Idea Engine
Preamble
Idea Engine asks whether a language model can help find ideas worth testing while lacking a causal model of what building those ideas would change. It uses proxies around time, subculture, rejection, and corpus density to produce test reasons with evidence and confidence.
Language Can Propose Futures
Most of the garden uses language models to make work possible: pages, code, audits, agents, briefs, maps, and proposals. Idea Engine holds up the counterweight. A model can describe possible futures in fluent detail. Description stays inside language. Consequence begins when someone acts.
That limitation matters most around ideas. A bad summary can be corrected. A weak page can be rewritten. A bad idea can consume months because it sounded plausible at the moment of selection.
Actions Change The Test
World-model research draws a sharper line. A world model learns how a state changes after an action. It observes, compresses, predicts, acts, and updates. It tries to answer a different question from a language model: if this action happens in this state, what future state follows?
Idea Engine borrows the lesson and lowers the claim. It builds proxy surfaces around traces that language can inspect: time, subculture, rejection, corpus density, and feedback.
The Edge Is Historical
The target is MAYA: most advanced yet acceptable. The idea needs enough familiarity to land and enough novelty to matter.
The problem is that a language model’s sense of acceptability is historical. It is trained on what has already been written, named, copied, rejected, celebrated, and made legible. It can search the archive of what became acceptable. The current edge has to be treated as unknown until action produces feedback.
The engine therefore treats the edge as a hypothesis. It asks where the boundary has moved before, which subcultures already accept a pattern, which objections would appear, and whether the idea sits in a dense, sparse, or alien part of a corpus.
Proxies Replace Prophecy
The seed version has four proxy methods.
Temporal Calibration asks how a domain’s acceptance edge moved across time.
Subcultural Arbitrage moves a familiar pattern from one world into another where it still has charge.
Adversarial Rejection Modeling names the objections an idea would face before pretending it has been accepted.
Corpus Density Proxy estimates whether a candidate sits in a crowded, sparse, or disconnected part of a known corpus.
Each proxy can be wrong. Every output needs evidence, confidence, category-error risk, and a falsifiable test.
The First Slice Moves Between Worlds
Subcultural Arbitrage is the first launch slice because it is plain enough to test. Find a pattern accepted in one domain. Translate it into a target domain where the same pattern is still underused, strange, or newly useful. Separate the familiar carrier from the novel payload. Attach evidence, confidence, risk, and a test.
The human still has to decide whether the transfer is worth attempting. The model can point at the gap. The cost of being wrong lands outside the model.
The Mirror Under The Garden
Idea Engine is a counterpoint to the rest of the garden. The garden depends on language models because they can produce, critique, retrieve, summarize, and structure at speed. This project asks where that power stops.
A language model can help me imagine more options. It can also make untested ideas sound as if they have already survived contact with reality. The engine earns its place by turning fluent possibility into a testable wager.
The output is a reason to act, reject, revise, or gather evidence. The future arrives after the action.