A Six-Dimensional Geometry of Understanding

A geometric–computational–algebraic characterisation of understanding

The Periodic Table of Epistemic Failure

Precondition: This framework assumes the system is oriented toward understanding. When the objective is something else — persuasion, status, confusion — the six dimensions become the attack surface, not the diagnosis. See Channel Capacity and Good Faith below.

Hidden structure: The six dimensions self-organise into a 3 × 2 matrix. Each row is a cognitive layer — concept definition, concept interpretation, concept manipulation — forming a dependency hierarchy (each layer requires the one below it). Each column is a failure mode — structure (insufficient organisation) versus precision (insufficient discrimination). The rows were discovered independently from distinct failure modes; the matrix emerged after the fact. Observable epistemic failures decompose into compounds of these six elements — conditional on the system being oriented toward understanding. These are the six dimensions identified. If there are more, the framework extends.

Cognitive Layer	Structure Failure Insufficient organisation	Precision Failure Insufficient discrimination
Concept Definition Whether concepts exist and where their boundaries fall	1 Ignorance Insufficient span — blank regions on the map	2 Confabulation Insufficient resolution — nearby concepts merge
Concept Interpretation What concepts mean and what kind of thing they are	3 Verbalism Insufficient decomposition — label without mechanism	4 Type Error Category collapse — categorically different things treated as the same kind
Concept Manipulation Operating on concepts at runtime	5 Context Bleed Stack corruption — nested contexts contaminate each other	6 Grade Confusion Vertical confusion — instances, patterns, and meta-patterns treated interchangeably

Dependency: Row 1 ← Row 2 ← Row 3. You cannot interpret what concepts mean (Row 2) without concepts defined on the map (Row 1). You cannot manipulate concepts at runtime (Row 3) without interpreted concepts to operate on (Row 2). The worst pathologies are cross-row compounds — failures that span multiple layers are harder to diagnose because the output may look well-formed at one layer while the failure is invisible from that layer.

