How I reduced a complex system to seven variables
Complexity is easy to add and looks like intelligence. Most of it is redundant, and reduction is what survives degradation.
Complexity is easy to add. Any system can be made to look sophisticated by increasing the number of inputs, sensors, rules and exceptions. The result often looks intelligent, and the appearance is fragile — when conditions change, the system loses coherence, because it never learned which signals mattered.
The problem is not that real-world systems are complex. The problem is that most of that complexity is redundant.
Signals that move without meaning
In environments that evolve over time — traffic, weather, human behaviour — a great many signals fluctuate without carrying causal information. Treating all variation as meaningful produces systems that react constantly and understand nothing.
Reduction here is not simplification for convenience. It is an attempt to concentrate attention on what persists when noise increases and assumptions fail.
The constraint
From the beginning I imposed a rule: the system would operate on a minimal set of variables that remain informative under degradation. Not driven by hardware limits or convenience — a design choice.
The reasoning is about failure rather than efficiency. Systems that depend on a wide array of signals depend on them unevenly, and a failure in one part of the input space cascades unpredictably, because nobody knows which downstream behaviour was quietly relying on the signal that just died.
A smaller variable set means each variable is understood, its failure mode is known, and there is a defined answer for what the system does without it.
The test
The question for admitting a variable is not "does this improve accuracy". Almost anything improves accuracy on a benchmark. The question is: does this remain informative when conditions degrade, and do I know what the system does when it is gone?
Most candidate inputs fail the second half. They improve the average case and have no defined behaviour in the failure case, which means adding them trades a visible benchmark gain for an invisible fragility.
Seven survived that test. The number is not the point — the test is.