Go wide, then go deep.

01 · the trouble with first ideas

When people are asked to invent something new, they do not start from nothing. Thomas Ward showed this by asking people to draw animals from another planet: nearly everyone drew things with legs, eyes, and bilateral symmetry.¹ Imagination is structured by the examples we already hold, and it stays close to them. Worse, showing people an example locks them in further: designers shown a flawed sample design reproduced its flaws even when explicitly told to avoid them.² The literature calls this fixation, and it is not a character defect. It is how memory works.

Language models inherit a version of the same tendency for a related reason: they are trained to produce what is probable, and the probable answer to a familiar problem is the familiar solution. Ask Fable 5 to “brainstorm” without structure and it will give you fluent, sensible, central ideas, which is precisely what you did not need. The fix is not a better adjective in the prompt. The fix is running exploration the way the research says exploration actually works.

02 · what the research says

Three findings carry this note. First, variance plus selection beats refinement. Girotra, Terwiesch, and Ulrich measured it directly: the quality of a team’s best idea is driven by the spread of ideas generated and the ability to pick well, and individuals generating independently before any group discussion produced better best-ideas than groups working together from the start.³ Dean Simonton’s equal-odds rule points the same way across creative careers: the hit rate stays roughly constant, so the people with the most masterpieces are the people with the most attempts.⁴

Second, distance unlocks. Gick and Holyoak’s famous radiation experiments found that only about 10 percent of subjects solved a hard medical problem cold, but roughly three quarters solved it once they were nudged to use an analogous military story they had just read.⁵ The solution existed in another domain; the work was the transfer. Third, exploration and exploitation compete for the same budget. James March’s classic result is that organizations systematically over-exploit what they know and under-explore what they don’t, because exploitation pays sooner and more predictably.⁶ Your instinct to grab the first workable answer is March’s finding operating inside one conversation.

03 · the exploration protocol

Here is how I run it with Fable 5 on any problem that deserves more than the obvious answer. It follows the shape of the research: wide first, distant on purpose, adversarial in the middle, deep at the end, and no judging until the map exists.

Go wide with forced distance. I do not ask for ten ideas; I ask for five directions that are mutually exclusive at the level of the underlying thesis, and then I make the model check its own spread: for each pair, say what one direction believes about the world that another denies. If two directions survive that check as siblings, they merge, and the model owes me a replacement. This single move kills the ten-variations-of-one-idea problem, because distance is now the specification rather than a hope.

Import from far domains. The Gick and Holyoak result becomes a prompt: which three industries solved a structurally identical problem under harsher constraints, and what would their solution look like here? A distributor’s routing problem has been solved by blood banks, by airline crew schedulers, and by concert tour logistics, and the transfer is where original-looking strategy usually comes from. Originality, examined closely, is mostly excellent importing.

Probe the extremes. Solve it with ten times the budget, then with none. Solve it as a monopolist, then as the smallest player in the market. Corner solutions are rarely buildable, but they reveal which constraints were real and which were furniture, and the best middle answer is usually a corner answer wearing sensible clothes. Then run the adversary pass from Field Note 03: for the leading directions, let every player on the map make their best response, and watch which directions die.

Finally, go deep on the strangest survivor. This is the most underused move in working with these models. Everyone samples; almost nobody deepens. Take the direction that feels least comfortable and push three levels: what would this require, what breaks first, what does it look like after the break is fixed? Depth is where a strange idea either becomes a real one or teaches you why it can’t, and either result is worth more than another lap of shallow sampling. Only after the deepening do the selection criteria enter, and they enter explicitly: expected profit, added value, and what the design survives.

problem: a $40M distributor, margins compressed 3 points in 2 yrs

wide     "Give me five directions that are mutually exclusive at
          the thesis level. For each pair, state what one believes
          that the other denies. Merge any siblings and replace them."
          → price architecture rebuild · unbundle service & charge
            for it · sell their operating data as a product ·
            route-density partnership with a competitor · exit the
            low-density half of the territory

import   "Which three industries solved margin compression under
          harsher constraints, and what does their move look like
          here?"  → airlines (yield mgmt), blood banks (pooled
          logistics), tour promoters (route economics)

extremes "Solve it as a monopolist. Now as the smallest player.
          What constraint turned out to be furniture?"

deepen   "Direction three feels wrong. Go three levels deep:
          what would it require, what breaks first, and what does
          it look like after the break is fixed?"

select   only now: expected profit · added value · survives
          every player's best response  (field notes 01 + 03)

04 · what it returns

The honest answer is that most of what exploration produces gets thrown away, and that is the point: you are paying a few hours of model time to buy variance, because variance is where the best idea lives. In the worked example above, the direction that survived was the strange one, the data product, and it survived because the deepening pass turned “sell the data” into something specific enough to test with the added-value math from Field Note 03. That is also not a coincidence of the example: in the Foundry, this protocol is how upper-right findings get onto the opportunity matrix at all. The profitable-and-common quadrant is almost never where the first idea lands, because the first idea is, by construction, the one everyone in the industry already had.

There is a compounding effect worth naming. Exploration transcripts are assets. The four directions that died carry priced reasoning about why they died, and half of the next engagement’s map is usually already drawn in the last engagement’s discards. An agent that keeps and indexes its own dead branches explores faster every quarter, which is a small, quiet moat that costs nothing to build except the discipline of not deleting things.

05 · the honest limits

Exploration is a budget, not a religion. March’s result cuts both ways: over-exploration is a real failure mode, and a problem whose answer is already known deserves execution, not a five-direction search, so the protocol runs on decisions where being wrong is expensive and the obvious answer has already disappointed. Novelty is not value: a strange idea that fails the added-value test is just strange, which is why selection criteria from Field Notes 01 and 03 are part of the protocol rather than an afterthought. And the variance math assumes real independence between directions, which is exactly what the distance check exists to enforce; skip it, and the slider in figure 1 is lying to you, because ten correlated ideas are one idea with ten haircuts. The discipline, as everywhere in this series, is what converts a trick into a system.