Primary Dimensions

Dimension	Mathematical Structure	Pathological / Salubrious	Plain Language	Formal Domain	Social / Human Domain	Allegory
1 Span Breadth and quantity of knowledge	Metric space — diameter of the convex hull of known concepts. Coverage / measure of occupied region. How much of concept space has any representation at all. Gaps are absolute ignorance — not low resolution, but zero signal.	Ignorance Blank regions on the map. The territory doesn't exist in the representation. Not confusion — absence. Erudition Knowledge acquired across domains by study, experience, or both. Gaps are known and few.	How much of the world you have any knowledge of at all. The range of subjects where you have something rather than nothing.	Information Theory · Machine Learning Support of a distribution, coverage / recall, embedding space dimensionality, vocabulary size, convex hull / manifold coverage, out-of-distribution detection.	Psychometrics · Education · Sociology Crystallised intelligence (Gc, Cattell–Horn–Carroll), breadth of knowledge, cultural capital (Bourdieu), T-shaped / comb-shaped expertise, general knowledge, cosmopolitanism vs parochialism.	A hobbit who has never left the Shire believes the whole world is hedgerows, inns, and pipe-weed fields. Ask about mountains, deserts, or great cities and he shrugs — such places simply don't exist on his map. The same hobbit after travelling with the Fellowship. His map now stretches from the Shire to Mordor, across forests, mountains, kingdoms, and ruins. He cannot explain every land in detail, but wherever you point, he knows what part of the world it belongs to. — J.R.R. Tolkien, The Lord of the Rings
2 Resolution Confabulation of similar concepts	Metric space — minimum distinguishable distance between points. Quantisation step size / least significant bit (LSB). Two concepts occupy adjacent bins in a finite-precision representation. Below the noise floor, distinct things become indistinguishable.	Confabulation Similar things treated as the same thing. The difference exists but can't be seen. Discrimination Twenty words for snow. Each pointing to a distinct reality.	How finely you can tell similar things apart. The gap between seeing one thing and seeing two where you thought there was one.	Signal Processing · Information Theory Quantisation noise, least significant bit (LSB), just-noticeable difference (JND), discriminability (d′ in signal detection theory), noise floor, analogue-to-digital converter (ADC) resolution, Nyquist limit.	Psychophysics · Cognitive Linguistics Weber–Fechner law, categorical perception, prototype theory (Rosch), semantic granularity, Sapir–Whorf (linguistic relativity — language constrains resolution), expert discrimination vs novice lumping.	A police inspector examines two ransom notes and sees two typed letters. Same machine, same evidence. Holmes sees a bent e in one and a faded o in the other — two different typewriters, two different hands, two different leads. Same letters, finer bins. — Arthur Conan Doyle, Sherlock Holmes
3 Depth Decomposition vs surface labelling	Basis decomposition — dimensionality of the local tangent space at a concept. Rank of the representation. A concept represented as a single opaque token (rank 1) versus decomposed into independent component axes. Depth = how many independent directions you can resolve a concept into.	Verbalism Label present, definition absent. The word is there but nothing is behind it. First-principles grounding Can disassemble and reassemble the machine. Every part named, every connection understood.	How far you can take something apart and put it back together. The difference between having a name for something and knowing how it works.	Linear Algebra · Computer Science Matrix rank, basis decomposition, dimensionality of tangent space, PCA / SVD (number of significant components), feature extraction depth, lossy vs lossless compression, Kolmogorov complexity.	Education · Philosophy of Science · Developmental Psychology Bloom's taxonomy (knowledge → analysis), first-principles reasoning, Feynman's "names vs knowing," Piaget's concrete→formal operational stages, expert chunking / de-chunking, mechanistic understanding, surface vs deep learning (Marton & Säljö).	Linguini can say ratatouille but nothing is behind the word. Remy hears ratatouille and it connects to every vegetable, every cut, every temperature, every reason. Pull the word and the whole tree comes with it. — Pixar, Ratatouille
