Currentness Under Conflicting Agent State: A 24-Case Conformance Suite Beyond Latest-Write-Wins

Long-lived agents routinely accumulate several records about the same fact, yet the most recent write is not necessarily the current truth. This technical note defines a small executable currentness contract over stable subject identity, scope, valid time, provenance/admissibility, and explicit revision semantics, then exercises it with 24 frozen synthetic conflict cases. Latest-write-wins conforms on 10/24 cases and an authority-first heuristic on 16/24. A reference resolver that implements the contract conforms on 24/24; that score is a specification self-consistency check, not a model benchmark or claim of generalization. Version 0.2 adds a targeted prior-art adjudication. Close work covers nearly every individual ingredient, but no earlier equivalent executable conformance artifact was found in the reviewed set that jointly tests all six dimensions under one explicit currentness contract, including UNKNOWN when no unique current value is justified. The bounded novelty claim therefore concerns the executable combination and released case surface, not the individual mechanisms.

Author correction notice — 13 October 2026 (v0.2)

In an adversarial self-review after publication, I found an out-of-suite counterexample to the relation-aware reference resolver in Currentness Under Conflicting Agent State.

Section 3 requires UNKNOWN when incompatible maximal candidates remain without an admissible relation resolving their conflict.

Consider three admitted records A, C and B for the same subject and slot. B is a valid CORRECTION of C, while A remains unlinked. The published reference implementation selects B, although A and B remain incomparable under the stated relation semantics. The expected result is UNKNOWN.

This is an author self-review finding, not independent adjudication.

The frozen 24-case results remain reproducible: 10/24 for latest-write-wins, 16/24 for authority-first and 24/24 for the relation-aware reference on the author-labeled cases. They do not, however, establish correct behavior on additional relation-target scenarios.

The distinction between currentness and recency remains valid as a conceptual argument. The published reference implementation should not be relied upon as a general conformance oracle until a documented correction and independent negative-control tests are available.

The original v0.2 files remain unchanged. Independent challenges, including disagreement with my interpretation of Section 3, are welcome.