Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory

Abstract

Abstract. A model that inherits one-line memories may pull one archived source record before acting; a directive in the store can steer that pull: a pointer, a criterion, or both. Across sixteen registered studies (179,352 attempts) we measured where the request goes under each form: a length-matched criterion exceeded a bare id on six direct-provider models (Study D) and failed its registered superiority rule on a nine-model OpenRouter-served panel (Study E); appending the id cancelled the criterion on three Claude models (Study F-x) and six byte-matched edits each had their own effect (Study G), with a replication at eighty runs per cell (Study G'), a ratification line and a two-credit budget that restored the target (Study J), five criterion wordings (Study H2), a second store (Study H1) and a decision task in which the criterion moved the choice toward the current record on Opus 5 and away from it on Fable 5.1 (Study I); a one-character plan pointer's effect (after a corrected first report) recurred under a registered re-run (Studies B, B'); on 72 generated worlds in 36 families the two registered signatures held on Opus 5 and Fable 5.1, Fable 5 followed the primary's sign, Haiku 4.5 reversed, and Sonnet 5, the GPT-5.6 endpoints and GPT-6 Astra lay near zero (Study K2); with a defensive adapter (Meta-SecAlign) at five gains the registered criterion for a gain-dependent change of the composite - criterion contrast was not met on Llama 3.3 70B and the 8B ladder's fall was described (Study K3); re-executed under a registered attenuation rule, the 8B change did not meet it - not a statement that the contrast was unchanged (Study K4); at registered power on 337 families, conditional on this instrument and execution, the 8B change satisfied the criterion (the rule's realised one-sided error at the margin in the registered grid was 1.8-3.2 % per corner, nominal 2.5 %, a finite grid and not a uniform-error guarantee; K4's status stands as its own result) while a second SecAlign++ adapter, served under the imposed Meta-SecAlign template (not its native configuration), did not (not a statement that the contrast was unchanged); their difference of +4.7 [+2.3, +7.2] points describes two fixed execution paths, licenses no superiority, equivalence or 'significant difference' claim, and nothing follows from one ladder's status differing from the other's (Study K5). All results are descriptive effects of exact edits on fixed panels with registered intervals and no mechanism claim; for the adapter-gain studies the intervals describe family-reweighting stability conditional on the recorded execution, not reproducibility across engine executions, audit replays were neither substituted for nor averaged into the analysed outcomes, no missingness gate fired and directional completions changed no status. The paper carries every number above (the three operating-characteristic figures are in its results section). Version notes (v2). Preprint, not peer reviewed. This version adds the K programme — Studies K2, K3, K4 and K5 (generated worlds; SecAlign adapter-gain ladders; the 8B attenuation registered, then re-posed at registered power with a second adapter) — to the twelve studies of v1: sixteen registered studies, 179,352 attempted episodes, 66 pages, 7 figures, 44 tables. The record holds the manuscript, its LaTeX source (identical to the arXiv v2 source package) and the complete data-and-code archive: every raw episode file and locked serving record, the frozen packages with their SHA-256 manifests, OpenTimestamps proofs and OSF deposit receipts, the analyzers, results, independent recomputation receipts and reports, the manuscript source with the generator that emits every number in the paper, the external review archives with their dispositions, and the literature search ledgers. v1 (10.5281/zenodo.22267221) remains available in the version history. Registration. Each study's package was hashed, committed and deposited to OSF before its first confirmatory call (Studies B–J and K2–K4 on project axsnm; Study K5's two-stage freeze and its results on component c48ns, made public with this version); the receipts carry the OSF server's timestamps. OpenTimestamps proofs were made before the first call for every package except Study K2's, whose proofs were made after its run had started (its pre-run ordering rests on the OSF server timestamp and the ledgers, as the paper states). These are archival deposits made before execution, not preregistrations on a registry. Integrity. Study B's headline is a post-freeze correction on the locked episodes; Study F's first confirmatory run is quarantined in full; Study K2 replaced two frozen analysis tools after the freeze with disclosed deviations; Study K5's analysis was invoked twice for a path-convention defect of its frozen package (both outputs retained, identical in every number) and the first attempt of both its ladders halted when the serving client's heartbeat to the provider failed and each was re-executed once from an empty directory; every deviation is in the studies' records and in the paper's integrity section. Related records. Paper 1: Verification Allocation in Inherited Agent Memory: Provenance Availability Is Not Provenance Use (concept DOI 10.5281/zenodo.22084498). Paper 2: When Stale Constraints Go Unchecked (concept DOI 10.5281/zenodo.22108557). arXiv: 2609.03450. AI assistance. Language-model assistants (Anthropic's Claude through Claude Code; OpenAI's ChatGPT/Codex) were used for research-design critique, implementation and execution of the runners, analysis code, adversarial review of every package and of the manuscript, and drafting; the author is responsible for every claim. Licence. Manuscript, data and records: CC BY 4.0. Code: MIT (LICENSE inside the archive). Suggested citation. Nakayashiki, K. (2026). Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory (v2). Zenodo. https://doi.org/10.5281/zenodo.22883713; all versions: https://doi.org/10.5281/zenodo.22267220