Technical research collection

The Machine Intelligence Canon

A dependency-ordered route connecting mathematical foundations, learning systems, representation, reasoning, agents, production controls and human judgement. Each paper combines conceptual argument, technical mechanism, visual explanation and executable work where code exposes the mechanism.

FoundationsFormal tools, representation, optimisation and learning systems
Intelligent systemsReasoning, search, memory, agents and compound architectures
Operational judgementEvaluation, governance, control and human authority

How to read the collection

Begin with the opening definition of intelligence, then follow the series in order when later concepts depend on earlier mechanisms. Each paper carries the argument through explanation, visual structure, worked application and executable artefact.

Series I: Intelligence, computation and representation

Research sequence
  1. 001What Would Count as Intelligence?
  2. 002Problems, Tasks and Environments
  3. 003Computation as State Transformation
  4. 004Algorithms, Heuristics and Learned Procedures
  5. 005Representation Before Intelligence
  6. 006Feedback, Control and the Cybernetic Loop
  7. 007Learning, Reasoning and Search Are Not Synonyms
  8. 008From Programme to Agent

Series II: Search, planning, logic and reinforcement

Research sequence
  1. 009Search as Structured Exploration
  2. 010Heuristics, A* and the Price of Guidance
  3. 011Adversarial Search and Strategic Intelligence
  4. 012Planning Before Action
  5. 013Knowledge Representation and Logic
  6. 014Reinforcement Learning and Sequential Choice
  7. 015Cognitive Architectures Before Foundation Models
  8. 016The Bitter Lesson and the Neuro-Symbolic Reply

Series III: Mathematical language for learning systems

Research sequence
  1. 017Functions, Composition and Computation Graphs
  2. 018Vectors as Directions, Not Lists
  3. 019Matrices as Machines
  4. 020Eigenstructure and Persistent Directions
  5. 021Low Rank, Singular Values and Compression
  6. 022Derivatives, Gradients and Credit Assignment
  7. 023Optimisation Landscapes Without the Mountain Myth
  8. 024Graphs, Networks and Message Passing

Series IV: Probability, information, causality and decisions

Research sequence
  1. 025Probability as Disciplined Uncertainty
  2. 026Bayes' Rule as Evidence Accounting
  3. 027Distributions and the Cost of a Point Answer
  4. 028Entropy, Information and Compression
  5. 029Variational Inference and Useful Approximation
  6. 030Causation Is More Than Prediction
  7. 031Decision Theory Under Irreversibility
  8. 032Abstention as a Positive Capability

Series V: Classical machine learning and scientific evaluation

Research sequence
  1. 033Linear Models and the Discipline of a Baseline
  2. 034Trees, Forests and Boosting
  3. 035Kernels and Similarity as Computation
  4. 036Clustering and the Politics of Distance
  5. 037Dimensionality Reduction and Manifold Claims
  6. 038Time Series, Drift and Anomaly
  7. 039Bias, Variance and the Geometry of Generalisation
  8. 040Evaluation Is a Measurement System

Series VI: Neural networks and deep learning

Research sequence
  1. 041The Artificial Neuron and What It Leaves Out
  2. 042Backpropagation as Distributed Credit Assignment
  3. 043Activation, Normalisation and Residual Flow
  4. 044Regularisation and Stable Explanations
  5. 045Convolution, Invariance and Equivariance
  6. 046Recurrence, Gating and the Vanishing Past
  7. 047Representation Geometry in Deep Networks
  8. 048How to Run a Deep-Learning Experiment

Series VII: Representation learning and generative models

Research sequence
  1. 049Autoregression and the Power of the Next Step
  2. 050Autoencoders and the Bottleneck Bargain
  3. 051Variational Autoencoders and Structured Latent Space
  4. 052Adversarial Generation and the Moving Target
  5. 053Diffusion and Learning to Reverse Noise
  6. 054Flow Matching and Continuous Generative Transport
  7. 055Contrastive and Multimodal Representation Learning
  8. 056Evaluating Generators Beyond Beauty

Series VIII: Transformer anatomy

Research sequence
  1. 057Attention as Content-Addressed Routing
  2. 058Many Heads, Grouped Queries and Shared Keys
  3. 059Position, Order and Relative Geometry
  4. 060The Residual Stream as a Shared Workspace
  5. 061Feed-Forward Networks, Gating and Sparse Experts
  6. 062Tokenisation and Its Hidden Tax
  7. 063Causal Masks, KV Caches and Generation
  8. 064A Transformer Forward Pass You Can Intervene On

