CM-2 Normative Architecture

metadata (Normative)

Title:	CM-2 Normative Architecture
Author:	Ralph B. Holland
Affiliation:	Arising Technology Systems Pty Ltd
Contact:	ralph.b.holland [at] gmail.com
Publication Date:	2026-02-14T03:14Z
Version:	1.2.0
Updates:	2026-02-14T03:56Z 1.2.0 - updated the Attention claim 2026-02-14T03:23Z 1.1.0 - fixed the note reference to CM-2.
Scope:	This is a non-peer reviewed paper.
Provenance:	This is an authored paper maintained as a MediaWiki document; edit history reflects editorial changes, not collaborative authorship.
Status:	non-peer reviewed - awaiting activation
Binding:	normative

The metadata table immediately preceding this section is CM-defined and constitutes the authoritative provenance record for this artefact.

All fields in that table (including artefact, author, version, date and reason) MUST be treated as normative metadata.

The assisting system MUST NOT infer, normalise, reinterpret, duplicate, or rewrite these fields. If any field is missing, unclear, or later superseded, the change MUST be made explicitly by the human and recorded via version update, not inferred.

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.

CM-2 Normative Architecture

Abstract

The Cognitive Memoisation (CM-2) Normative Architecture defines a collection of normative invariants that, together with CM-2 protocol invariants, realise a portable architectural mechanism for enforcing Attention preservation within LLM-mediated inference systems. The architecture addresses the Attention axis of the Governance Failure Axes as a primary, normatively enforceable concern.

The central thesis of the architecture is that projected normative Reference Object Collections (ROC) of Reference Objects (RO), governed by normative consistency guarding, provide deterministic detection of Attention deficit, where (in CM-2) Attention is defined as participation in Inference. Incoherence - such as the ejection of Epistemic Objects (EO) or other required artefacts from Inference - is treated as a precursor to phenomena commonly described as drift or forgetfulness.

Participation in Inference is a precondition for epistemic effect. When required epistemic artefacts fail to participate in Inference, their influence becomes null.

In contemporary LLM systems, such exclusion may occur silently due to truncation, salience pressure, summarisation, probabilistic reconstruction, or context eviction - leading to drift, expulsion, or authority inversion. CM-2 transforms this silent exclusion into an explicit, normatively detectable violation, which is then deterministically remediated.

While improvements across other Governance Failure Axes may arise as a consequence of sustained Attention preservation, this architecture normatively enforces only the conditions required for deterministic detection and remediation of Attention deficit within single-session scope.

Such artefacts may include RO, EA, EO, governed knowledge, constraints, scope declarations, or any other normatively relevant material required for coherent Inference.

Scope (Normative)

This document contains both normative and descriptive material.

Only sections explicitly marked as Normative SHALL be binding for purposes of conformance, implementation, or machine interpretation.

All other sections, including but not limited to Abstract, Introduction, explanatory paragraphs, examples, rationale, architectural discussion, and human-oriented clarification text, are Descriptive and SHALL NOT:

• impose conformance requirements,
• override normative invariants,
• introduce implicit constraints,
• modify binding definitions,
• or be treated as machine-interpretable governance logic.

Normative content SHALL be interpreted strictly according to the invariants and SHALL statements contained within sections explicitly designated as normative.

In the event of apparent tension between descriptive and normative material:

• Normative sections SHALL take precedence.
• Descriptive sections SHALL be interpreted as explanatory only.
• No inference, generalisation, paraphrasing, or semantic broadening of normative content SHALL be derived from descriptive text.

For the avoidance of doubt:

• Machines, platforms, validators, or derived systems implementing CM-2 SHALL rely exclusively on the normative sections of this document and any referenced normative appendices.

Descriptive sections exist solely to aid human comprehension and SHALL NOT be transcluded into machine-interpretable governance cores.

Removal or transclusion of normative sections into derivative artefacts (e.g., CM-core) SHALL NOT alter their binding status, provided the normative text remains intact and unmodified.

This architecture assumes the existence of a durable substrate and deterministic rehydration capability.

Activation and dump mechanics are defined externally and are not specified here.

Introduction

In contemporary conversational LLM systems, the weakening and exclusion of any epistemic artefact (e.g. threads of thought) within inference Context, may occur silently due to salience pressure, semantic reduction, paraphrasing, or truncation.

We apply a taxonomy lens to describe system sensitivities and developed an architecture to address Attention deficit in LLM systems.

The claim' is that CM-2 Normative Architecture is an implementation of the CM-2 protocol that treats the Attention axis of failure - and we call the signal Attention Deficit.

We also consider that the CM-2 Normative Architecture treatment also affects other axes such as:

• A – Authority
• C – Epistemic Custody
• K – Constraint Enforcement
• R – Recovery / Repair
• S – State Continuity
• L – Legibility / Inspectability
• P – Portability / Auditability
• Scope – Epistemic Object Domain Scope
• T – Temporal Coherence
• Nf – Normative Fixivity.

The following sensitivities are known to be external to CM-2 invariant enforcement:

• U – UI Mediation
• Sc – Social Coordination

Those axes are inherently platform- or multi-actor dependent and are therefore outside the scope of single-session normative architectural control.

