Other fields
It doesn't pull in expertise from other fields unprompted.
File 04 · Diverge
When everyone runs the same agents, the edge is position: how much opportunity space you search, and how close to reality you land.
The common view says compute, proprietary data and a willingness to drop guardrails decide who wins once intelligence is cheap. It misses a fourth.
It's curation, not probability
A model searches only the opportunity space its prompt and context open up. By default it does the least work that produces an acceptable answer.
A thousand agents running that way don't widen the search. They repeat it. Every answer lands on the same floor.
Widening starts with what the default never pulls in unprompted: expertise from other fields.
Then the research beyond what's required, past the point where an acceptable answer already exists.
Then solutions far outside the question's frame, where nothing in the prompt says to look.
The edge is expanding the opportunity space on purpose: forcing the outside expertise, the extra research and the far-off frames the default never reaches for.
What the default skips
By default a model does the least work that produces an acceptable answer. Three things it doesn't do unprompted. Force them.
It doesn't pull in expertise from other fields unprompted.
It doesn't do research beyond what's required.
It doesn't look for solutions far outside the question's frame.
LEFT SHUTThe least work that passes. That's consensus curation.
Compute and divergence
Compute matters most for well-defined, deterministic work, where it narrows toward the answer faster. Divergence spends compute too, but on widening the opportunity space instead of compressing it.
The two are complements, not rivals.
Folded in
Same author, framework or training data is correlation, not confirmation. Past a point, more agreement should lower confidence.
For when both sides have reason to bend the record. Training data is consensus residue, not evidence. The result is a data point, never a verdict. This corpus is what lets a moonshot stand up.
Every working exploit encodes a truth about how a system really runs under its official story.
Why now
As model output floods the world and trains the next models, consensus feeds on itself and drifts further from the ground truth. The reality delta grows.
The labs approach the agent problem from inside their own incentives. Infrastructure built on your own methodology means any agent system you run follows rules you actually agree with.
What it allows
In a saturated market, standing out at all has value. Standing out on something real compounds.
FOLDED IN · EACH ONE BREAKS WHEN