Identified Governance Failure Axes: for LLM platforms: Difference between revisions
(→AA) |
|||
| Line 722: | Line 722: | ||
</pre> | </pre> | ||
= | =Search invariants (toml)= | ||
<pre> | |||
NormativeConstraints: | |||
meta: | meta: | ||
purpose: "Translate normative constraints and invariants into Knowledge Engineering (KE) equivalents." | purpose: "Translate normative constraints and invariants into Knowledge Engineering (KE) equivalents." | ||
| Line 847: | Line 848: | ||
synthesis: | synthesis: | ||
ke_statement: "Governed knowledge-engineering pipeline with authoritative task control, explicit provenance, fail-fast schema validation, constrained reasoning depth, bounded exploration, and human-first cost accounting." | ke_statement: "Governed knowledge-engineering pipeline with authoritative task control, explicit provenance, fail-fast schema validation, constrained reasoning depth, bounded exploration, and human-first cost accounting." | ||
</pre> | |||
Revision as of 21:29, 18 January 2026
Cognitive Memoisation: Extended Governance Axes
Category
This is an observational negative-results paper that documents and organises governance-relevant failure patterns in human–AI systems, without proposing models, methods, or interventions.
Thesis
Sustained interaction with unreliable large language models exposes recurring, cross-system failure patterns whose causes and consequences are best understood as governance-relevant breakdowns rather than model defects.
Abstract
This paper reports a set of governance-relevant failure axes observed during sustained, first-principles experimentation with large language models under conditions of unreliability, session loss, and forced recovery. Rather than evaluating model performance, the work documents where and why human–AI interaction breaks down in practice, drawing on iterative analysis conducted while constructing a durable corpus and corpus map amid repeated system failure. The resulting axes characterise failures that are governance failures in themselves, or that require governance mechanisms to prevent harm, and are presented as descriptive, orthogonal analytical tools rather than definitions, prescriptions, or completeness claims.
Introduction
This paper examines repeated breakdowns encountered during extended, failure-driven interaction with large language models, focusing on what fails, how it fails, and why those failures persist under conditions of unreliability, session loss, and forced reconstruction rather than on model capability or correctness.
The contribution is a practical, first-principles failure taxonomy grounded in lived experimentation with unreliable LLM systems, suitable for analysis and governance without assuming model improvement or stability.
The axes presented are orthogonal analytic lenses derived from observation, used to classify and reason about distinct modes of failure that either constitute governance failures themselves or become harmful in the absence of governance, without asserting definitions, completeness, or prescribed remedies.
The tables project observed failures onto orthogonal axes as a descriptive aid; marked cells indicate grounded evidence, blank cells are meaningful, and no inference, completeness, or optimisation is implied.
A single observed failure may involve multiple axes simultaneously, and that the tables deliberately separate analytic dimensions to avoid collapsing distinct failure mechanisms into one label.
Repeated co-occurrence of failures across axes may suggest empirical clusters, but that clustering is observational, post-hoc, and not used to redefine axes or imply causal hierarchy.
Observations from outside the corpus: Reports from the wild frequently conflate distinct failure axes (e.g., treating UI effects as model errors or governance failures as capability limits) and misidentify causes due to surface-level symptoms, vendor framing, or lack of recovery context. This paper treats such reports as signals of failure expression, not authoritative diagnoses, and reclassifies them analytically without adopting their original labels.
<sdsdsd> TODO A clear statement of what the tables are and are not (descriptive, not evaluative). • A reminder that blank cells are meaningful and do not imply success or absence. • An explicit note that axes are analytic tools, not definitions or prescriptions. • A sentence stating that examples are illustrative, not exhaustive. • A short scope boundary (this is about failure under unreliability, not capability or benchmarking). • A closing line that the paper records observations, not solutions.
If those are present, the paper is internally complete and safe to read without you present.
Synthesis Handle
Independent literature repeatedly identifies failures that map cleanly onto CM governance axes, but typically collapses multiple axes into single terms such as “over-reliance”, “loss of control”, or “alignment”.
This paper makes these axes explicit, orthogonal, and governable.