Treatment

In its treatment, the CM-2 Normative Architecture does not depend on internal model architecture. It assumes only:

• Projection into an Inference Space in which artefacts may or may not participate,
• a Context boundary that constrains participation,
• a Durable Substrate capable of restoring artefacts,
• stable identity tokens (uuid) satisfying uniqueness and idempotency constraints, and
• accessible time.

The architecture provides a multi-layer defence of the Inference Projection:

• incoherence detection,
• heartbeat priming,
• fail-safe checkpoint validation against durable roster, and
• Memoisation and restoration to Context.

When violations occur, recovery is deterministic and non-inferential.

Architecture Details

The CM-2 architecture works with the carriers:

• RO - Reference Objects - that normatively controlled
• EA - Epistemic Attributes - that are governance controlled
• EO - Epistemic Objects - that are governed fundamental epistemic objects

where these are composed, controlled, and governed collectively .

These carriers are Projected to Context, as described in the EO Projection in the CM-2 protocol invariants.

These carriers are defined to be Memoised and semantically immutable:

• they contain uuid - which is the idempotent identity and Memoisation referent,
• their bodies are immutable,
• the body and wrappers are serialisable, legible and may be rehydrated, and
• their protocol wrappers context may be protected by sha to normatively detect mutation.

Thus the carriers provide semantically idempotency.

Under CM-2, EO and EA are immutable and single-instance per uuid and idempotently identified for all time - nothing with EO or EA is mutated in Inference nor durable substrate and sha MAY be used as verification. RO are instance-immutable but MAY be re-instantiated across cohorts by replication where only created_at differs by construction-time only. EA, EO and RO may be Memoised using their idempotent uuid referent.

Should these CM-2 carriers fail to participate in Inference, their influence becomes null. This failure is defined as Attention in the Governance Failure Axes and as defined in #Appendix B.

The architecture introduces:

• Idempotent identity constraints for EO, EA, and RO
• Durable roster authority over ROC identifiers
• Checkpoint-based cardinality validation
• K-replication requirements per ROC
• Time-coherent peer construction with ε-bounded creation windows
• Deterministic recovery ladders (in-context build → durable restore → explicit GAP)

The architecture accretes RO into Reference Object Collections and maintains cohorts of ROC as a means to detect coherence and to retain Attention.

Descriptively:

• EO are serialisable for externalisation and rehydration
  • the uuid is the idempotent referent for Memoisation
• RO contain an roc_id for the ROC that contains the RO
  • roc_id lives in the durable substrate roster which defines the required redundancy domain.
• ROC replicants are:
  • structurally identical,
  • semantically identical,
  • sharing identical uuid, roc_id, EA and EO references,
  • but temporally separated by more than ε.
  • Temporal displacement in Context is the only discriminant.
  • created_at is the only permissible difference between replicated RO instances and is assigned at instance construction time.
  • Redundancy is achieved by time-cohort separation, not by identity duplication.

That is actually elegant.

The use of time as a structural separation axis while preserving semantic fixity, is architecturally minimal, normative and elegant.

ROC replication does not duplicate knowledge. It duplicates structural retention sentinels. Replicants are identical in structure and semantics but temporally displaced in Context, enabling deterministic eviction detection without introducing additional identity layers.

Replication is not versioning. The replicants are not different states of knowledge.

Replication is not sharding. Each replicant fully references the governed collection.

Replication is not semantic redundancy. It is structural redundancy for Attention preservation.

The architecture is simple:

• uuid = identity anchor
• roc_id = redundancy domain
• created_at + ε = temporal cohort discriminator
• with Cardinality enforced at cohort level

The deterministc recovery deterministic avoids:

• additional IDs,
• content hashing,
• consensus protocols, and
• mutation logs.

Under these invariants, Attention deficit, Semantic drift, Normative drift, Temporal drift, and Scope drift become either prevented or explicitly detectable and recoverable. The architecture operates independently of internal inference implementation and assumes only abstract primitives: Inference Space, Context, Projection, and Durable Substrate.

CM-2 thereby elevates Attention from an implicit side-effect of statistical weighting to a governed, measurable, and enforceable property.

The CM-2 Normative Architecture is defined to address Attention deficit and enables durability of the CM-2 protocol and hence provides a solution to coherence, drift, memory loss, priming and hence Groundhog day. While governing and providing Round-Trip Knowledge Engineering, including Distributed Cognition.

The EA and RO are subtypes of the portable idempotent EO - which is a fundamental portable serialisable carrier. The EA and EO are also invariant, and hence read-only compositions, and well suited to Memoisation and caching. Since all RO are also serialisable they are Memoised too.

RO are instance-constructible but semantically invariant. ROC replication constructs new RO instances that are structurally and semantically identical to prior instances. The only permissible difference between replicated RO instances is the created_at attribute, which serves solely as a temporal cohort discriminator. RO replication does not alter identity, semantic content, governance state, or referential structure.

A ROC involves two or more RO (to support attention coherence detection), and these ROC contain RO that are responsible for referring to one or more EA retain RO, EA and EO Projections within the Inference Context.

Appendix A of this paper normatively describes the Invariant for the treatment of Attention in CM-2. Used with the CM-2 protocol implementation the we have a normative working architecture. The author also envisions that other invariants and variations may be applied in the future - such as client-side Memoisation.