Series IX: Language-model training and post-training

Research sequence
  1. 065Data Mixtures Are Part of the Model
  2. 066Pretraining Objectives and What They Reward
  3. 067Scaling Laws Without Extrapolation Theatre
  4. 068Instruction Tuning as Behavioural Reorientation
  5. 069Preference Learning, RLHF and DPO
  6. 070Reinforcement Learning for Reasoning
  7. 071Distillation, LoRA and the Private Skill Model
  8. 072Training Lineage as a Governed Supply Chain

Series X: Language-model use and component engineering

Research sequence
  1. 073Prompting as Engineering
  2. 074Structured Output and Function Calling
  3. 075Decoding, Sampling and Controlled Generation
  4. 076Context Windows, Position Extension and Context Rot
  5. 077Fine-Tune, Prompt, Retrieve, Calculate or Call a Tool?
  6. 078Model Selection and Routing Without Benchmark Theatre
  7. 079Hallucination Engineering and Claim Support
  8. 080LLM-as-Judge as a Fallible Instrument

Series XI: Efficient inference and frontier architectures

Research sequence
  1. 081The Anatomy of Inference Latency
  2. 082Quantisation and the Shape of Lost Precision
  3. 083Speculative Decoding and Verified Acceleration
  4. 084Inference Cascades and the Economics of Escalation
  5. 085Mixture-of-Experts Routing as a Systems Problem
  6. 086State-Space Models and Selective Memory
  7. 087Hybrid and Diffusion Language Models
  8. 088Private, Edge and Federated Inference

Series XII: Reasoning, search, verification and test-time compute

Research sequence
  1. 089What Is a Reasoning Trace Evidence Of?
  2. 090Decomposition and the Shape of a Solvable Problem
  3. 091Sampling, Self-Consistency and Search Over Answers
  4. 092Outcome Verifiers, Process Verifiers and Their Blind Spots
  5. 093Adaptive Test-Time Compute
  6. 094Solve, Verify or Escalate?
  7. 095Reflection and Self-Correction Under Controlled Evidence
  8. 096The Reasoning Stack: From Proposal to Defensible Answer

Series XIII: Interpretability, metacognition and machine self-knowledge

Research sequence
  1. 097Probing Is Not Yet Explaining
  2. 098Circuits and the Search for Mechanism
  3. 099Superposition, Sparse Features and Causal Inertness
  4. 100Representation Geometry and Belief Subspaces
  5. 101Introspection Without Taking the Report on Faith
  6. 102Metacognition as Calibration, Control and Information Seeking
  7. 103Self-Models and Safe Self-Modification
  8. 104Machine Psychometrics Without Anthropomorphic Leakage

Series XIV: Retrieval, context and knowledge systems

Research sequence
  1. 105Retrieval as Inference Under a Budget
  2. 106Embeddings, Sparse Search and Hybrid Retrieval
  3. 107Reranking and the Last Mile of Evidence Selection
  4. 108RAG, GraphRAG and Relational State
  5. 109The Context Compiler
  6. 110Governed Context Reconstruction
  7. 111Freshness, Supersession and Temporal Entitlement
  8. 112Semantic Caches Without Privacy or Evidence Leakage

Series XV: Agents, planning and tool use

Research sequence
  1. 113A Model Is Not an Agent
  2. 114ReAct and the Observation-Action Loop
  3. 115Plans, Policies and Replanning
  4. 116Tool Use as Learned API Interaction
  5. 117Typed Actions and the Deterministic Authority Kernel
  6. 118Computer Use and the Fragile Visual Interface
  7. 119Code Agents and the Repository as an Environment
  8. 120The Certainty Gradient

Series XVI: Memory, world state and long-horizon execution

Research sequence
  1. 121World State Is Not Memory
  2. 122Episodic, Semantic, Procedural and Working Memory
  3. 123Memory as an Attack Surface
  4. 124Long-Horizon Drift and Markovian Reconstruction
  5. 125Event-Sourced Case State
  6. 126Dependency Graphs, AND/OR Plans and Mission Control
  7. 127Checkpointing, Replay and Recoverable Execution
  8. 128Agent Plus Operational Twin

Series XVII: Multi-agent systems and protocols

Research sequence
  1. 129When Is More Than One Agent Justified?
  2. 130Choosing Topology from Task Structure
  3. 131Decision-Sufficient Communication
  4. 132Debate, Diversity and the Baseline Problem
  5. 133Learning Dynamic and Trust-Aware Topology
  6. 134MCP from Resources to Tool Security
  7. 135A2A Tasks, Artifacts and Inter-Agent Authority
  8. 136The Autonomy Ledger for Agent Teams

