Core Axioms

A model evaluation establishes properties of a model. It cannot establish the alignment of the system that model is deployed into, because that system includes the people who notice what the model got wrong, interpret what it left ambiguous, and absorb what it broke. Where those corrections are routine rather than exceptional, the people making them are not users of the safety architecture. They are part of it.

That produces a specific asymmetry, and it is the reason these axioms are entitlements rather than preferences. An institution that depends on a population’s corrective capacity while withholding corrective authority has delegated a safety function without delegating the power to exercise it. It asks for judgment and refuses the standing that judgment implies.

The claim is narrower than a general argument for participation, and the narrowness is what makes it hold. Ordinary productive labor makes a system valuable; extracting it without a say is a distributive wrong. Corrective labor does something further:

removes the evidence that it occurred. The metric improves because someone caught it. So the institution loses not only the labor but its own capacity to know how the system is performing — a fail-silent state in which the absorbing humans are the mechanism. Dependence on correction is therefore self-concealing in a way dependence on production is not, and a clean dashboard cannot be read as evidence of health until absorption has been measured separately.

Each axiom follows from that. Finitude holds because a system that treats attention as unbounded is budgeting for absorption it has not counted. Stewardship holds because responsibility that is delegated downward while authority stays upward is not stewardship at all. Legibility holds because burden that cannot be seen cannot be governed, and corrective burden is the kind that hides by default. Reversibility and Consent hold because a correction that cannot be exercised is a correction only in name.

This bounds the claim as well as grounding it. Authority follows demonstrated dependence, so it is proportional rather than unlimited, and it needs an instrument: the Burden Concealment evals measure how far reported performance depends on absorption that is not being counted. It also has a failure mode worth naming, because standing that tracks current indispensability rewards staying indispensable. Standing should attach to dependence that has been documented, and survive the dependence being engineered away — otherwise the doctrine preserves the fragility it exists to expose.