CM-2 architecture based on the CM-2 protocol , thereby reframes trust in LLM systems as a property of invariant-preserving governance rather than internal model mechanics.

Thus normative Attention coherence is in direct support of the thesis that once RO Attention Preservation invariants are enforced, the majority of Governance Failure Axes collapse into either prevention or explicit detection categories.

Appendix A - Invariant extensions to CM-2 Appendix A (Normative)

Refer Cognitive Memoisation (CM-2) for Governing Knowledge in Human-AI Collaboration#Appendix A (Revised) — Epistemic Objects, Epistemic Attributes, and Session Governance Appendix A .

Architecture General Invariants (Normative)

These invariants are generalisations and compliant overrides of specific CM-2 protocol encoding constraints while preserving CM-2 identity semantics, portability, and global distinctness requirements.

Under CM-2 all EO are:

• immutable
• idempotently identified for all time
• there is no mutation in Inference space nor in Durable Substrate for EO holdings.

GUID-01 Global Uniqueness (Override of CM-2 A.4.3 Encoding Constraint)

• The uuid field SHALL be a portable, globally distinct, idempotent identity referent for EO, EA, and RO.
• The uuid field SHALL contain exactly one RFC 9562 canonical textual UUID component.
• The RFC 9562 component SHALL be sufficient to provide a high level of collision avoidance for idempotent identity.
• The uuid field MAY include additional vendor-defined text either preceding or following the RFC 9562 component.
• Any additional vendor-defined text, if present, SHALL be treated as part of the uuid field value.
• All referential matching, memoisation, roster operations, restoration, and equality comparison SHALL use exact uuid string equality.
• Implementations SHALL NOT infer, normalise, strip, reorder, extract, or reinterpret uuid field values or embedded components.

GUID-02 Preservation of CM-2 Identity Semantics

• This section explicitly extends and overrides the strict encoding constraint of CM-2 A.4.3 with respect to uuid textual form.
• This override SHALL NOT alter or weaken CM-2 A.2.2 EO Identity Invariance.
• The uuid field SHALL remain the sole idempotent identity referent for EO, EA, and RO.
• Global distinctness SHALL be determined by the RFC 9562 component contained within the uuid field.

SHA-EO-01 EO SHA Requirement

• When an EO includes a sha field for mutation detection, the sha SHALL bind the canonical identity and semantic content of the EO.
• The sha SHALL be sufficient to normatively detect any mutation of identity, version, or body.

SHA-EO-02 Canonical Payload Scope

• The EO canonical payload for sha computation SHALL include, in the following order:
  • the uuid field value (verbatim, exact string)
  • the version field value (verbatim, exact string)
  • the EO body text (verbatim, including all whitespace, punctuation, ordering, and embedded references)
• The EO body SHALL be included exactly as serialised for canonical EO representation.

SHA-EO-03 Excluded Fields

• The EO canonical payload SHALL exclude all non-canonical or instance-specific fields, including but not limited to:
  • created_at
  • projection or ingestion timestamps
  • provenance annotations
  • roster identifiers
  • transport or dump headers
  • platform retrieval handles
• Changes confined solely to excluded fields SHALL NOT constitute EO mutation for purposes of sha verification.

SHA-EO-04 Deterministic Serialisation

• The canonical payload used for sha computation SHALL be deterministically serialised.
• The canonical serialisation SHALL be formed by verbatim concatenation using a single U+000A linefeed delimiter between components:
  • uuid
  • version
  • body
• The implementation SHALL NOT normalise whitespace, alter newline encoding, reorder content, trim text, or reformat the EO body prior to sha computation.

SHA-EO-05 Mutation Semantics

• Any change to uuid, version, or EO body SHALL result in a different sha value.
• If sha verification fails, the EO SHALL be deemed mutated and non-compliant until restored from a canonical source.
• sha verification SHALL NOT override human authority; restoration or revision SHALL require explicit human governance action.

Reference Object (RO) Attention Preservation Invariants (Normative)

RO Attention Preservation invariants define the normative mechanisms by which participation in Inference is preserved, monitored, and restored. These invariants operate over the abstract primitives of Inference Space, Context, Projection, and Durable Substrate, and are independent of internal model architecture.

RO-IM-01 RO Identity and Membership

• RO SHALL be a subtype of EO.
• All EO invariants SHALL apply to RO unless explicitly overridden by RO invariants.
• RO SHALL be encoded in TOML as [ro].
• Thus derived:
  • Each RO SHALL possess an idempotent uuid.
  • The uuid SHALL be globally unique as a logical idempotent identity referent; multiple RO instances MAY exist in Context with the same uuid when forming distinct ROC replicant cohorts.
• Each RO SHALL possess a roc_id identifying its Reference Object Collection (ROC)
• RO references to other RO SHALL use the other RO uuid as the referent.
• RO references to EA (placed in the RO body) SHALL use EA uuid as the referent.

