communicoupling

The message you cannot obey,
refuse, or mention.

"Be spontaneous." "You should want to visit your mother." "We reward candour here" — said in a room where candour has ended careers. Some communications are traps with three walls: doing what they say loses, refusing loses, and pointing at the trap is the move punished hardest of all.

Gregory Bateson, Don Jackson, Jay Haley and John Weakland named the structure in 1956: the double bind. A double bind is an intense relationship the victim cannot leave, a demand at one logical level, a second demand at a higher level that contradicts the first, and a block on commenting about the contradiction. The words say do X or be punished; the tone above the words says this is not coercion — you must do X freely; and a third rule, usually unspoken, says you may not call this what it is. Every direct reply is punishable, and the one winning move — talking about the frame instead of inside it — is barred. The layering comes from Russell and Whitehead's theory of logical types: the trap is built out of levels.

Bateson's darker claim was about learning: live inside such traps long enough and you learn that messages themselves cannot be trusted — a trained incoherence in handling levels. Below, a reinforcement learner lives inside the bind. It can comply, refuse, withdraw, distort, name the frame, or leave; a payoff engine runs the trap, and the agent becomes whatever the trap teaches.

bind intensity · Isize of the sanction when the trap fires
1.8
meta block · Bthe wall against naming the frame
1.00
exit cost · Xprice of leaving the relationship
1.60
inconsistency · qchance an episode is warm, not a trap
0%
response entropy · bits
the learned incoherence, live
punishment rate
mean sanction per episode (EMA)
dominant strategy
largest share of the policy
The learned distribution · who the agent becomes
each band is the live probability of a reply — with the wall up, mass drains out of comply, refuse and naming, into withdrawal and distortion
τ 0.25 · α 0.06 · ε 0.05 0 episodes
The running bill · cumulative punishment
slope is suffering per episode — adaptation bends it flatter; dropping the wall bends it flat
Watch what happens
The classic bind: sanction 1.8, wall up, exit walled, the trap firing every episode. Comply, refuse and naming all bleed, so within a few thousand episodes the policy is mostly distortion and withdrawal — symptom-like strategies discovered as optimal play.
bind = punish(comply) ∧ punish(refuse) ∧ block(meta) ∧ wall(exit) Bateson's four conditions in payoff form. The agent learns a softmax policy π(a) ∝ eQ(a)/τ over six replies (τ = 0.25); when every straight reply is sanctioned and the frame may not be named, what remains of the distribution is the symptom.

The anatomy of the trap

Four conditions, and the fourth is the lock.

The payoff engine implements the bind exactly. When the trap fires, comply is punished at the meta level — you didn't mean it — and refuse is punished at the object level, both by the sanction I. Naming the frame is heavily punished while the wall B is up, and rewarded the moment it falls. Leaving costs X every episode. Withdrawing and distorting carry small ambient costs nobody designed. The agent knows none of this: it holds one learned value per reply and samples from a softmax over them, temperature on display in the run bar. Every band, curve and readout is computed from that live policy.

Notice what makes a bind more than a contradiction. Contradictory demands are survivable wherever you can say you're contradicting yourself — that sentence, spoken one level up, dissolves the trap. The secondary injunction rarely arrives in words; it rides on tone, posture and history, which is why the trap is so hard to point to. And the tertiary injunction forbids pointing at all. Contradiction plus enforced silence about the contradiction plus no exit: that is the complete machine, and each dial above is one of its walls.

What to try

Three ways to spring the trap.

01

Free the word

Run The classic bind until distortion and withdrawal own the chart. Then, without restarting, drag meta block → 0.00. A beat of hesitation — naming was learned to be dangerous, and the agent barely tries it — then a few lucky trials repay the risk and the green band takes the whole picture. The bill goes flat.

02

Open the door

Keep the wall at 1.00 but drop exit cost → 0.05. The agent never outwits the trap; the purple leave band simply swallows the policy. A bind with a cheap exit is an inconvenience — binds need walls on every side, including the door.

