Principle
Warrant: Foundation
This work is early.
A representation of industrial capability is not built by declaration. It is built by accumulating observations of real work, over years, and by being wrong in ways that can be measured and corrected. What follows is the part of that most companies leave out.
01 · Where the evidence currently runs out
Principle
Warrant: Foundation
Named plainly, because naming it is the discipline.
The representation is calibrated against a limited history. Confidence intervals reflect that, and they are wider than they will eventually be. They are reported at their real width rather than a comfortable one.
Where a benchmark is anchored to published incident and industry data rather than to an observed worker cohort, it is a theoretical reference and is described as one. It is not presented as an empirical finding.
Validation requires that a judgement be checked against what actually happened after deployment. That loop takes as long as real projects take. Until it closes for a given class of work, estimates in that class remain provisional.
Epistemic coverage, the fraction of industrial reality inside the representation at all, is the smallest of the five state variables today. Regions not covered are reported as not covered rather than interpolated across.
Four of this firm's claims are unresolved in the theory rather than thin in the evidence, and those are set out separately, in what is not known.
02 · What this means in practice
Provisional is a state, not a caveat.
An output marked provisional is not a weaker version of a finished output. It is an accurate description of what the evidence currently supports. The marking is removed when two things are true: the formal stopping rule is satisfied on its own terms, and a real placement outcome has validated the prior.
Neither can be waived because a customer would prefer a cleaner document. A system whose confidence responds to commercial pressure is not measuring anything.
03 · What a full description would have to carry
Warrant: Derivation
Coverage is measured against the description the problem actually requires, not against the description that is convenient to produce. Any account of what a person can do resolves onto eight dimensions, and a claim that is silent on any of them is silent about how much it is worth.
Coverage is stated per dimension, and the dimensions where it is thinnest are named rather than averaged away.
Each dimension taken separately, and the sort into exposed and protected
A judgement known to be unreliable is treated with suspicion. A judgement wrongly believed to be reliable is acted upon.
An instrument that reports certainty it does not possess is worse than no instrument.
A judgement known to be unreliable is treated with appropriate suspicion. A judgement wrongly believed to be reliable is acted upon. The second does far more damage, and it is the failure mode that measurement systems in this field most commonly produce.
That constraint is not a caveat attached to the work. It is the discipline that makes the work worth anything at all.