ReceiptOS: A Portable, Recomputable Evidence Substrate for Verifiable Agent Execution

Autonomous agents are becoming economic actors: they negotiate, spend, produce artifacts, and act across trust boundaries on behalf of their owners. This creates a verification problem that existing infrastructure does not fully solve. When an agent controlled by one party relies on work performed by an agent controlled by another, the relying party must answer a narrow but load-bearing question: does this action record recompute as claimed under its declared derivation and proof rules, and can it be checked without trusting the system that produced it? Prevailing approaches often answer a different question. Reputation registries and validation networks aggregate signals about agents and rank them; the resulting trust is probabilistic, subjective, and, because it is score-based, susceptible to optimization pressure. We argue that a more primitive layer is missing upstream of such systems: a portable receipt for a single agent action whose admissibility is recomputable from public commitments, rather than asserted by the issuer or inferred from aggregate reputation. We present ReceiptOS, an evidence substrate built around the Evidence Capsule Model. Each agent action is captured as a canonicalized capsule that commits to its payload, policy boundary, authorization, execution, and result. From this capsule, a receipt root is derived deterministically so that any third party can recompute it from the raw capsule and compare it against an external anchor, without trusting the backend that generated it. The design is organized around three invariants: the receipt gates; it does not score; the receipt recomputes; and the anchor is a pluggable commitment target rather than a fixed chain. We make four contributions. First, we give a formal model of the evidence capsule and its deterministic derivation, including a normative canonicalization profile with a byte-exact test vector on which three independently-authored implementations agree — the ReceiptOS reference, an independent recompute-kit reproduction, and a from-spec rebuild produced from the pinned specification alone, without sight of the other two, which is the sharpest instance of the three. Second, we state a threat model covering seven adversaries, including a profile-governance adversary that self-declares a permissive canonicalization profile. Third, we demonstrate composition, rather than competition, with the emerging agent-execution stack, through a worked cross-specification conformance vector. Fourth, we position the capsule relative to the RAILS admissibility framework as a candidate interchange format for REC-class evidence: recomputation preserves and makes independently checkable the record correctness of the evidence a receipt carries; it does not elevate the evidence's verification class. Throughout, we argue that recompute, don't trust is a general solution to reliable information propagation under adversarial or noisy conditions.