had the text in front of them, which is a different condition from ours, so we take the point about noticing rather than the result about self-explaining. Priya, mid-reconstruction, may simply not feel a gap where there is one. Restoring the source removes the need to notice: the reference finds the gap for you.
11. Keep a short record#
Write five lines and keep them with your annotated page.
| Field | What to write |
|---|---|
| Condition | “Self-administered and unwitnessed; reference closed,” or “one look back.” |
| Observation | “Without the reference, I produced…” |
| Counts | How many P, R, C and X. |
| Interpretation | “This supports the claim that I can…” |
| Limit | “It does not show that I can…” |
A useful limit names the evidence you are missing. “Explain a mechanism I had not just read” is one; “I don’t really understand locks” is not.
An examiner’s record of Priya might read: Condition — self-administered, unwitnessed, reference closed on her own account. Observation — she named a rotating barrel and pins of differing heights, and stated that correct heights permit rotation. Counts — P 4, R 2, C 1, X 1. Interpretation — she retained the outcome of the mechanism and one of its components. Limit — the single C is the finding that matters; nothing here shows she holds the boundary that makes the outcome follow.
Reading with a purpose#
Primary evidence: Rozenblit, L., and Keil, F. (2002). “The misunderstood limits of folk science: an illusion of explanatory depth.” Cognitive Science, 26(5), 521–562. Full text from the authors’ lab. Read Study 1 and ask: of their five ratings, which one actually takes the source away? This is an empirical finding, demonstrated across twelve studies in the one paper, and Study 1 rests on sixteen graduate students at one university.
A complementary argument: Chi, M. T. H. (1998). “Self-explaining: the dual processes of generating inferences and repairing mental models.” Online manuscript prepared for R. Glaser (ed.), Advances in Instructional Psychology (Lawrence Erlbaum), dated 15 September 1998. Manuscript PDF. Read “What self-explaining is not” and the section on self-repair, and ask how much of a reconstruction is repair rather than retrieval. It supplies both empirical findings and a theoretical argument; note that its own title page describes it as in press, so the published chapter’s title and pagination may differ.
Carried from Week 1: Clark, A., and Chalmers, D. (1998). “The Extended Mind.” Analysis, 58(1), 7–19. Read the paper. Week 1 asked when a resource belongs inside the activity. Ask something different this week: if a resource can genuinely participate in thinking, what exactly are we measuring when we take it away? This is a philosophical argument, not evidence about reconstruction.
Technical grounding for the lock only: Wikipedia, “Pin tumbler lock”.
What this leaves unresolved#
One task cannot settle the question. Priya’s page could be thin because she does not hold the mechanism, or because she writes tersely, or because eight minutes was short, or because a written paragraph is not the form in which she thinks. A single reconstruction cannot distinguish those, and the unwitnessed condition leaves the strongest result open to a simple alternative explanation.
So the next move is to compare the same person’s performances across conditions rather than trust one of them. Week 4 asks why the same person looks more or less knowledgeable under different conditions, and reuses this lock under two or three of them. Apply this week’s lesson first in the Week 3 investigation, which takes the same protocol further and asks you to defend one judgement from what you actually produced.