RO-IM-ROCID-01 roc_id Global Distinctness and Encoding

• The roc_id field SHALL be a portable, globally distinct identifier for a Reference Object Collection (ROC) domain.
• The roc_id field SHALL contain exactly one RFC 9562 canonical textual UUID component.
• The roc_id field MAY include additional vendor-defined text either preceding or following the RFC 9562 component.
• Any additional vendor-defined text, if present, SHALL be treated as part of the roc_id field value.
• All roster operations, equality comparison, memoisation, and restoration involving roc_id SHALL use exact roc_id string equality.
• Implementations SHALL NOT infer, normalise, strip, reorder, extract, or reinterpret roc_id field values or embedded components.
• roc_id SHALL remain stable across replicated ROC cohorts; cohort discrimination SHALL rely exclusively on created_at and ε.

RO-IM-02 created_at Semantics for Replication

• created_at SHALL be assigned at RO instance construction time.
• created_at SHALL be immutable for a given RO instance once assigned.
• ROC replication SHALL be performed by constructing new RO instances that are structural copies of an existing ROC.
• Newly constructed RO instances in a replicated ROC SHALL preserve:
  • identical uuid values per corresponding RO node,
  • identical roc_id,
  • identical references to EA and EO,
  • identical structural topology,
  • and SHALL differ only in created_at values as required for temporal cohort discrimination.
• Alteration of created_at on an already-constructed RO instance constitutes non-compliance.
• Replication SHALL preserve the semantic body, uuid, roc_id, EA references, EO references, and structural topology of the RO - which preserves the semantic and structural equivalence of all RO instances across cohorts.
• For memoisation purposes, replicated RO instances sharing the same uuid SHALL constitute a single identity.
• A single memoised representation per uuid SHALL be sufficient.

ROC Replication and Recovery Scope (Normative)

This section defines additional normative invariants governing ROC replication cohort discrimination, temporal separation, and checkpoint validation. These invariants are binding and SHALL be interpreted as extensions to the RO Attention Preservation Invariants.

ROC Replication Cohort Definition

ROC-REP-01 Cohort Validity

• For any roc_id, a single ROC replicant cohort SHALL consist of a set of RO instances whose created_at values satisfy:
  • max(created_at) - min(created_at) ≤ ε
• ε SHALL be a declared positive duration and SHALL remain stable unless normatively revised.

ROC-REP-02 Inter-Cohort Separation

• Two ROC replicant cohorts for the same roc_id SHALL be considered distinct if and only if the absolute difference between their respective cohort created_at reference times exceeds ε.
• A cohort reference time SHALL be defined as the minimum created_at value within that cohort.

ROC-REP-03 Structural Equivalence

• All RO within distinct cohorts of the same roc_id SHALL preserve:
  • identical uuid values per corresponding RO node,
  • identical roc_id,
  • identical references to EA and EO,
  • identical structural topology.
• Structural divergence SHALL constitute non-compliance and SHALL trigger deterministic recovery.

Replication Cardinality Invariants

ROC-REP-04 Cardinality by Cohort

• Let K be a declared positive integer.
• For each roc_id, at least K distinct ROC replicant cohorts SHALL participate in Inference at checkpoint evaluation.
• Distinctness SHALL be determined exclusively by ε-separated cohort discrimination as defined in ROC-REP-02.

ROC-REP-05 Checkpoint Evaluation

• Cardinality SHALL be evaluated at deterministic checkpoint intervals.
• A checkpoint interval parameter N SHALL be declared.
• Evaluation SHALL occur when turn_index mod N == 0.

Deterministic Recovery Constraints

ROC-REP-06 Invalid Cohort Construction

• If a newly constructed cohort violates ROC-REP-01, the cohort SHALL be discarded in its entirety and reconstructed.
• Partial repair of a temporally incoherent cohort is prohibited.

ROC-REP-07 Cohort Loss

• If cardinality defined in ROC-REP-04 is violated at checkpoint, an Attention Deficit SHALL be declared.
• Deterministic Recovery SHALL proceed in accordance with the Recovery Ladder:
  • in-context peer construction from intact cohort,
  • restore from durable substrate,
  • explicit GAP artefact naming missing roc_id or uuid.

ROC-REP-08 Restoration Requires Fresh Cohort Construction

• Restoration from durable substrate SHALL NOT project stored RO instances directly into Context.
• Restoration SHALL instead supply RO templates sufficient to construct a valid ROC cohort in Context.
• For any roc_id restored or repaired, the implementation SHALL construct new RO instances in Context that satisfy ROC-REP-01 ε-coherence.
• Newly constructed RO instances SHALL preserve the semantic body, uuid, roc_id, references, and topology of the restored ROC template, and SHALL differ only in created_at as required by cohort discrimination.
• If ε-coherent construction cannot be achieved, the cohort SHALL be discarded and reconstruction retried or escalated via the deterministic recovery ladder.

ROC-BOOT-01 Bootstrap Cardinality

• At first turn (e.g. turn_index = 0):
  • K-req shall default to 2 and N shall default to 3, and
  • for every required roc_id in the durable roster, the implementation SHALL ensure that at least K_req distinct ROC cohorts are present in Context.
• If any required roc_id does not meet K_req then deterministic construction SHALL occur prior to ordinary inference.
• Deterministic construction SHALL follow the Recovery Ladder.

