2. Constraint Without Final Authority

The constraint the investigation found in logic was not quite the same as the constraint we found in probability or causality.

In logic, once enough of the representation remained fixed, a conclusion could follow or a counterexample could remain. Probability allowed several possibilities to stay open while constraining the weights they could carry. Bayesian updating constrained how those weights changed under supplied priors, likelihoods, evidence, and a model.

Causality added something different again. Under a causal model, observational information could sometimes determine an intervention quantity. In other cases it could establish that the quantity was not identified: more than one intervention answer remained compatible with what had been supplied.

The methods also had something more concrete in common. They made enough of the relevant structure available for inspection and manipulation. In the small logical cases we could enumerate the relevant possibilities and look for a counterexample. Probability let us construct and partition possibility spaces and follow how weight changed within them. Causal diagrams kept structural relations visible so that interventions and their consequences could be followed through the model.

Once enough of the arrangement remained fixed, other parts of the inquiry were no longer independently open. A conclusion followed, a counterexample remained, weights were constrained, an intervention quantity was identified, or several answers were shown still to fit.

This is the sense in which Part II found constraint without final authority.

The conditions carrying the constraint remained examinable. A premise could later become doubtful, or never become convincing in the first place. A probability model could omit an important possibility. Or a causal graph could represent the situation badly.

Their consequences were still constrained while they remained in place.