Determinism over heuristics

A heuristic is a bet that the average case is the case you will get. It is a good bet most days, and on the day it is wrong you have no way to reconstruct why. Determinism is the cheaper insurance: given the same input and the same declared epoch, the system produces the same output, and a disagreement becomes a diff instead of an argument.

The objection is that determinism costs performance. In the substrate layer this is usually false. Most of what looks like a performance win in a context pipeline is a caching win, and caching requires exactly the property determinism gives you — a stable key. We measured this while building Contex: making the compiler deterministic made it faster, because it made the output cacheable.

Determinism is not a claim about intelligence. Models are not deterministic and we do not pretend otherwise. The claim is narrower and, we think, more useful: everything beneath the model should be. The nondeterminism budget of a system should be spent in one place, deliberately, and accounted for.

We hedge here once: we do not know whether this holds at the scale of a multi-tenant store. ContexDB is the place we intend to find out, and the specification is public partly so that someone can tell us it does not before we spend a year on it.