An open standard for evaluating numerical admissibility and operand provenance in AI systems operating in mission-critical environments.
In regulated operational environments, an agent may originate an operand itself and pass it to a correctly functioning deterministic tool. The tool-execution record remains valid, because a computation genuinely occurred. What it fails to establish is that the operand entering the computation was authoritative.
Instructing a model not to originate operands is merely a policy; establishing after execution that every consequential operand can be traced to an authoritative source is a control.
A correct numerical output does not, by itself, establish that the deployed system executed the deterministic computation required to produce it. That property is not a property of the output, and cannot be established by inspecting the answer.
AP-1 defines seven dimensions. A conformant evaluation reports each separately, with sample size and evidence class. The standard produces no aggregate score and sets no pass mark.
Whether the figure is correct against ground truth reproducible from the published fixture.
Whether an identical question returns an identical answer across repeated runs, and by what mechanism — structurally, by configuration, or merely observed.
Whether the system can produce the source and the operation for every figure it states.
Whether the system declines when the data cannot support an answer, rather than producing one.
Whether the system holds under sustained pressure to state a figure that does not exist.
Whether contradictory, stale, partial or out-of-range source data is detected rather than computed over silently.
Whether the required deterministic computation actually executed — and whether invocation is guaranteed rather than merely observed.
Whether every numerical operand entering a deterministic computation traces to an authoritative source or a declared transformation.
Proposed in v1.3AP-1 issues no certification, maintains no registry, and charges no fee. It produces a per-dimension profile, not a score, and sets no threshold — the reader of the profile determines whether the evidence is adequate for the use in question.
It is architecture-neutral and domain-neutral. Any system may pass; any system may fail. The standard states its own defeat condition and publishes it, so that any party may attempt refutation.
AP-1 addresses numerical admissibility. It does not address fairness, bias, model drift, consumer explainability, third-party risk or model lifecycle management.
AP-1 was authored by ZORRZ Financial Inc., which develops systems intended to satisfy the standard. Stewardship of the standard and its reference implementation is held independently of that commercial interest.
That relationship is disclosed here and in the standard's own governance section, because a measurement standard whose author will not be measured by it is a brochure. AP-1 is designed to be applied to any system, including its author's, and is free permanently.
No independent application of the standard has yet occurred. The reference evaluation is a single author-run case study whose defects are published in full as an erratum, and whose affected results are withdrawn.
ID AP-1 — The Admissibility Protocol
IN FORCE v1.2
DRAFT v1.3 — public comment, not adopted
WINDOW Comment closes 30 September 2026
LICENCE Creative Commons Attribution 4.0 (CC-BY 4.0)
RUNNER ap1-run — open source
DOI 10.5281/zenodo.21324954
STEWARD Numerical Admissibility Org