Decomposing complex systems into parts that can be understood
Our first initiative. Prism studies decomposition itself: when a system is too complex to model whole, which way of splitting it preserves the behaviour you were trying to explain?
Research question
When a system is too complex to model as a whole, which decompositions preserve what actually matters — and how would you know before committing to one?
Overview
Most interesting systems resist being modelled all at once. The usual response is to break them into parts. But a decomposition is a choice, and a bad choice quietly discards the behaviour that made the system worth studying — usually without announcing that it has done so.
Prism treats the decomposition as the object of study rather than a preliminary step. Given a system and a question, we ask which way of splitting it keeps the answer intact: which components can be understood in isolation, which interactions cannot be separated without distortion, and how to tell the difference before a model is built on top of the choice.
The name is literal. A prism separates light into components you can measure individually, and the separation is lossless — the parts recombine into what you started with. Most decompositions of real systems are not like that. Prism is an attempt to measure how far from that ideal a given decomposition falls.
It is a representative first initiative for Wonder Constant Labs for a plain reason: nearly every question the Lab expects to take on will begin by dividing a system too large to hold whole, and the division is where the behaviour under investigation is most easily destroyed without anyone noticing.
Approach
- Start from synthetic systems with known structure, where a correct decomposition can be checked against ground truth, before touching systems where it cannot.
- Score decompositions by reconstruction: how faithfully the reassembled parts reproduce the whole, and how much uncertainty each choice introduces.
- Build tooling so a decomposition can be stated, executed, and audited as an artefact — rather than assumed inside someone's analysis script.
- Report the cases where the approach fails as prominently as the cases where it works.
Instruments
Decomposition harness
In progressAn internal harness for declaring a candidate decomposition, running it against a system, and recording the result with its provenance. Early and internal-only.
Synthetic system generator
In progressGenerates systems with a known component structure and controllable interaction strength, so a decomposition can be scored against ground truth.
Data
Synthetic reference set
In progressA growing collection of generated systems with recorded structure. Not yet stable enough to publish, and not yet large enough to be interesting.
Models
Reconstruction error estimator
PlannedAn estimator for how much of a system's behaviour survives a given decomposition, with an interval rather than a point value.
Log
March 1, 2026 · Initiative opened
Scope defined around decomposition as the primary question rather than as a preliminary step. Problem selection and reading underway.
May 19, 2026 · Harness scaffold
First internal utilities for declaring and evaluating candidate decompositions against generated systems.
July 6, 2026 · Ground-truth generator
Synthetic system generator produces systems with known component structure, so reconstruction can be scored rather than argued about.
Known limitations
Stated before anyone has to ask.
- Everything measured so far has been measured on systems we generated ourselves. Results on synthetic systems say very little about real ones.
- The scoring approach assumes the question being asked of the system is fixed. A decomposition that is right for one question is often wrong for another, and we do not yet handle that.
- No external review. Nobody outside the company has checked any of this.
Findings to date
No results published. Prism is in development; the work to date is problem framing and instrumentation. When there is something to report, it will appear here with its methods and its uncertainty attached.