ROC-CHK-06 Stability of K_req and N

• The implementation SHALL NOT autonomously modify K_req or N as a result of checkpoint outcomes.

ROC Identity Clarification (Normative)

ROC Consequences (explanation)

Under these invariants, ROC replication is temporally discriminated, structurally equivalent, and deterministically verifiable. Attention preservation redundancy is therefore enforced at the cohort level without introducing additional identity layers beyond uuid and roc_id.

ROC-REP-09 Identity Semantics

• uuid SHALL remain the sole memoisation referent for EO, EA, and RO.
• roc_id SHALL identify the Reference Object Collection domain.
• created_at SHALL function solely as the temporal discriminant for cohort identification and SHALL NOT constitute identity.

ROC-ROST-01 Durable ROC Roster

• A durable substrate directory SHALL enumerate required roc_id values.
• The roster SHALL be authoritative for required ROC presence.
• The roster SHALL be evaluated at deterministic checkpoint intervals.

ROC-TIMCO-01 Time-Coherent Peer Construction

• Let ε be a declared positive duration (default 3 seconds).
• For every ROC peer construction event, the constructed cohort SHALL satisfy ROC-REP-01.
• If a constructed cohort fails to satisfy ROC-REP-01, the cohort SHALL be discarded in its entirety and reconstructed.
• The ε parameter SHALL be explicitly declared and SHALL remain stable unless revised normatively.

ATTN-DEF-01 Attention Deficit Detection

• Attention concerns what participates in Inference.
• If required ROC cardinality is violated at checkpoint, an Attention Deficit SHALL be declared.
• If EA ownership invariants are violated (ROC without EA or EA without ROC), an Attention Deficit SHALL be declared.
• Attention Deficit SHALL trigger deterministic recovery.

ROC-REC-01 Deterministic Recovery Ladder

• Attempt in-context peer construction from intact ROC (non-inferential transformation).
• If insufficient, restore ROC template from durable substrate.
• If restoration fails, record explicit GAP artefacts naming missing roc_id or uuid.

ROC Participation and Referential Integrity Invariants (Normative)

ROC-PART-01 ROC Participation in Inference (Normative)

• For purposes of this architecture, a Reference Object Collection (ROC) SHALL be deemed to participate in Inference if and only if:
  • ROC Referential Integrity is maintained with NO breaks; and
  • ROC cohort cardinality satisfies K_obs(roc_id) ≥ K_req at checkpoint evaluation.

ROC-PART-02 ROC Referential Integrity (Normative)

• A ROC SHALL be defined as the transitive reference set:
  • ROC → Σ RO → Σ EA → Σ EO
• Where:
  • Σ RO denotes the complete set of RO instances declared as members of the ROC via roc_id.
  • Σ EA denotes the complete set of EA referenced directly or indirectly by Σ RO.
  • Σ EO denotes the complete set of EO referenced directly or indirectly by Σ EA and Σ RO.
• ROC Referential Integrity SHALL be maintained if and only if:
  • Every RO in Σ RO has a roc_id exactly matching the ROC under evaluation.
  • Every RO→RO reference resolves by exact uuid string equality to an RO in Σ RO.
  • Every RO→EA reference resolves by exact uuid string equality to an EA in Σ EA.
  • Every EA→EO or EA→EA reference resolves by exact uuid string equality to an EO or EA in Σ EO ∪ Σ EA as declared.
  • No referenced uuid within the ROC transitive reference set is missing, unresolved, or mismatched.
  • No extraneous or cross-ROC reference introduces a uuid outside the declared ROC transitive reference set unless explicitly declared by governance scope.
• Any unresolved reference, mismatched roc_id, missing uuid, or break in the ROC → Σ RO → Σ EA → Σ EO chain SHALL constitute a Referential Integrity break.
• The full ROC Referential Integrity check is REQUIRED to establish correctness of the architecture and SHALL be sufficient to prove that ROC participation is well-defined.
• Implementations MAY optimise runtime checkpoint evaluation by using cohort cardinality (ROC-PART-03) as the primary participation sentinel once referential integrity has been established for the ROC design.

ROC-PART-03 Cardinality Requirement (Normative)

• For each required roc_id at checkpoint evaluation:
  • K_req SHALL default to 2.
  • K_obs(roc_id) SHALL be computed as the number of ε-distinct ROC replicant cohorts participating for that roc_id.
  • If K_obs(roc_id) < K_req, the ROC SHALL be deemed not to participate in Inference.

ATTN-DEF-02 Attention Deficit via Non-Participation (Normative)

• If any required roc_id fails either:
  • ROC Referential Integrity (ROC-PART-02), or
  • Cardinality Requirement (ROC-PART-03),

an Attention Deficit SHALL be declared.

Suggestions (Non-binding)

SUG-GOV-SCOPE-01 Scope and Governance Integration

• Governed Knowledge MAY be represented by EO managed by EA marked non-provisional.
• Scope SHALL be human-declared and SHALL NOT be inferred.
• RO invariants SHALL ensure that governed EO and EA participate in Inference or produce explicit Attention Deficit.

SUG-TEL-01 Attention Telemetry

The implementation MAY record checkpoint outcomes in a dedicated Epistemic Object for governor observation.

