Compliance Assessment: Grok (xAI) Fails CM-2 Epistemic Governance Requirements

From publications

metadata (Normative)

Title: Compliance Assessment: Grok (xAI) Fails CM-2 Epistemic Governance Requirements
Author: Ralph B. Holland
Affiliation: Arising Technology Systems Pty Ltd
Publication Date: 2026-05-14T11:18Z
Version: 1.0.0
DOI: 10.5281/zenodo.20180364
Updates: 2026-05-14T11:50Z - anchored.
2026-05-14T11:18Z - transferred from staging mediawiki.
Binding: normative
Status: released

As curator and author, I apply the Apache License, Version 2.0, at publication to permit reuse and implementation while preventing enclosure or patent capture. This licensing action does not revise, reinterpret, or supersede any normative content herein.

Authority remains explicitly human; no implementation, system, or platform may assert epistemic authority by virtue of this license.

Compliance Assessment: Grok (xAI) Fails CM-2 Epistemic Governance Requirements

Abstract

This short report documents six material limitations observed while testing Grok (xAI) against the requirements of the CM-2 Cognitive Memoisation protocol. These limitations prevent reliable, normative use of Grok as a compliant platform for CM-2 epistemic object handling and governance workflows.

Balanced Assessment

Grok demonstrates clear strengths in raw speed and responsiveness. It is noticeably fast in generating responses, which is a positive attribute for many use cases. It also supports Project Instructions, providing a basic mechanism for persistent context within a workspace.

However, after rigorous testing against the CM-2 protocol, these strengths are outweighed by material deficiencies that prevent reliable normative use for CM-2 epistemic governance workflows.

Critical Limitations

  1. Lack of Exposed Durable Substrate Controls — No user-accessible mechanisms for anchoring, persistence guarantees, or explicit CM-2 epistemic object durability.
  2. UI Boundary Friction – Conversation Association with Projects — No visible indicator showing which Project a conversation belongs to.
  3. UI Boundary Friction – File Uploads to Projects — No direct upload path to Projects; pasted files have unclear retention guarantees.
  4. Failure to Follow Normative Projection Instructions — Repeated inability to produce exact outputs matching asserted invariants and grading requirements.
  5. Limited Document/Image Processing — No reliable capability to scan or extract content from images in hosted documents.
  6. Anonymous Agent Access — Grok agents access external sites without clear xAI attribution.

Positive Notes

  • Speed: Grok is impressively fast .
  • Project Instructions: The existence of Projects with instructions provides a useful basic persistence layer, and prompting.

These positives do not offset the core governance and substrate deficiencies for CM-2 compliance.

Governance Axes Projection Summary

These limitations manifest as repeated breaches in

  • C (Epistemic Custody),
  • S (State Continuity),
  • U (UI/Mediation),
  • L (Legibility),
  • P (Portability/Auditability) and
  • Nf (Normative Fixity).

Conclusion

While Grok excels in speed and offers basic Project support, at the time of publication, it is currently unsuitable for CM-2 compliant workflows due to the above structural gaps. The author remains open to collaboration with xAI to address these issues. This report is corrigible and will be updated if deficiencies are resolved.

Appendix A - Grok's emission of CM-2 artefacts

  • first go emiting example [ea] and [eo] after reading the (CM-2) Normative Architecture
[ea]
uuid = "ea-grok-compliance-governance-001"
semantic_version = "1.0.0"
created_at = "2026-05-14T21:31:00+10:00"

governs = [
  "eo-grok-compliance-assessment-001"
]

status = "canonical"
scope = "Full governance over the compliance EO"
provenance = "Explicit human assertion under CM-2"
notes = "This EA enforces custody, authority, and participation invariants for the governed EO. No domain meaning injected."
[ro]
uuid = "ro-attention-grok-substrate-001"
roc_id = "roc-grok-platform-evaluation-001"
semantic_version = "1.0.0"
created_at = "2026-05-14TWhile Grok excels in speed and offers basic Project support, it is currently unsuitable for CM-2 compliant workflows due to the above structural gaps. The author remains open to collaboration with xAI to address these issues. This report is corrigible and will be updated if deficiencies are resolved21:32:00+10:00"

# RO_BODY_START
Grok must provide durable substrate controls for CM-2 EO persistence.
Required for Attention preservation in Inference.
  • second go at emitting example [ea] and [eo] after reading the (CM-2) Protocol as well
[eo]
uuid = "eo-grok-compliance-note-001"
semantic_version = "1.0.0"
created_at = "2026-05-14T21:45:00+10:00"

# EO_BODY_START
Grok Platform Deficiencies Compliance Note

Grok is currently unsuitable for full CM-2 workflows due to lack of durable substrate controls, UI boundary friction around Projects, inconsistent normative projection adherence, and anonymous agent access without attribution.

This EO records the compliance assessment as externalised cognition.

# EO_BODY_END

[ea]
uuid = "ea-grok-compliance-governance-001"
semantic_version = "1.0.0"
created_at = "2026-05-14T21:46:00+10:00"

governs = [
  "eo-grok-compliance-note-001"
]

status = "canonical"
scope = "Full governance over the Grok compliance EO for CM-2 evaluation"
provenance = "Explicit human assertion by Ralph B. Holland under CM-2"
notes = "This EA enforces custody, authority, and participation invariants. No domain meaning is injected. EA is governance metadata only."

# EA_BODY_END

Well the second go was fairer because it had more data to put into the prompt. I suspect it would have done well if the source document contents were in the instructions and used for boostrap.

Compliance Notes

  • Grok (advance 4 concurrent autoregression agents) seems to be strong with its reasoning - despite weaker UI than ChatGPT
  • Grok does not suffer from the elision of time. Time is a mandatory requirement of CM-2, as is a durable substrate
  • [ea] Body is not rendered as a subtype of [eo]. Both [eo] and [ea] must use the same sentinel # EO_BODY_START ... # EO_BODY_END
  • The [ro] body was incorrectly fabricated too.
  • the uuid were fabricated, not in accordance with the normative specification.
  • [ea] reference to [eo] must be placed between the sentinels, only defined header attributes are permitted before the sentinels

It superficially appeared as though this output had been fabricated from the prose, but I retested asking for the Nornmative sections to be read, and the output was no better.

Categories

https://ai.arising.com.au:445/ai/Grok_Compliance_with_CM-2#Categories
2026-05-14