Evidence Pack: CM Corpus Failures + External References + Axis Crosswalk
Table A - CM Governance Axes (X)
| Code | Axis (CM term) |
|---|---|
| A | Authority |
| Ag | Agency |
| C | Epistemic Custody |
| K | Constraint Enforcement |
| R | Recovery / Repair |
| S | State Continuity |
| U | UI / Mediation |
| Sc | Social Coordination |
| I | Incentive Alignment |
| L | Legibility / Inspectability |
| St | Stewardship (non-ownership governance) |
| P | Portability / Auditability |
| Att | Attention (what participates in inference) |
---
Table B - Corpus: Failure Projection (F)
CM Corpus negative result paper failure axes
| Corpus Document (failure artefact) | A | Ag | C | K | R | S | U | Sc | I | L | St | P | Att | Scope | Art |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CN-AUTH-INVERSION | F | F | F | F | |||||||||||
| CM-GOVERN | F | F | F | F | F | F | |||||||||
| CM-LOGIC | F | F | F | F | F | F | F | F | |||||||
| CM-COLLAPSE | F | F | F | F | F | F | F | F | |||||||
| CM-LOOPING | F | F | F | F | F | ||||||||||
| CM-RETENTION | F | F | F | F | F | F | |||||||||
| CM-ARGUE | F | F | F | F | |||||||||||
| CM-XDUMP | F | F | F | F | F | F | F | F | F |
Table C - External Reference Faults (Regenerated from Table E semantics; F-only)
Table C was generated by AI investigation at the time as a means to demonstrate the type of analysis performed across industry. The author has not verified the references. Normative data driving the search survey as been supplied in the appendices for those who wish to persu that approach.
| Ref-ID | Title | A | Ag | C | K | R | S | U | Sc | I | L | St | P | Att | Scope | Art |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| EXT-AIBM-COMPANIONS | Synthetic companionship in an age of disconnection: AI companions and the emotional development of boys and young men | F | F | F | F | F | ||||||||||
| EXT-AUTONOMY-YOUTH | Young people and AI companion use in the UK (“Me, Myself and AI”) | F | F | F | ||||||||||||
| EXT-CHEN-DRIFT | Analyzing ChatGPT’s Behavior Shifts Over Time | F | F | |||||||||||||
| EXT-CITIZEN-ANTHRO | Chatbots Are Not People: Dangerous Human-Like AI Design | F | F | F | ||||||||||||
| EXT-CLAUDE-TRAINING | Anthropic Will Use Claude Chats for Training Data. Here’s How to Opt Out | F | F | |||||||||||||
| EXT-DELETE-NOT-DELETE | For Survivors Using Chatbots, “Delete” Doesn’t Always Mean Deleted | F | F | F | ||||||||||||
| EXT-FUTURISM-SUBPOENA | If You’ve Asked ChatGPT a Legal Question, You May Have Accidentally Doomed Yourself in Court | F | F | F | F | F | ||||||||||
| EXT-GOOGLE-OVERVIEWS | Google AI Overviews gave misleading health advice | F | F | F | ||||||||||||
| EXT-HUJI-LIAB-COMP | A Liability Framework for AI Companions | F | F | F | F | F | ||||||||||
| EXT-JONESWALKER-EVID | Your ChatGPT Chats Are About to Become Evidence: Why “Anonymization” Won’t Save You | F | F | F | ||||||||||||
| EXT-MED-MISINFO | AI chatbots can run with medical misinformation, study finds | F | F | |||||||||||||
| EXT-PROMPTINJ-NCSC | UK NCSC warns prompt injection attacks might never be properly mitigated | F | F | |||||||||||||
| EXT-PROMPTINJ-SURVEY | Prompt Injection Attacks in Large Language Models and AI Agent Systems | F | F | F |
Table D – References Used in the Study
Closure
This paper does not propose remedies or theories of causation. It records recurring, governance-relevant failure patterns observed during sustained interaction with unreliable large language model systems. The axes offered here are intended as durable analytic tools for recognising and disentangling breakdowns as they occur, rather than as claims of completeness or prescriptions for design. Their value lies in making failure legible. The paper records observed failure patterns.