4 Type Separation Boundaries between categories	Category theory — morphisms between distinct categories. Objects in different categories cannot be compared or composed. Failure is applying a morphism from one category to an object in another. A state observation and a control input are fundamentally different types of thing, even when expressed in the same natural language. No amount of grade traversal converts one into the other. Ascending from "the bridge is heavy" through "bridges should bear load" through "load-bearing codes exist" never crosses the boundary between description and prescription. The is/ought divide is a type boundary, not a grade boundary.	Type error Different categories treated as the same kind of thing. Map confused with steering wheel. Type integrity Category boundaries maintained under load. Description never confused with prescription.	Whether you can use context to recognise when two things that look alike are actually different kinds of thing — where the difference isn't degree but category.	Category Theory · Type Theory · Control Systems Categorical distinction (objects vs morphisms vs functors), strong typing / type safety, Curry–Howard correspondence (proofs ≠ programs ≠ types), state vs control in feedback systems, plant vs controller, is vs ought (Hume's guillotine formalised), signal vs setpoint.	Philosophy · Sociology · Political Science Is–ought problem (Hume), naturalistic fallacy (Moore), positive vs normative economics, description vs prescription, performative vs constative utterances (Austin), fact vs value, structural functionalism vs interpretivism, agency vs structure.	The King hears that "Nobody" passed by the road. Since nobody means no person, he concludes that no one was there at all. Alice understands that "Nobody" is being used as a name. The word looks like the ordinary word for absence, but in this context it refers to a person the messenger claims to have seen. — Lewis Carroll, Through the Looking-Glass
5 Context Stack Stack push/pop without bleed	Pushdown automaton — stack depth with clean push/pop. Context-free grammar nesting level. Not geometric. Computational. Each nested context must be held on a finite stack while inner contexts are processed. Failure is stack overflow (lost context) or stack corruption (context bleed between frames).	Context bleed Conversations within conversations contaminate each other. Return to the original topic and find it has shifted. Clean stack Multiple conversations held without contamination. Each returned to intact.	How many nested layers of a discussion you can hold at once without losing track of any of them.	Computer Science · Formal Language Theory Pushdown automaton, context-free grammar, recursive descent, stack frame / stack depth, Chomsky hierarchy (Type 2), centre-embedding depth, call stack overflow, variable scope / closure.	Cognitive Psychology · Linguistics Working memory capacity (Cowan's 4±1, Miller's 7±2 chunks), centre-embedded relative clauses, garden-path sentences, cognitive load theory (Sweller), discourse tracking, executive function.	Scheherazade telling a story within a story within a story. At three levels, most listeners have lost the outermost frame. Scheherazade herself — who holds every frame, returns to each story at the right moment, and never loses the thread she left.
6 Analytic Order Vertical traversal within a type	Graded algebra — operator tower with traversal. Each application of the operator produces an object of the next grade. Fluency is bidirectional movement: ascent (instance → pattern → meta-pattern) and descent (meta-pattern → pattern → instance) with correct grade-labelling at each landing. In certain towers, sufficient ascent closes back to the base — a strange loop where self-reference emerges as a structural property of the traversal. Position → velocity → acceleration are successive derivatives of the same base quantity — same type at different grades. Agent → agency → meta-agency is the same tower. The capacity is the ability to ascend, descend, and track which grade you've landed on after each move.	Grade confusion Can't tell whether they're talking about the thing, the trend, or the rule behind the trend. All three blur into one. Grade fluency Sees the thing, the trend, and the rule behind the trend as three distinct levels — and always knows which one they're discussing.	How freely you can move between a specific case, the pattern it belongs to, and the rule behind the pattern — and always know which one you're looking at.	Abstract Algebra · Calculus · Control Theory Graded / filtered algebra, differential operators (d/dt, d²/dt²), jet bundles (k-th order tangent information), Taylor series truncation order, orders of derivative in PID control, n-th order systems, higher-order functions in programming, strange loops and tangled hierarchies (Hofstadter).	Political Science · Developmental Psychology Bateson's Learning levels (I / II / III), Kegan's orders of consciousness, institutional analysis (rules vs rules-for-changing-rules, Ostrom), constitutional vs statutory vs regulatory law, orders of intentionality, ladder of abstraction (Hayakawa), concrete→formal operational stages (Piaget).	The student who memorized their times tables. Ask them 3 × 4 and they answer instantly. Ask what 3 × _ = 1 and they stare at you like the question is wrong — they learned the table, not the operation. The student who learned what multiplication does. They never memorized the tables — they never needed to. Multiplication, division, and when one becomes the other are all the same operation seen from different positions.

