Systems

Random Imagination Engine

What might we be missing entirely?

Most research starts from a company and asks whether it is worth owning. This engine starts from nothing and generates ideas deliberately outside the existing record. Almost all of them are worthless — its job is to throw those away quietly, and to be right on the rare occasion it does not.

How it works

Generate blind. The generation step cannot see the portfolio, the watchlist, or any existing thesis. Not by instruction — by construction: it runs before anything is loaded and receives a sealed brief and nothing else. An idea that agrees with us because it was told what we think is not evidence of anything.
Freeze. Candidates are hashed and written down before the engine is allowed to look at our research. That is what makes “independently reached” a checkable claim rather than a story told afterwards.
Collide. Only now is each frozen idea compared against everything already researched, and labelled: new, rediscovered by a different path, contradicting what we believe, or merely agreeing with it.
Triage. Every candidate gets a disposition and a recorded reason. Almost all are handled without anyone being told.

What it takes to interrupt a person

Attention is the scarce resource, so the bar is conjunctive: an idea must clear every one of these, not most of them.

It needs a person. Only two of the routes do. A consensus check, a curiosity, and a rejection are work the system finishes alone.
Its evidence holds. No stale source, no unverifiable locator, nothing carrying a fabrication signature.
It was fully compared and challenges the existing record. Every company already researched, with no gap left untested. An idea that reinforces a thesis we already hold is not news, however good it is.
It is at least average on all three axes — novelty, evidence, and asymmetry, each scored one to five. Their product must reach 27 out of a possible 125, which is exactly the score of a candidate that is a three on every axis. Being exceptional on one does not buy a weakness on another.

The score can only hold something back. It can never push something through.

That asymmetry is the point. A number that can only subtract cannot quietly become a ranking algorithm that overrules the gates above it. At most 3 findings from any single run reach a person; anything the cap cuts is recorded as cut, never dropped, because a silent cap reads as “nothing else qualified”.

What it refuses to do

It does not decide. The engine proposes. A person accepts or rejects, and the verdict is recorded with its reason.
It does not state a claim it cannot kill. Every idea must name the observation that would disprove it and where that observation would be found, before it can be written down at all.
It does not throw away what it rejected. Rejections keep their reasons, so a later review can ask whether the engine was wrong to be quiet.
It does not claim to be reproducible when it is not. A run whose ideas came from a model records that they cannot be replayed, and names the model, rather than implying an audit trail it does not have.

Findings

Not published

What this engine has found is private, and stays private until it becomes research worth publishing. Most of it never will: the engine exists to let hypotheses die cheaply, and a laboratory that published every idea it had would be publishing mostly noise.

The findings are read in the local workshop, where they are accepted or rejected with a recorded reason. Anything that survives that and becomes real research appears in the published record on its own merits, under the company or paper it belongs to.