Series XVIII: Computer vision, multimodal systems and embodied intelligence

Research sequence
  1. 137Computer Vision from Pixels to World Structure
  2. 138Multimodal Fusion and Unequal Evidence
  3. 139Vision-Language-Action Models
  4. 140Grounding Through Consequence
  5. 141World Models That Survive Intervention
  6. 142Embodiment as Constraint, Sensorium and Skill
  7. 143Simulation, Digital Twins and the Reality Gap
  8. 144Active Inference as a Design Hypothesis

Series XIX: Evaluation, reliability and recovery

Research sequence
  1. 145Benchmark Construct Validity
  2. 146The Evaluation Stack from Component to Outcome
  3. 147Trajectory Assurance Beyond Final Answers
  4. 148Failure Injection as the Primary Evaluation
  5. 149Drift Sentinels for Decisions, Not Just Models
  6. 150Reproducibility and Counterfactual Replay
  7. 151The Unknown Outcome and Postcondition Verification
  8. 152Normal Accidents and Graceful Recovery in Compound AI

Series XX: Security, identity, containment and incident response

Research sequence
  1. 153Prompt Injection Is a Trust-Boundary Failure
  2. 154Threat Modelling the Agent Supply Chain
  3. 155Capabilities, Scoped Identity and Least Authority
  4. 156The Three Principals and the Confused Deputy
  5. 157Memory Poisoning, PII and Context Leakage
  6. 158Guardrails and Model-Independent Containment
  7. 159Red Teaming and Incident Response for Agentic Systems
  8. 160Causal Attribution and the Five-Chain Ledger

Series XXI: AI governance, model risk and assurance

Research sequence
  1. 161Why Governance Is an Engineering Discipline
  2. 162An Operational AI Risk Taxonomy
  3. 163Regulation as Requirements Engineering
  4. 164Model Risk for Foundation and Compound Systems
  5. 165Fairness Across the Decision Pipeline
  6. 166Explanation for Users, Reviewers and Regulators
  7. 167The Assurance Case for Agentic Systems
  8. 168Decision Rights, Accountability and Claims Discipline

Series XXII: Production AI platforms, operations and inference economics

Research sequence
  1. 169The AI Platform as a Product
  2. 170From MLOps to Compound-AI Operations
  3. 171A Vendor-Neutral Enterprise Agent Fabric
  4. 172Durable Workflows for Stochastic Workers
  5. 173Observability That Explains Decisions
  6. 174Model Release, Canarying and Behavioural Rollback
  7. 175Buying Thought as a Portfolio
  8. 176The Operating Model for Shared AI

Series XXIII: Data, events, legacy integration and software-engineering agents

Research sequence
  1. 177Data Contracts Before Model Contracts
  2. 178Events, CDC and the Reconstruction of Change
  3. 179Knowledge Publication as a Governed Pipeline
  4. 180Legacy Integration Patterns for Agentic Systems
  5. 181The Strangler Pattern for AI-Enabled Change
  6. 182Software-Engineering Agents Beyond the Demo Task
  7. 183Test Migration as a Governed Agent Factory
  8. 184Document Intelligence as Evidence Reconstruction

Series XXIV: Regulated banking intelligence

Research sequence
  1. 185Banking as Information, Time and Risk Transformation
  2. 186The Banking Technology Estate Sets the Automation Frontier
  3. 187The Front-Office Copilot and Governed Adviser
  4. 188Credit Underwriting with Auditable Reasoning
  5. 189KYC and Identity as Evidence-Bound Workflow
  6. 190Financial-Crime Intelligence and the Hard-SAR Constraint
  7. 191Payments, Fraud, Markets and the Irreversibility Gradient
  8. 192Regulatory Obligations, Controls and Executable Workpapers

Series XXV: Human-machine cognition and the research frontier

Research sequence
  1. 193Extended Intelligence and the Human-Machine Unit
  2. 194Cognitive Sovereignty and Epistemic Agency
  3. 195Theory of Mind and Social Modelling in Machines
  4. 196Collective Intelligence Across Humans and Agents
  5. 197Artificial Metacognition Under Irreversible Choice
  6. 198Continual Learning, Self-Improvement and Identity Stability
  7. 199The Road to General Agents Is a Stack of Open Problems
  8. 200The First Defensible Contribution