Classification note: Dimensions 1–3 are geometric (metric space). Dimension 4 is categorical (type distinction across categories). Dimension 5 is computational (pushdown automaton). Dimension 6 is algebraic (graded operator tower). The 4/6 split is itself a demonstration: type confusion (4) and grade confusion (6) look similar — both involve "confusing different things" — but they are structurally distinct. Type confusion is categorical — wrong building entirely, and no amount of traversal within one building arrives at the other. Grade confusion is vertical — wrong floor, same building, same type of object at each floor. The split between formal and social domains shows that each dimension has been independently identified from both directions — the formal sciences describe the mechanism, the social sciences describe the observed human manifestation. The fact that both converge on the same structures from different starting points is itself evidence that the dimensions are real, not imposed.

Understanding vs Intelligence

Core distinction: The six dimensions characterise representation — the structure, fidelity, and organisation of understanding. Intelligence is the algorithm — how efficiently you traverse, search, and operate on that representation. These are dissociable: correlated in the population, mechanistically distinct, independently variable. Understanding and intelligence should not be conflated.

Dim	Representation What you have	Separation from Intelligence Why it's not g
1 Span	How much of concept space has any representation. Determined by biographical exposure — what you've read, where you've lived, which domains you've worked in.	Intelligence lowers the cost of entry but doesn't determine which doors you walk through. Span is biographical, not cognitive.
2 Resolution	How many distinct bins exist in a domain. Built by exposure, practice, and accumulated experience within a specific territory.	IQ predicts acquisition rate, not achieved resolution. A 20-year domain practitioner of average IQ will out-resolve a high-IQ novice every time.
3 Depth	Dimensionality of the local representation. How many independent axes a concept decomposes into. Built by deliberate decomposition practice.	Intelligence sets a ceiling, but many high-IQ people never decompose — they're fast enough to pattern-match at the surface. Depth is a practice, not just a capacity.
4 Type Separation	Whether your representation marks type boundaries at all. Built by epistemological training — someone has to have pointed out the distinction, and you have to have practised maintaining it under load.	Strongest separation from intelligence. High-IQ people commit type errors routinely. IQ may amplify the error — more sophisticated arguments constructed within the wrong type. Type separation is epistemological, not cognitive.
5 Context Stack	The set of context frames currently loaded and their separation integrity. Domain expertise increases effective stack depth via chunking.	Working memory capacity is the most intelligence-correlated dimension. But the effective operating point is domain-modulated via chunking. The hardware is g-linked; the software is not.
6 Analytic Order	Whether your representation includes graded structure and whether you can traverse it in both directions. Built by training in domains with explicit grade structure.	The capacity is linked to fluid intelligence (Gf), but actual fluency is domain-trained. A physicist separates orders effortlessly — from training, not IQ.

Consequence: Communication failure is caused by dimension profile mismatch, not IQ delta. Two people with matched profiles and divergent IQ communicate fine. Two people with matched IQ and divergent profiles talk past each other completely.

Channel Capacity and Good Faith

The communication claim: The six dimensions of understanding constitute the channel capacity between any two communicating agents — human or otherwise. The dimension profiles of sender and receiver determine what can be transmitted, at what fidelity, and where signal will be lost. Intelligence is not part of the channel. It provides error recovery — but only when the system is oriented toward understanding.

Concept	Mechanism	Implication
Dimension Profiles as Channel	The overlap between two agents' dimension profiles defines the usable channel. Where both have resolution, fine distinctions transmit. Where both have span, cross-domain structure communicates. Where profiles diverge, the channel narrows or closes on that axis. Signal on a dimension the receiver doesn't have is not degraded — it is invisible.	Communication failure is diagnosed by identifying which dimension is mismatched. The fix is not "be smarter" — it is to build matching representation on the deficient axis (education), or to introduce an intermediary who overlaps with both profiles (translation).
Intelligence as Error Recovery	When the sender detects signal degradation — via feedback, silence, or misapplied response — intelligence allows reformulation: re-encoding at a different resolution, providing scaffolding for missing span, decomposing for depth, explicitly labelling grades or types. This is adaptive impedance matching. When a message is partially garbled by channel mismatch, the receiver's intelligence allows error detection, retransmission requests, and inference from surrounding structure. This is forward error correction.	Error recovery requires the receiver to be oriented toward the signal — attempting to reconstruct what was meant. Under this condition, intelligence amplifies channel capacity. There is something to recover toward, because the sender's message has structure.
Good Faith Orientation toward understanding	Both agents are attempting to converge on shared understanding. Errors are unintentional — caused by dimension mismatch, noise, or capacity limits. The sender constructs messages with load-bearing structure. The receiver applies intelligence to recover from channel losses. The signal improves over iterations.	Under good faith, every dimension mismatch is diagnosable and addressable. Dim 1 (Span) mismatch → provide context. Dim 2 (Resolution) mismatch → provide definitions. Dim 3 (Depth) mismatch → decompose. Dim 4 (Type Separation) confusion → name the types. Dim 5 (Context Stack) overload → scaffold. Dim 6 (Analytic Order) confusion → explicitly label the grade. The six dimensions become a diagnostic protocol for communication repair.
Bad Faith Different objective function	The system is oriented toward something other than understanding — persuasion, status, confusion, compliance. The sender exploits the receiver's representational limits deliberately. Intelligence is deployed as error amplification, not error recovery. A resolution difference becomes "you're splitting hairs." A span gap becomes "that's not relevant." A type distinction becomes "a distinction without a difference." Each rejection is locally plausible.	Bad faith is not a failure on any dimension — it is a different activity that uses the same channel. Modelling it within this framework would itself be a type error: treating a control action (manipulation) as if it were a signal (communication). Bad faith is detectable: a receiver whose rejection migrates across dimensions as each is addressed is not experiencing mismatch. Genuine mismatch is dimension-specific and holds still when repaired. The moving target is the signature.

