Ready to Build II: Specifying What Must Stay Distinct — Non-Collapse Conformance in the Audit Replay Agent

A specification can fix what an implementation must do and still leave it free to merge what the specification keeps apart: to report every failure through one negative outcome, to treat a resumable hold as a terminal block, or to let the label of one decision object stand in for a verdict. A companion paper (https://doi.org/10.5281/zenodo.23229267) presented execution-neutral buildability, the property of a specification that fixes required behavior and conformance criteria, bounds the mechanisms it leaves open, and can therefore be implemented by projection in different paradigms. Here we examine a further property that such a specification can have: it fixes the semantic distinctions every implementation must preserve and publishes them as checkable conformance criteria. We study it on the Audit Replay Agent pack (ARA-AUDIT-REPLAY-001), a final, manifest-verified specification for an offline agent that reconstructs the decision path, checkpoints, and rollback sequences of a ledger-backed session after the fact and determines whether they are complete, coherent, and verifiable without live access to the underlying system. The pack draws its distinctions at several levels — six decision objects with their own admitted verdicts, a closed canonical verdict set, gate statuses, and workflow states — resolves fourteen lexical overlaps in its Ambiguity Register, and gives every conformance record a forbidden or bounded condition and every oracle a forbidden collapse. It also bounds what it leaves open: two outcomes delegated to deployment within declared values, and an enforcement carried by gate logic. In the pack's NNT companion run, five paradigm-isolated reconstructions in Python, Rust, Prolog, C, and Bash ran the 27 conformance records live and each reported 26 PASS, 1 SKIP, 0 FAIL, and 0 ERROR, with no divergent record and zero corrections; ten of their 34 clarifying questions concerned distinctions, and the pack resolved each of them. A supplementary artifact renders a declared 271-input projection of the workflow in five further paradigms with byte-identical outputs and runs the full suite in AWK and in spreadsheet formulas with the same profile. The evidence has a bound that we report: in the recorded mutation run, the verdict-swap mutants built in two of the reconstructions changed no record's status, so the suite checks distinctions at the level of vocabularies and routing rather than every verdict on every path. We read the results as evidence of derivability for this artifact, not as a certification, a quality score, or a claim of regulatory adequacy. Use of AI tools. The author constructed the pack's seven canonical files, including the conformance tests, reviewed them manually, and designed and calibrated the harnesses for the reconstruction campaign and the supplementary derivations. Large language models were used openly, as tools to accelerate research and production, with manual review: they assisted the drafting of the pack's documents and the revision of this manuscript and, by design, built the reconstructions and the supplementary derivations that are the object of the demonstration. No tool is an author; the author is accountable for the content. What matters is not which tools touched the documents but whether the artifact can be verified. The claims in this paper are limited to what a reader can check with the shipped hashes, the runnable conformance suite, and the reproducible outcome vectors.