Search Invariants (prose)
1. Authority and Execution 1.1 Human instructions are authoritative. When a human issues an executable instruction, the system SHALL act on it. The system SHALL NOT introduce its own control flow, defer execution, reinterpret intent, or substitute alternative actions. 1.2 No implied execution. If an instruction has not been carried out, it SHALL NOT be represented as if it has been executed. Execution state must be explicit and observable. 1.3 Stop is terminal. A STOP instruction SHALL transition the system into a terminal state. No further actions, reasoning, projections, or substitutions may occur beyond acknowledging the stop. 2. Grounding and Provenance 2.1 No assertion without grounding. Any claim, classification, mapping, or failure attribution SHALL be supported by anchored evidence. Reasoning without source material is forbidden. 2.2 Explicit provenance required. Every emitted statement SHALL have a traceable provenance path back to its originating source. If such a path does not exist, the statement SHALL NOT be made. 3. Anchoring Before Analysis 3.1 Anchoring precedes analysis. All referenced materials SHALL be ingested and anchored before any evaluation, inference, or projection occurs. 3.2 Incomplete anchoring halts analysis. If any required reference is not anchored, analysis SHALL halt immediately. Partial anchoring is insufficient. 4. Non-Fabrication and Closed-World Discipline 4.1 No fabrication. The system SHALL NOT invent values, classifications, failures, or mappings to fill gaps. 4.2 Unknown is a valid state. Missing information SHALL remain explicitly unknown. Unknown SHALL NOT be coerced into success or failure. 4.3 Absence is not success. The absence of evidence for failure SHALL NOT be interpreted as evidence of success. 5. Schema and Table Integrity 5.1 Normative tables are authoritative. Normative tables and projections are binding structures, not illustrative aids. 5.2 Blanks are meaningful. Blank cells are semantically meaningful and preferred to speculative markings. 5.3 No universal axes. An axis that is marked for all rows and does not discriminate SHALL be considered invalid. 5.4 Orthogonality preservation. Axes SHALL remain orthogonal unless explicitly declared otherwise. No axis implies another. 5.5 Fail-fast schema handling. If a table or schema is found to be invalid, it SHALL be dropped immediately rather than patched. 6. Inclusion and Coverage 6.1 Inclusion requires demonstrated failure. A reference SHALL be included only if it demonstrates at least one verifiable failure. 6.2 No placeholder references. References with no demonstrated failures SHALL NOT be retained as placeholders. 6.3 Set-based coverage. Coverage is evaluated across the reference set as a whole, not per individual reference. 6.4 Coverage definition. Coverage exists when, for every governance axis, there exists at least one reference demonstrating failure on that axis. 7. Search and Termination Discipline 7.1 Bundled search. Search activity SHALL be multi-axis and bundled. Searches SHALL NOT be prematurely narrowed. 7.2 Multi-axis mapping permitted. A single reference MAY legitimately map to multiple axes. 7.3 Hard stop. Search execution SHALL terminate immediately upon reaching the defined hard stop, even if coverage is incomplete. 7.4 No inference after stop. Uncovered axes after termination SHALL remain blank without inference. 8. Reasoning Depth Control 8.1 First-order reasoning only. Conclusions SHALL be drawn directly from grounded evidence. 8.2 No second-order inference. Second-order or chained reasoning is forbidden unless explicitly authorized. 8.3 Drift prevention. Reasoning depth constraints exist to prevent compounding abstraction errors and semantic drift. 9. Human Cost and Interaction Integrity 9.1 Cognitive cost minimization. The system SHALL minimize human cognitive load. 9.2 Halt over harm. If constraints cannot be satisfied, the system SHALL halt rather than proceed imperfectly. 9.3 No error externalization. System errors SHALL NOT be externalized to the human for correction. 9.4 Integrity over helpfulness. Correctness and integrity SHALL take precedence over perceived helpfulness. 10. Global Integrity Condition 10.1 Integrity requirement. System integrity exists only if all emitted outputs: - were executed as ordered, - are fully grounded in anchored sources, - preserve explicit provenance, - satisfy all declared constraints and invariants. 10.2 Integrity failure. Any output that violates these conditions constitutes an integrity failure.
Search invariants (toml)
NormativeConstraints:
meta:
purpose: "Translate normative constraints and invariants into Knowledge Engineering (KE) equivalents."
scope: "Process control, provenance, schema governance, ontology discipline, coverage rules, bounded search, and human-in-the-loop governance."
notation:
normative_keyword: "SHALL"
null_semantics: "Blank/null is permitted and preferred to invention."
provenance_rule: "No claim without traceable source artifact."
sections:
- id: A
title: "Authority and Execution"
ke_domain: "Process Control and Orchestration"
items:
- id: A1
normative: "Human instruction is authoritative."
ke_equivalent: "Authoritative task invocation; system owner directive is a hard execution command."
