The Black Box

Definition

When a system can only be studied from the outside — a "black box" — any finite set of observations is compatible with infinitely many internal explanations ("isomorphs") that fit the data equally well; behavior alone cannot pick out the "true" one. A related limit is complementarity: refining an observation along one dimension can degrade it along another, so that two mutually irreducible views of "the same" system can both be legitimate and neither be complete.

In the Book

Weinberg works through the functional relation T = f(a): if observations show T staying constant while a varies wildly, either T doesn't really depend on a ("overcompleteness") or it depends on something else entirely that isn't being tracked ("incompleteness") — and no amount of staring at existing data tells you which. He shows two different equations that both fit the same three data points and concludes "the box is black. We cannot 'see inside' to say which is the 'true' structure." He then builds the complementarity case with a toll-booth traffic camera: a fast shutter speed gives an accurate position reading but blurs velocity past usefulness; a slow shutter gives accurate velocity but blurs position; any choice of shutter speed is a compromise, and a different observer choosing a different speed genuinely sees a different, complementary picture of the same car. He extends this beyond physics into social science, citing anthropologists Robert Redfield and Oscar Lewis, who studied the same Mexican village a generation apart and produced irreconcilably different accounts — not because either was wrong, but because their observational choices structured what they could see.

Why It Matters

This dissolves the instinct that disagreement about a system must mean someone made an error. Two rigorous, data-consistent models of the same system can differ because they made different but equally legitimate scope and precision trade-offs, and pushing for more data or a "final" answer sometimes just relocates the trade-off rather than resolving it. It is a useful check against overconfidence in any single dashboard, KPI, or model as "the" true description of a complex system.