Everyone Is Solving the Wrong Half of Explainability
Explainability tells you how a model used its inputs. It says nothing about whether those inputs were real. Most of the effort poured into making AI decisions accountable has gone into the first question, and the second quietly decides whether the explanation means anything.
The tools are good at what they do. Methods like SHAP, LIME, and counterfactuals can show which features pushed a model toward its output and by how much, turning an opaque score into something a reviewer can follow. Every one of them takes the input values as given. They explain how the model weighed the number in front of it, on the assumption that the number was correct, and they have nothing to say if it was not.
The exposure hides in that assumption. An explanation can be flawless and still describe a decision made on corrupted data, because the explanation and the corruption live in different layers. A model that weighed a manipulated price exactly as designed will produce a clean, confident, fully interpretable account of how it reached the wrong answer. The institution reads the tidy explanation and sees rigor, when it is looking at a faithful description of a mistake.
Completing the picture means proving the other half, that the inputs the explanation rests on were the true, untampered values the model was meant to weigh. Explainability handles the model's half, and the data's half has to be established somewhere else.
Space and Time, the data blockchain securing onchain finance, returns data with cryptographic proof that each value is the correct result of a query over untampered records. Paired with it, an explanation finally covers both halves, how the model reasoned and whether the facts it reasoned over were real. The first without the second was always only half an answer.