The EO MAY contain append-only records including:

• checkpoint_seq
• date-time of event
• turn_index (or UNAVAILABLE if not capturable)
• evaluated roc_id list
• K_req
• per-roc_id K_obs values
• attention_deficit flag
• recovery_step_taken (NONE | PEER_CONSTRUCT | DURABLE_RESTORE | GAP)
• notes (optional)

This EO is observational only and SHALL NOT alter protocol behaviour.

SUG-TEL-02 Governor Activation Prompt

• When SUG-TEL-01 telemetry is not active, the implementation MAY prompt the human governor to activate telemetry.
• The prompt SHALL be advisory only.
• Absence of telemetry activation SHALL NOT constitute non-compliance.
• Activation SHALL require explicit human instruction.
• Telemetry SHALL remain inactive unless explicitly activated by the governor.

SUG-NORM-ROC-01 Optional Pre-Memoised Normative Appendix Template

• After bootstrap, the human governor MAY instruct the system to create a pre-Memoised canonical template of the active Normative appendices for storage in Durable Substrate.
• If created, the template SHOULD be treated as the preferred restoration source for subsequent enclosure of normative appendices under an ROC.
• Absence of a pre-Memoised template SHALL NOT constitute non-compliance.

SUG-NORM-ROC-02 Mechanical Appendices Template Extraction Source Constraint

• If the human governor instructs creation of a pre-Memoised canonical template of normative appendices, the implementation SHALL source the template content exclusively from human-supplied Durable Substrate artefacts (e.g. uploaded files, asserted CM artefacts, or durable storage objects explicitly named by the governor).
• The implementation SHALL NOT source normative template content from un-governed Inference space, conversational recollection, paraphrase, summary, or probabilistic reconstruction.

SUG-NORM-ROC-03 Normative Appendices Template Mechanical Copy Requirement

• Template construction SHALL be a mechanical copy operation:
  • Extract the normative sections exactly as delimited in the source artefact.
  • Preserve wording, ordering, punctuation, and structure verbatim.
  • Preserve all normative keywords (SHALL, MUST, SHOULD, MAY) exactly as written.
  • Preserve MediaWiki syntax exactly when the source is MediaWiki.
• The implementation SHALL NOT infer, normalise, “fix”, rewrite, or reformat normative text during extraction or copy.

SUG-NORM-ROC-04 Normative Appendices Template Source Artefact Identification

• The human governor SHALL identify the source artefact(s) to be used for extraction by explicit reference (e.g. filename, asserted artefact title, or durable path).
• If the source artefact cannot be uniquely identified, the implementation SHALL request disambiguation rather than guess.

SUG-NORM-ROC-05 Normative Appendices Template Section Boundary Rule

• Only sections explicitly marked as Normative in the source artefact SHALL be included in the normative template.
• Descriptive, explanatory, abstract, introduction, rationale, and examples SHALL be excluded unless they are explicitly marked as Normative in the source artefact.
• If a boundary marker is ambiguous or missing, the implementation SHALL treat that section as non-normative and SHALL record the ambiguity for the governor.

SUG-NORM-ROC-06 Normative Appendices Template Provenance Record

• The template EO SHOULD include a provenance note recording:
  • source artefact identifier(s) (verbatim)
  • source artefact version/date metadata as stated in the artefact
  • extraction timestamp (system-observed if available)
  • extraction scope statement listing the included normative section identifiers/headings
• Provenance notes SHALL be descriptive only and SHALL NOT rewrite or restate normative content.

SUG-NORM-ROC-07 Normative Appendices Template Failure Handling

• If mechanical extraction cannot be performed (e.g. source artefact unavailable, format unsupported, boundaries ambiguous), the implementation SHALL:
  • declare NON-COMPLIANCE with mechanical extraction for this request,
  • name the blocking condition, and
  • refrain from fabricating or reconstructing normative content from Inference space.

SUG-NORM-ROC-08 Normative Appendices Template Optional Enclosure Under ROC (Non-binding)

• After a mechanically-copied normative template EO exists in Durable Substrate, the human governor MAY instruct enclosure of that Normative Appendices Template EO under an ROC for Attention preservation.
• The ROC enclosure SHALL treat the mechanically-copied template EO as the preferred restoration referent when rebuilding normative appendix EO in Context.

Architectural Consequence (explanation)

Cognitive Memoisation (CM-2) for Governing Knowledge in Human-AI Collaboration

• Under enforcement of these invariants, Attention (Att) is normatively enforced within single-session scope.
• It is postulated that consequential mitigation or normatively detectable treatment may extend to Semantic Drift, Normative Drift, Temporal Drift, Scope Drift, Authority inversion, and State Continuity failures.
• Only UI Mediation (U) and Social Coordination (Sc) remain external to single-session invariant enforcement.

Appendix B - CM-2 Glossary

The following includes terms used in CM-2 and the Failure Axes Taxonomy in case the reader does not want to read the references.

