KCSA: knowledge-constrained state abstraction for LLM-agent traces — frozen artifacts, provenance manifests, and code (v1.0.0)

Frozen numerical artifacts, provenance manifests, and source code supporting the manuscript "Knowledge-Constrained State Abstraction: Finite Contracts and Conditional Guarantees for LLM-Agent Traces" (submitted to Knowledge-Based Systems). The deposit contains (i) the frozen outputs the manuscript's numerals are bound to — the native and augmented MRP for the 829-state airline snapshot, the per-record native audit, the state-aliasing report, the canonical epsilon-sweep over all 829 states, the retail construction-split numbers and sweep, the RQ5 evidence-triage results, and the frozen confirmatory bootstrap output behind the reported interval endpoints; (ii) the consistency gate that binds every numeral in the manuscript to an artifact path, selector and unit, and recomputes it; (iii) the producer scripts for the deposited artifacts; and (iv) per-file SHA256 checksums. Regenerating the claim scan against the shipped snapshot reproduces the reported counts:672 numeric occurrences, 428 exempt, 242 verified, 2 document-verified, 0 unresolved, and 123 passing gate unit tests. Scope and limits: the numerical blocks discussed in the manuscript are candidates. Authenticated production certification and validation of block evidence were not established, and the corresponding result does not claim semantic safety or analyst benefit. The deposit makes the numbers traceable and recomputable; it does not upgrade their epistemic status. Agent traces derive from the third-party tau-bench benchmark, which is not redistributed here. The full 1.1 GB working output tree is not included; only the frozen artifacts cited by the manuscript are. No provider credentials are included — the code reads API keys from the environment. Code: Apache License 2.0. Numerical artifacts and manifests: CC BY 4.0.