The Observability Manifest: Warrant and Reliance in Operational Evidence

Two questions stand behind an operational claim, and they are not the same question. Whether a measurement reliably says what it says it says. And whether it can be used for what it says it can be used for. The first is answerable by the thing doing the measuring. The second is not, because sufficiency depends on a decision the measurement has no knowledge of. Observability has mature machinery for the first and almost none for the second, and the gap is rarely visible when a claim is made and frequently unrecoverable afterward: what a party required before relying on a number is not the kind of thing that survives in a log. Each element of the problem has been treated somewhere. Measurement definition, provenance, context-of-use qualification, professional reliance disclosure, and evidence sufficiency all have established treatments, agent decision-basis and outcome verification have acquired them since 2026, and consumer-authored acceptance records exist. A cross-disciplinary review, refreshed 13 September 2026, did not locate their conjunction: an acceptance authored by the relying party, bound to the version of the producer declaration governing the evidence used, scoped to a stated decision purpose, recording the limitations accepted, and held so that neither the producer's declaration nor the consumer's acceptance can be manufactured from provenance, reconstruction, or third-party assessment. This paper proposes the Observability Manifest, a claim-scoped inventory of the promises in force when an operational assertion is made and the reliance placed on them. Four blocks record what a producer declares about its own measurement, what a relying party required and accepted for a purpose, what a comparison rests on, and what is known about whether those conditions still hold. From four relations a claim depends on and three states the promise governing any of them can be in, the paper derives three routes by which warrant fails: a relation never declared, one declared and exceeded, and one adequate when bound but not when relied upon. Worked cases show each arising while every component operated correctly. The contribution is normative composition rather than representation. Existing provenance models can likely carry the information; what is proposed is a set of invariants constraining authorship, version binding, purpose-scoped acceptance, and what reconstruction may assert. The paper claims efficacy under stated conditions and does not claim that adoption improves outcomes or that the invariants are irreducible to existing models. Both are stated as tests others can run. The discipline is NOT new. Practitioners have always reconstructed the basis of a measurement before staking anything on it, and rationed what they would defend by whether they could reconstruct it under questioning. That worked because the person reading a number could usually find the person who built it. What this paper proposes is the same discipline written down.