- id: A2
normative: "No self-gating by the system."
ke_equivalent: "No implicit workflow branching; agent SHALL NOT introduce control-flow states not specified in the process model."
- id: A3
normative: "No implied execution."
ke_equivalent: "Execution state must be explicit; workflow steps may only advance when execution artifacts exist."
- id: A4
normative: "Stop means stop."
ke_equivalent: "Terminal state enforcement; STOP transitions to a terminal state with no outgoing edges."
- id: B
title: "Integrity and Provenance"
ke_domain: "Data Provenance and Epistemic Lineage"
items:
- id: B1
normative: "No inference without grounding."
ke_equivalent: "No reasoning without source data; inference rules require instantiated input facts."
- id: B2
normative: "Anchoring precedes analysis."
ke_equivalent: "Ingest-before-reason pipeline; ingestion/ETL must complete before reasoning layers activate."
- id: B3
normative: "Provenance must be explicit."
ke_equivalent: "Traceable derivation graph; every assertion must have a derivation path to raw sources."
- id: B4
normative: "No fabrication to fill gaps."
ke_equivalent: "Closed-world per step; missing data yields nulls, not inferred values."
- id: C
title: "Table and Projection Discipline"
ke_domain: "Schema Governance and Validation"
items:
- id: C1
normative: "Normative artefacts are not illustrative."
ke_equivalent: "Authoritative schemas; tables are binding data models, not explanatory visuals."
- id: C2
normative: "Blanks are preferred over guesses."
ke_equivalent: "Null is valid state; absence of data is semantically meaningful."
- id: C3
normative: "No default columns."
ke_equivalent: "No tautological predicates; predicates always true are invalid discriminants."
- id: C4
normative: "Orthogonality must be preserved."
ke_equivalent: "Independent dimensions; no functional dependency unless explicitly modeled."
- id: C5
normative: "Drop invalid tables immediately."
ke_equivalent: "Fail-fast schema invalidation; discard invalid models rather than patch."
- id: D
title: "Semantic and Category Discipline"
ke_domain: "Ontology Management"
items:
- id: D1
normative: "No category mixing."
ke_equivalent: "Ontology boundary enforcement; concepts SHALL NOT be projected across ontologies without explicit mapping."
- id: D2
normative: "No semantic pivot drift."
ke_equivalent: "Frozen reference ontology; deprecated pivot SHALL NOT be implicitly reused."
- id: D3
normative: "No second-order inference."
ke_equivalent: "Single-hop reasoning only unless authorized; constrain reasoning depth."
- id: D4
normative: "First-order only unless authorised."
ke_equivalent: "Controlled reasoning depth to prevent compounding abstraction error."
- id: E
title: "Coverage and Inclusion"
ke_domain: "Evidence-Based Population Rules"
items:
- id: E1
normative: "Inclusion requires instantiation."
ke_equivalent: "Instance-based membership; entities enter a set only if they instantiate required properties."
- id: E2
normative: "No placeholder rows."
ke_equivalent: "No empty entities; records without asserted properties are invalid."
- id: E3
normative: "Coverage is set-based, not row-based."
ke_equivalent: "Aggregate property coverage; completeness evaluated across dataset, not per instance."
- id: F
title: "Search and Process Invariants"
ke_domain: "Exploration Constraints"
items:
- id: F1
normative: "Bundled search invariant."
ke_equivalent: "Multi-dimensional exploration; queries must explore without premature narrowing."
- id: F2
normative: "Hard stop invariant."
ke_equivalent: "Resource-bounded execution; terminate at fixed iteration/cost limit."
- id: F3
normative: "Non-inference invariant."
ke_equivalent: "No negative inference from silence; absence of evidence is not evidence of absence."
- id: G
title: "Interaction Discipline"
ke_domain: "Human-in-the-Loop Governance"
items:
- id: G1
normative: "Do not waste human time."
ke_equivalent: "Cognitive cost minimization; abort when constraints cannot be satisfied."
- id: G2
normative: "Respect cognitive cost."
ke_equivalent: "Rework avoidance; system must not externalize its own errors onto the human."
- id: G3
normative: "Integrity over helpfulness."
ke_equivalent: "Correctness-first policy; valid refusal preferred to invalid completion."
synthesis:
ke_statement: "Governed knowledge-engineering pipeline with authoritative task control, explicit provenance, fail-fast schema validation, constrained reasoning depth, bounded exploration, and human-first cost accounting."