# Review note: Agentic Institutions and Agentic Representation v0.4

4 November 2026 · CounterProof Research

**Review completed.** These are editorial reviews of conceptual drafts; the proposed studies have not run.

The two conceptual preprints received sequential model-assisted criticism on 4 November 2026. A Z.ai model, reported by the endpoint as `glm-5.2`, reviewed the earlier v0.4 texts and selected source excerpts. It offered five optional improvements and identified no necessary correction supported by that material. A directed follow-up considered the supplied clarifications addressed. Those changes narrowed what the proposed studies could establish, distinguished resource allowances from actual expenditure, clarified reader outcomes and sampling, and marked two incomplete philosophical readings as provisional.

A Xiaomi model, reported by the endpoint as `mimo-v2.6-pro`, then reviewed the post-GLM substantive texts. The current review disclosure and earlier review answers were withheld from this first pass; historical review disclosures remained. The revised prose already reflected the GLM sequence. Different providers and fresh requests do not establish independent errors, and this was not a comparison on an identical manuscript snapshot.

Xiaomi proposed further changes to role separation, exposure controls, measurement-failure handling, the documentary basis of an internal example, and the wording of the decision rules. The editorial response distinguishes useful improvements from claims the supplied text does not establish: publication of a schema does not prove that a particular researcher saw it; the scoring text does not establish that only the candidate condition needs transformation; and Representation already limits reader gains without assessed-adequacy gains to a communication claim. The implemented edits make those boundaries clearer without endorsing the stronger reviewer premises.

Other clarifications concern materiality, the first study's measurement scope, consistent organisational disclosure, and record-integrity safeguards. Timestamp verification must connect the precise file bytes to the anchored proof; a digest mismatch can reveal a nonmatching file–proof pairing. Privacy depends on the commitment construction and disclosed information. The proposed integrity comparison concerns the incremental value of anchoring over signed logs, not whether either arrangement earns its cost over ordinary versioned storage. The existing scoped Popper reference is retained with its stated reading limits.

**Xiaomi correction check:** The reviewer considered the changes adequate and found no further necessary correction in the supplied passages. It accepted retaining the scoped Popper citation. The editor checked the complete manuscripts for the older commercial description, which the excerpt-only follow-up could not exclude; it is absent from both final texts.

The follow-up is not adopted wholesale. Its expectation that one study condition would suffer more measurement failures remains unestablished; the new bounds address either direction. It also describes several existing definitions and qualifications as newly added. The preserved diffs, rather than that narrative, establish the revision history. Optional wording suggestions did not require further manuscript changes. The full-text gaps for Ankersmit (2012) and Zammito (2005) remain explicitly provisional.

One follow-up launch stopped at the bench's containment check before any model request: its generated hostname was too long. A fresh, shorter run name passed the same checks without a policy or control change. That failed launch is preserved separately from the completed model call.

The reviewers received supplied text and selected excerpts through API calls, without autonomous access to files, browsers, repositories or experimental data. Most cited sources were not provided in full. Raw substantive responses, frozen review packs and the delivered Xiaomi bundle are retained in the editorial record; the exact wrapper-added Xiaomi API-message envelope was not retained. Endpoint model labels describe reported service identities, not independently verified model weights. The earlier raw judgments remain part of the record when editors disagree with them.

These reviews are criticism for the author's consideration, not journal peer review, a security audit or empirical validation. Neither proposed first study has run. Editorial responsibility remains with the author. The web revisions remain v0.4 drafts without new DOIs; existing version-specific DOIs identify the deposited v0.3 papers.
