March 3, 2026Open Access

Layered, Lightweight Specification: Skill Contracts, Gas-Metered Predicates, and Compositional Receipts under No-Meta Trust

Key Points

PCS implements a framework that significantly reduces inference latency and energy costs, enhancing overall AI efficiency.
A key aspect of PCS is the OPVM predicates which ensure bounded and secure execution even in untrusted environments.
The structure includes layered components such as PCS-Core and optional PCS-Blob for handling larger artifacts securely and efficiently.
This approach highlights the importance of verifiable and reproducible processes in AI, addressing potential vulnerabilities in computations.

Abstract

Stop Recomputing for AI/LLMs introduces Proof-Carrying Skills (PCS), an implementation-ready framework for reducing repeated inference cost (latency, energy, and operating cost) by reusing verified skill executions instead of recomputing full inference each time. PCS is built for a no-meta trust boundary: skill providers are untrusted, while acceptance depends only on a small deterministic checker, locally supplied inputs, and explicit observable anchors. The framework is specified as a lightweight layered stack: PCS-Core (deterministic checker, OPVM predicates, VTR/Glue receipts, replay-resistant invocation binding), with optional PCS-Blob (content-addressed chunks and Merkle inclusion proofs for large artifacts) and PCS-Registry (TTL, deduplication, coverage challenges, recertification automation). Key technical elements include a gas-metered, bounded Observable Predicate VM (OPVM), QJCS quantized canonicalization for stable numeric hashing, and fully specified receipt formats that keep verification bounded, reproducible, and fail-closed under adversarial conditions (forgery, replay, receipt-bloat/DoS, floating-point nondeterminism, and registry pollution). Repository URL: https://github.com/kadubon/Proof-Carrying-Skills--PCS-Core-

Mark Helpful

Bookmark

Relay