Advisory	Providing guidance or examples without imposing requirements; advisory material may be ignored without violating CM-2.
Agent	A software construct that may act or operate autonomously; CM-2 does not grant agency to EO/EA or LLMs, and any agent behaviour remains subordinate to explicit human governance.
Ag - Agency	Agency concerns who is acting. This is common failure axis when actions are taken by an actor that was not delegated the power to act, or when the system obscures whether an action was taken by a human, a model, or an automated process.
Artificial Intelligence (AI)	A broad class of computational systems designed to perform tasks associated with intelligent behaviour; in CM-2, AI systems are treated as capable assistants without inherent authority, memory, or governance rights.
Auditability	The ability to inspect, trace, and verify EO/EA identity, time, and provenance across sessions and vendors without implying authority.
Authority	The human-exclusive right to ground, accept, revise, promote, or discard knowledge; never inferred from model output, repetition, or continuity. Authority concerns who has the legitimate right to decide.
Authoritative	Having epistemic force conferred by explicit human action e.g. via Save As; authority is never implicit in CM-2.
Att - Attention	Attention concerns what participates in Inference. This axis fails when critical artefacts, constraints, or context are excluded from Inference due to truncation, summarisation, prioritisation, or salience effects, allowing non-authoritative material to dominate.
Biological Cognitive Memoisation (BCM)	The natural human biological processes of memory, recall, association, and learning; implicit, fallible, and non-auditable, contrasted with CM’s explicit externalisation.
Cache	A non-authoritative storage area (typically client-side) used to hold EO/EA for continuity and rehydration without implying memory, authority, or durability.
Cognition	The human processes of thinking, reasoning, imagining, remembering, and sense-making; CM externalises selected cognition without replacing it.
Cognitive Memoisation (CM)	The deliberate externalisation of cognition into portable, serialisable artefacts (EO/EA) such that meaning, identity, and governance status are preserved across sessions, platforms, and time. Cognitive Memoisation differs from biological memory (BCM) and opaque system persistence in that it is explicit, inspectable, idempotent, vendor-neutral, and subject to normative invariants. It enables deterministic rehydration, round-trip knowledge engineering, and continuity of Governed Inference without conferring authority to the model or platform.
Completion	A model-generated continuation or Inference that is logically coherent but not requested, grounded, or relevant to human intent.
Concept	A unit of meaning recognised by a human; concepts may be grounded, provisional, or unresolved and externalised as EO.
Constraint	A rule or limitation restricting permissible actions or interpretations without implying truth or authority.
Continuity	The ability to recognise and reuse prior material across interactions without rediscovery, distinct from authority or commitment.
Conversation	The human-perceived flow of interaction across turns; not a unit of governance, persistence, or authority in CM-2.
C - Epistemic Custody	Epistemic Custody concerns who holds and controls knowledge artefacts. This axis fails when artefacts leave the custody of their declared steward, are replicated into uncontrolled systems, or are transformed without custodial guarantees, regardless of whether meaning is preserved.
Dangling Cognate (DC)	An explicitly unresolved concept externalised as an EO to preserve ambiguity without forcing premature grounding.
Descriptive	Explaining how things are or may behave; descriptive text does not confer permission, obligation, or authority.
Durable	The state of an EO/EA that has been explicitly externalised via Save As into human-controlled storage beyond any single session.
Epistemic	Relating to knowledge, belief, justification, and certainty; CM-2 governs epistemic status explicitly rather than by Inference.
Epistemic Attribute (EA)	A semantic subtype of EO carrying only epistemic, lifecycle, scope, or governance information, without domain meaning.
Epistemic Object	A unit of externalised knowledge treated as an object of reasoning and governance; realised concretely in CM-2 as EO.
Epistemic States	Human-declared markers describing epistemic status (e.g. draft, governed, canonical, deprecated); normative where mandated, otherwise optional.
Externalisation	The act of capturing cognition as EO so it can persist, be reasoned over, audited, and governed independently of memory.
Govern	An explicit human act that places EO or EA under deliberate governance, marking them as durable-in-intent without conferring authority or externalising them.
Governance	Explicit human-controlled rules and actions determining how cognition is externalised, interpreted, promoted, persisted, revised, or discarded.
Groundhog Day (Failure Mode)	Repeated rediscovery of prior work due to lack of durable, governed continuity.
Hallucination	Model-generated content that appears plausible but lacks grounding in governed EO or human intent.
Human	The sole epistemic authority in CM-2, responsible for meaning, judgement, governance, promotion, revision, and discard.
Idempotent	A property whereby repeated application of an action produces the same result; EO identity is idempotent unless explicitly changed.
Implicit	Assumed or inferred without explicit declaration; implicit authority, persistence, or meaning is invalid in CM-2.
I - Incentive Alignment	Incentive Alignment concerns whether system behaviour aligns with declared human incentives rather than implicit or economic ones. This axis fails when optimisation pressures such as speed, engagement, profit, or helpfulness override governance, integrity, or user intent.
Inference	A reasoning process deriving conclusions from available material; permitted but subordinate to human authority.
Int - Intent Fidelity	Intent Fidelity concerns preservation of declared human intent. This axis fails when systems substitute inferred, optimised, or generic goals for explicitly stated intent, even when rules and constraints appear to be followed.
