Lateral Thinking
TL;DR
Lateral thinking — named by Edward de Bono in 1967 as the deliberate escape from the first pattern — moves sideways out of the dominant line of reasoning rather than deeper into it: provocation, reversal, challenging the assumption, random entry. In an IRG, it is best understood as an engineered momentum break. An unguided model’s options are all variations of its first idea, because next-token generation continues the pattern it started; a lateral node is required to produce an option that is structurally different rather than differently worded, and a divergence gate rejects a candidate set whose members share the assumption under challenge. Provocation is a means, not an output — a feasibility gate holds anything that leaves the graph to the ordinary rules. Reach for it when every option on the table is the same option in different words.
Four options, one idea
The options paper on the table has four entries and a recommendation, prepared after a quarter of rising losses on a payment channel. Lower the auto-approve threshold. Add a second reviewer above a dollar line. Extend the hold on first-time payees. Introduce a step-up challenge at submission. By the ordinary standards of the work these are four choices — different costs, different customer friction, different staffing implications — and the committee reads them that way for most of an hour. Then someone says the quiet thing: all four buy the same protection with the same currency. Look harder, and later, at a payment that has already been built.
The assumption underneath them is written nowhere, which is exactly why nobody has argued with it — that a payment’s risk is fixed by the time it arrives, and the only remaining lever is the intensity of inspection. Violating it on purpose (suppose the risk is not fixed; suppose the institution chooses when the payment arrives) opens a different region of the option space: proving a destination account at the moment a payee is added rather than at the moment money moves; a settlement-timing choice that makes the safe path the convenient one. Most of what comes out of a move like that will fail on cost, on regulation, or on operational reality, and it is meant to. What matters is that none of it was reachable by continuing down the four options already on the page.
An unguided model asked for more options returns more of the option it has already produced — next-token generation continues the pattern it started, so the fifth idea is usually the first idea rephrased. The lateral graph installs a momentum break: a node must name the assumption the set rests on, and a gate refuses a candidate set whose members all still share it.
How the graph works, step by step
Generate the first-pattern option. The graph begins by producing the obvious answer deliberately, and keeps it. This is not a warm-up. The dominant pattern has to exist as a named artifact before anything can be measured against it, and in most regulated settings the obvious answer is also a serious candidate that deserves to stay on the list. What the step buys is a reference point: from here on, “structurally different” means different from this, and the claim is checkable.
Name the load-bearing assumption. The critique node’s task is narrow and unusually hard — find the premise the first option depends on, not the premises it merely mentions. A serviceable test: negate the candidate assumption and ask whether the option still stands. “Reviewer time is expensive” is a constraint the option lives with; “risk is fixed at arrival and inspection is the only lever” is load-bearing, because the option collapses without it. The named assumption goes into the trace, which is what makes the rest of the run challengeable by someone who thinks a different premise was the real one.
Generate against the assumption. Generation is now constrained in the opposite of the usual direction: produce an option inconsistent with the named assumption. De Bono’s devices are the mechanics — reversal (run the sequence the other way), provocation (assume the thing that cannot be true and follow it a few steps), random entry (introduce an unrelated element and force a connection). The divergence gate then checks the result structurally rather than lexically: does this candidate actually violate the assumption, or is it the first option wearing new vocabulary?
Evaluate feasibility, and gate. Everything that survives divergence goes through ordinary evaluation — legal, regulatory, operational, cost. The feasibility gate is deliberately unforgiving, because provocation is a means and not an output: an option leaves this graph only by meeting the standard a conventionally generated option would have to meet. Most candidates die here, by design, and the trace keeps them — the discarded set is the evidence that the option space was actually searched.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Assumption gate | The named premise is load-bearing — negating it would collapse the first-pattern option, not merely inconvenience it | The node names a preference, a restated constraint, or a decoration nothing rested on — challenging something that was never holding weight |
| Divergence gate | At least one candidate is inconsistent with the named assumption, and the set’s members differ in mechanism rather than in wording or magnitude | Every member still shares the assumption under challenge — paraphrase presented as a set of alternatives |
| Feasibility gate | Each surviving candidate carries a full ordinary-rules assessment: legal, regulatory, operational, cost | A provocation is treated as a proposal, or a lateral option ships with less scrutiny than a conventional one would have drawn |
Where it fits — three use cases
1. Remediation options where every proposal is “more review”
Second-line remediation sets converge on added inspection with striking reliability: another approval tier, a wider sample, a longer hold. The lateral node forces the underlying premise into the open — usually that the defect is detectable only after the fact — and asks for one option that does not depend on it. The surviving candidate is often upstream of where anyone was looking, and it arrives with the rejected alternatives attached, which is what a reviewer needs to judge whether the search was real.
2. Product or process design under a constraint nobody questions anymore
Long-lived processes accumulate constraints whose original reason has expired: a sequencing rule inherited from a retired system, a document required by a policy that has since been rewritten, a handoff that exists because two teams once reported separately. Because these read as facts rather than as choices, they never enter the option set as variables. A node whose job is to name and violate one makes the constraint visible, and the feasibility gate immediately says whether it is still real.
3. Adversarial and red-team scenario generation
When generated attack scenarios all rhyme — the same channel, the same stolen-credential opening, the same escalation path — coverage is narrower than the count suggests, and that narrowness is invisible in a list of twenty. Naming the shared assumption (the adversary is external, the adversary is in a hurry, the adversary wants money now) and requiring a scenario that violates it widens the tested surface in a way the trace can show, which is exactly what a model-risk reviewer will ask about.
When to reach for it
Reach for lateral thinking when the option set has stopped moving — when successive proposals differ in magnitude rather than in kind, and the discussion has become about how much rather than about what. It sits naturally inside the first half of a Double Diamond, where divergence is the phase’s whole job, and beside dialectic, which works the same seam from the other side: dialectic opposes a position, while lateral thinking simply leaves it.
Its limits are the strongest thing to be honest about. Lateral output is raw and mostly discarded by design; the strategy’s value is not its hit rate but whether the one surviving option was reachable any other way — a question worth asking of every run, and answerable from the trace. It is also a poor last step. A provocation is not a proposal, and a lateral node should never be the node immediately before an action; in regulated work it belongs upstream, generating candidates that a conventional evaluation then has to accept or kill on the ordinary terms. Used that way it reduces and interrupts a specific failure — the option set that only looks like a set. It does not make the surviving option right, and the trace is there so that a reviewer can say so.
- Momentum break
- A node placed to stop generation from continuing its own pattern — the structural answer to a model whose fifth option is its first option rephrased.
- Load-bearing assumption
- The premise an option actually depends on, identified by negating it and asking whether the option still stands; the only assumption worth challenging.
- Provocation, not proposal
- Lateral output is an input to evaluation rather than a candidate for action — the feasibility gate holds it to the same standard any conventional option would face.