T: does the evidence cover where the decision lives

Our own measured law: transfer ≈ (construct preservation) × (evidence self-containment). Measured on a held-out sample of Cochrane risk-of-bias assessments: the one dimension whose supporting evidence lived outside the artifact being graded came back at κ = −0.23 - confidently anti-correct, not merely degraded. No dimension reached the human agreement ceiling, and that ceiling is itself inflated. The other two degraded toward it in proportion to how much of the construct survived and how much of the evidence sat inside the artifact: κ = 0.62 where the construct means the same thing in both domains and the paper states it outright, κ = 0.23 where the construct is adjacent and the evidence has to be inferred. That one inverted.

The measured law is a product, not a sum: transfer degrades multiplicatively in construct preservation and in evidence self-containment, so a claim can fail transport either by measuring a different construct than the decision needs or by never having looked at the part of its own support the decision will actually visit. The rungs below are keyed to the second failure - how much of the decision’s own support the evidence covers - and the figure after them draws it directly.

A similarity metric is symmetric: it reports how close two things are, and closeness runs the same in both directions. Containment is not. A region can contain another without being contained by it, and that asymmetry is the entire content of the relation T asks for. Evidence whose coverage is a superset of the decision’s region transports, because every configuration the decision can land in was already exercised somewhere inside the evidence. Evidence that merely sits near the decision, however that nearness is measured, licenses nothing about the configurations the sample never visited.

A distance metric cannot represent this failure. It can report that two regions are close on average, but it has no vocabulary for the actual problem: the decision lives partly where the evidence never looked, and averaging distance over the part that was covered hides exactly the part that was not. Containment states that directly, as one inequality, because it checks a superset relation rather than a proximity score. The relation is the wrong one only if the underlying regularity is genuinely global. If the same law holds everywhere, any sample transports regardless of where it was drawn, and T collapses to a constant - there is nothing left for it to discriminate.

Three cases against the same fixed evidence region: the decision’s region sits fully inside it, straddles its edge, or sits entirely outside it. Only the first transports. Each case is annotated with the single T rung that names it most exactly.

telemetry in the catalog: T 5, I 0. Exactly your population, exactly your period, exactly zero identification. Note this record cannot ALSO be the I witness - the first draft cited it for both rows, which made it an anti-witness for I. The I witness is obs-slope.

Breaks in reflexive domains, where the sign inverts: more live exposure CONSUMES the regularity rather than confirming it. Pricing is reflexive by construction.

Pricing is reflexive because the estimate is acted on, and acting on it changes the thing it estimated: competitors reprice against the move, and customers adapt to the new price level. Live exposure to the estimate does not confirm the regularity that produced it, it consumes it. Every period the estimate sets a price, the market that generated that number stops existing in the form it was measured in, and the next reading is taken on a population already conditioned by the last decision. That inverts the ordinary sign on evidence: more of it usually raises confidence, but in a reflexive domain more live exposure to the same estimate erodes the very regularity the exposure was supposed to be confirming. The consequence for this axis is exact and uncomfortable: a high T rung, exactly your population, exactly your period, is where the estimate is least trustworthy, not most - the one place T argues against itself.