communicoupling

Nodding is cheap.
Now teach it back.

Your student nods. Your colleague says "makes sense." Your teenager says "I know." One of them may understand; the nod will never tell you which. Gordon Pask built a theory of conversation in which the question has an operational answer: understanding of a topic exists exactly when the learner can teach it back — reconstruct the topic's derivation from their own conceptual structure — and word-perfect reproduction is merely memory of yours.

In Pask's conversation theory a concept is a cluster of procedures: the ways a topic can be re-derived from its neighbours in an entailment mesh, a web of this-follows-from-that. Teacher and learner converse on two levels at once — exchanging descriptions (what holds) and derivations (how it hangs together) — and only the second level carries understanding. An echo can be flawless and structurally empty; probing shatters it.

Below, a Teacher's mesh is serialised into messages and pushed through a lossy channel to a Learner who either parrots the sentences or rebuilds the structure. An assessment battery then asks both kinds of question — repeat what was said, and derive something by a fresh route — and the headline number is the gap between those two scores. Press play.

channel fidelity · φchance a message survives intact
100%
teachback rounds · Rteach → teach back → correct cycles
0
mesh size · Ntopics in the teacher's mesh
7
learner mode · μparrot echoes · teachback re-derives
parrot
apparent understanding
recall probes · repeat the lesson
actual understanding
derivation probes · rebuild the route
the gap
recall − derivation
The two meshes · structure vs reconstruction
left: the teacher's true entailment mesh · right: what the learner holds — dotted links mark disconnected sentences
true link correctly held false link missing this step
edge overlap true links held · false links carried
The conversation · messages, teachback, probes
[R] recall probe: repeat what was said · [D] derivation probe: rebuild a route, sometimes avoiding a named topic — every claimed chain is checked link by link against the true mesh
ready
Watch what happens
A perfect channel into a perfect echo: recall lands at 100% while derivation probes collapse to 25% — the few it gets right are the ones where saying nothing happens to be true. The 75pp gap is the whole diagnosis: fluent agreement, no structure behind it.
understanding(T) = re‑derive(T | own mesh) · gap = recall − derivation Pask's criterion: you understand a topic when you can rebuild its derivation inside your own entailment mesh. Recall measures what the channel delivered; derivation measures what the learner built. The first number measures agreement; the second measures understanding.

Two levels of one conversation

A concept is the ways you can rebuild it.

The teacher's mesh is a directed graph: A ⊢ B means topic B can be derived from topic A. Teaching serialises that structure into a lecture — one edge per message — and pushes it through a channel that, with probability 1−φ, drops a message or garbles its endpoint. The two learner modes differ in what they do with an arrival. The parrot files each sentence in order and can replay any of them on demand. The teachback learner discards the packaging and splices the link into a mesh of its own; asked anything, it answers by walking that mesh — from its structure, never from the teacher's serial order.

The assessment battery quizzes both the same way. Recall probes ask for a message back and grade the echo against what was sent. Derivation probes come honestly from the true mesh: multi‑hop questions the lecture never stated — derive F from B, derive D from A avoiding C — with every claimed chain checked link by link, so one false step names the bad link. Teachback rounds close the loop before the exam: the learner produces its mesh, and production is diagnostic — a missing link is a derivation it cannot make, a false link one the teacher watches fail. The discrepancies get re‑sent through the same lossy channel; nothing else in the instrument repairs anything.

What to try

Three moves that expose the gap.

01

Lay the gap bare

Press The nodding student. Apparent understanding climbs to 100%; actual sticks at 25%; the gap reads 75pp. Watch the log: every [D] probe answers "no derivation" while the transcript sits complete on screen. The parrot holds all eight sentences and links none of them.

02

Repair a rotten channel

Set mode to teachback, fidelity 0.55, rounds 0: recall 75%, derivation 58%. Now raise rounds to 3: derivation hits 100% while recall sits at 88% — the gap goes negative. The learner understands a route it cannot recite; its mesh derives around a lost sentence.

03

Localise the lie

Press The misconception. Recall: perfect. Then three [D] probes snap at the same step — A ⊢ D is no link — and the false shortcut glows red on the learner's panel. The battery finds the wrong edge without ever being told where to look.

Being understood vs being agreed with

The nod measures the channel. Teachback measures the mesh.

Every teacher has met the student whose recall is immaculate and whose understanding is vapour; every meeting has ended with a round of "makes sense" that dissolved by the following Tuesday. The nod is a report on the channel — your words arrived — and channels can run clean into a learner who builds nothing. Pask's test cuts past it with one move: ask for the idea back in the other person's terms, by a route you never taught. The colleague who can re‑derive your architecture decision from the constraints, the student who can reach the theorem while avoiding the lemma you used — these are derivation probes, and they cannot be passed from a transcript.

The same instrument runs at home. Couples therapists have partners restate each other's grievance in their own words before replying — teachback, prescribed clinically, because "I heard you" is a recall claim, and the fights that repeat are the ones only ever recalled. Onboarding ends the day the new engineer teaches the system back; a committee that adopts a strategy no member can re‑derive will follow it to the letter and abandon it at the first case the letter fails to cover. The repair loop matters as much as the test: having to produce the idea exposes exactly which links never arrived — why explaining a bug to a colleague finds the bug, and why the fastest way to learn anything is agreeing to teach it.

In the channel cluster

Teachback generates redundancy on demand.

Beside noise and redundancy, teachback is a feedback code: the lecture spends redundancy blindly, repeating everything a little, while a teachback round spends it exactly where production failed — how three rounds beat a channel that eats half of every lesson. From the black box it takes its method: the learner's mesh is unobservable, so the battery reads structure out of responses to probes. And like gesture and response, it locates meaning in the return leg: the lesson means what the learner can give back, and conversation is the only channel that gets to check.

The mapping

Mechanism ↔ social life.

In the modelIn the world
the entailment meshWhat a concept is to someone — its web of "because"s, each topic held up by its neighbours.
serialised messagesThe lecture, the explanation, the strategy deck — structure flattened into a sequence of sentences.
parrot recallThe nod, the fluent echo, "makes sense" — your words returned in your order.
derivation probes"Explain it to me a different way." "Why would it break without this?" Routes never taught.
the gapThe distance between being agreed with and being understood.
teachback roundsConversation doing what transmission cannot: production exposing exactly what never arrived.

Where it tears

Limits.

The instrument assumes the teacher is right.

Grading every chain against a fixed true mesh presumes the teacher's structure is the correct one. Real concepts are contested and open‑textured — classrooms should expect to be corrected by good students, and science depends on it. Pask's full theory lets both participants revise the mesh; a learner whose "wrong" edge is an improvement scores here as a misconception. The simulation models tutoring, and only the agreeable kind.

Teachback is expensive, and institutions know it.

One learner at a time, minutes per topic, a teacher attending closely enough to spot which link failed. Recall testing is cheap, parallel, and gradable by machine — which is why schools, certifications, and onboarding checklists default to it while knowing it measures the wrong quantity. At scale the gap is a budgeting decision: institutions buy apparent understanding because the actual kind does not fit in an exam hall.

Understanding‑as‑derivability is one operationalisation.

Knowing how to ride a bicycle, recognising a face, judging when a sentence is finished — little of this survives translation into nodes and derivation links, and Pask's machinery covers less of the mind than its generality suggests. The mesh picture fits knowledge that argues: claims held up by reasons. Where know‑how, perception, and taste live, teach‑it‑back may test verbal fluency about a skill rather than the skill itself.