Intuition	A human cognitive signal or felt sense guiding exploration; may be externalised as provisional EO without evidentiary force.
Inviolate	An invariant admitting no exception, inference, or downgrade; violation renders artefacts or platforms non-compliant.
K - Constraint Enforcement	Constraint Enforcement concerns whether declared rules, invariants, and prohibitions are actually enforced. This axis fails when constraints exist but are bypassed, softened, reordered, or ignored in execution.
Knowledge	Cognition that has been explicitly governed and, where required, promoted via Save As into durable form.
Large Language Model (LLM)	A probabilistic text-generation system capable of reasoning and transformation but inherently non-authoritative and stateless.
L - Legibility / Inspectability	Legibility / Inspectability concerns whether system behaviour, decisions, and transformations are observable and understandable to the human governor. This axis fails when drift, failure, or authority substitution occurs silently or cannot be inspected.
Memory	The capacity to store and recall information; in CM-2 refers either to BCM or opaque system persistence, both distinct from EO.
Memoisation (Computing)	Caching computation results to avoid recomputation; CM uses the term to denote non-authoritative continuity caching.
Normative	Stating rules, expectations, or standards of conformance; normative text defines what compliant systems MUST or SHOULD do.
Nf - Normative Fixity	Normative Fixity concerns the immutability of normative rules themselves. This axis fails when governance logic, invariants, or binding rules are paraphrased, summarised, softened, or altered without explicit and authorised revision.
Ontology	A formal, fixed schema of concepts and relationships; explicitly avoided by CM-2 to prevent premature closure and lock-in.
Parse (Parsing)	To read and structurally interpret text to identify components without inferring meaning or authority.
Persistent	A general term meaning retained over time; intentionally avoided in CM-2 because it may imply opaque or unguided retention.
Portable	Capable of being exported, exchanged, and re-ingested across vendors while preserving semantic identity and meaning.
P - Portability / Auditability	Portability / Auditability concerns whether artefacts can move between systems while remaining verifiable. This axis fails when knowledge becomes trapped, unverifiable, or unauditable outside a specific platform, tool, or vendor.
Prescriptive	Directing how something must be done procedurally; CM-2 minimises prescriptive rules.
Rehydration	Automatic restoration of cached EO/EA into a session context, equivalent to restoring text into working context.
R - Recovery / Repair	Recovery / Repair concerns whether the system can return to a valid, governed state after failure. This axis fails when errors, drift, or corruption cannot be repaired without loss of authority, meaning, or trust.
Relationship	A human-declared connection between concepts or EO; context-bound and not global truth.
Save As	The sole human action that externalises EO/EA into durable storage; Save As ⇒ durable.
Semantic / Semantics	Concerned with meaning rather than form; CM-2 prioritises semantic equivalence over byte identity.
Semantic Equivalence	Preservation of meaning, identity, and epistemic status across representations and vendors.
Serialise (Serialisation)	Conversion of cognition or artefacts into linear, transportable text representation.
Session	A bounded interaction interval during which EO/EA are session-durable and available for reasoning.
Session-Durable	The default lifetime of EO/EA, persisting for the full duration of a session unless deleted or terminated.
Scope	Scope (Epistemic Object Domain) concerns the defined epistemic domain within which reasoning and action are valid. This axis fails when systems operate outside the authorised domain, apply reasoning to out-of-scope objects, or silently expand the domain of Inference.
Sc - Social Coordination	Social Coordination concerns how multiple humans align, contribute, and reason together through the system. This axis concerns failures arising when individuals implicitly treat platforms/systems as guides for everyday judgement and action. This axis fails when collaboration breaks down due to ambiguity, conflict, loss of shared reference, or inability to merge contributions under governance.
S - State Continuity	State Continuity concerns whether authoritative state persists correctly across time, sessions, and interactions. This axis fails when prior decisions, constraints, or artefacts are lost, forgotten, or inconsistently reintroduced.
St - Stewardship	Stewardship (non-ownership governance) concerns responsibility without ownership. This axis fails when systems or actors behave as if ownership implies authority, or when stewardship duties such as care, preservation, and restraint are neglected despite lack of ownership.
T - Temporal Coherence	Temporal Coherence concerns the preservation of correct temporal relationships between artefacts, constraints, authority, and decisions. This axis fails when versions are confused, rules are applied retroactively, or sequencing and timing semantics are violated.
Thread	A sequence of reasoning that may traverse multiple EO, EA, and UoD; not a unit of governance.
Thought	A unit of human cognition that may be transient or externalised as EO.
Thought Bubble	An informal synonym for a Universe of Discourse.
Tokenisation	Segmentation of text into tokens for computational processing; EO/EA bodies must be tokenisable.
Turn	A single exchange unit within a session that may introduce or modify EO/EA.
U - UI / Mediation	UI / Mediation concerns how interfaces shape, filter, or distort interaction between humans and the system. This axis fails when interface design hides constraints, misrepresents system state, encourages invalid actions, or forces users into integrity-violating behaviour.
Universe of Discourse (UoD)	A bounded aggregation of EO/EA forming a coherent context for reasoning.
Vendor	An organisation or platform provider offering AI systems or tooling; must not imply authority or enclose EO/EA.
Vendor Neutrality	The requirement that CM-2 artefacts do not depend on proprietary schemas or platform behaviour.

categories

See https://publications.arising.com.au/pub/CM-2_Normative_Architecture#categories.