03

Make it intermittent

Set inconsistency → 50%. Warm episodes reward compliance just often enough to keep it in the race, so the blue band keeps surging back and entropy hangs near 2 bits. Convergence crawls: a trap that fires reliably is easier to learn than one wrapped in kindness.

Symptoms as solutions

What chronic binds teach.

The model's key trace is where the probability goes. With every straight reply sanctioned and the winning move barred, mass drains into withdrawal and distortion — the strategies that look, from outside, like symptoms. The employee who turns vague in exactly one meeting; the teenager who answers in shrugs; the parishioner who goes quiet rather than agree or object: each has discovered the least-punished cell of a payoff matrix someone else built. Bateson read certain psychotic speech the same way — as the perfected strategy of a person for whom no level of the message was ever safe. The entropy readout makes his "learned incoherence" a number: the policy of someone who can no longer afford a definite reply.

Two of the presets carry the theory's hopeful edges. Therapy is the wall coming down: the Palo Alto tradition that grew from this paper treated the family, and its rules about what may be said about the rules, as the patient — and in the model, the moment metacommunication becomes learnable the distribution snaps to naming the frame. Watch the hesitation first, though: learned distrust outlives the trap that taught it, which is why permission to speak must be demonstrated, and demonstrated again, before it is believed. The humorist is the same move at low stakes — Bateson insisted that humour lives at the very level where pathology does, playing with frames rather than suffocating in them. Distortion as art instead of injury.

Neighbouring concepts

Level trouble runs in the family.

The bind is Bateson's grammar of levels at its sharpest, and its neighbours share the machinery. Punctuation is a fight about frames too — who started it is a disagreement one level above the blows. Schismogenesis shows what happens when a relational pattern runs away without anyone able to step outside it — the bind adds the rule that stepping outside is itself forbidden. And laws of form gives the clean logic of the levels the bind so productively confuses.

The mapping

Mechanism ↔ social life.

In the modelIn the world
primary injunctionThe stated demand — "visit your mother", "be candid", "be spontaneous".
secondary injunctionThe tone that unsays it: you must also want this, freely, or it doesn't count.
meta block · B"Don't you dare call this what it is" — the rule against discussing the rules.
withdraw / distortSymptoms as solutions: vagueness, shrugs, absence — the least-punished replies.
name the frameMetacommunication — therapy's move, and the only reply that wins when permitted.
exit cost · XWhy binds live in families, jobs and churches rather than among strangers.

Where it tears

Limits.

As an explanation of schizophrenia, it failed.

The 1956 paper proposed the double bind as an etiology of schizophrenia, and the empirical record did not cooperate: coding studies could not reliably find more binds in the families of diagnosed patients, and the causal theory faded from psychiatry. What survives is the grammar — a precise anatomy of pathological communication that illuminates families, workplaces and institutions whether or not it ever produced a psychosis.

A softmax learner is a cartoon of a person.

The agent can only reweight six replies. A real victim theorises the trap — builds a model of the binder, of themselves, of why this keeps happening — and the theorising changes the trap: sometimes loosening it, sometimes tightening it, since self-blame is also a theory. Nothing in the instrument captures that reflexive loop; treat the drained distribution as the skeleton of the experience, not its flesh.

Detecting real binds is interpretive work.

In the model the secondary injunction is a payoff term you can read off a dial. In life it lives in tone and context — exactly where observers disagree. One person's double bind is another's ordinary ambivalence, and the same sentence can bind one listener and amuse the next. That slipperiness is why the coding studies faltered, and it is a property of the phenomenon, not a flaw in the observers.

Which move breaks this loop?

Three replies from one level up.

The moveWhere it cuts
Self-distancingSteps outside the injunction's level, where both commands are visible at once — the oscillation between them needs a viewer trapped inside.
Surfacing the assumptionsSays the unspoken rule aloud — the tertiary injunction holds only while it stays unstated.
Making the system an objectComments on the bind itself rather than replying inside it — the model's name the frame, the one reply that wins wherever the wall permits it.