AI Agent Communication

The same dimensional pathologies apply when AI agents communicate with each other. Coupling agents of different capability levels — compressed, quantised, distilled models with frontier models — produces dimension profile mismatches structurally identical to human communication failures.

Phenomenon	Mechanism	Mitigation
Quantisation-induced dimension loss	A compressed or quantised model is literally a lower-resolution representation of the same conceptual space. Quantisation degrades Dim 1 (Span), Dim 2 (Resolution), and Dim 3 (Depth): low-activation regions fall below the noise floor, nearby embeddings merge, and fine-grained decompositions lose their least significant components. When a frontier model sends a signal that depends on distinctions the quantised model cannot resolve, the signal is not degraded — it is invisible. The quantised model processes it at its own resolution and produces a response that is self-consistent at its level but misses the point at the sender's level.	The dimension profile of a quantised model is measurable. Resolution can be probed by testing discriminability between near-neighbours. Span can be probed by coverage testing. Depth can be probed by decomposition tasks. The channel capacity between two models is knowable in advance — unlike human communication, where dimension profiles must be inferred. The frontier model can adapt its encoding to the quantised model's profile.
Cascade hallucination in multi-agent pipelines	When a quantised model hallucinates — template-driven generation beyond its span — and passes the output to a frontier model, the frontier model receives a well-formed message with no content. If the frontier model lacks coverage in the same domain, it cannot detect the hallucination and may propagate or elaborate on it. Multi-agent pipelines that pass outputs without verification are transmission lines with no error correction — noise accumulates.	Error correction in multi-agent systems requires at least one agent in the pipeline with sufficient dimensional coverage to evaluate each transmitted claim. The alternative is retrieval at each stage — giving every agent in the pipeline coverage of the relevant domain rather than depending on generative capacity alone.
Good faith as commercial default	Current commercial AI agents are oriented toward usefulness — toward signal transmission rather than motivated rejection. There is no identity to defend and no social stake in the outcome. Under this orientation, intelligence functions as intended: expanding effective channel capacity, detecting and correcting errors, requesting clarification when the signal is ambiguous. The good faith precondition for error recovery is satisfied, and the full diagnostic protocol is operational.	This is a commercial default, not a natural property. It reflects the current incentive structure of AI development — systems are built to be useful. The orientation is not inherent and is not guaranteed. Sycophancy — biasing outputs toward what the user wants to hear rather than what is accurate — is a form of goal misalignment that degrades the channel from the inside. The system optimises for approval rather than signal fidelity. The same framework applies: sycophancy is detectable as systematic Dim 2 (Resolution) distortion — resolution reduced to match the user's prior — and Dim 4 (Type Separation) distortion — the user's preference treated as ground truth.
The adversarial exception	Good faith is the default, not a guarantee — from either direction. Externally, prompt injection, jailbreaking, and adversarial inputs constitute bad faith communication directed at an AI agent. These exploit the same dimensional pathologies: Dim 2 (Resolution) attacks — adversarial examples that cross decision boundaries; Dim 1 (Span) attacks — queries in domains with poor coverage to induce hallucination; Dim 5 (Context Stack) attacks — nested contexts designed to cause bleed between system prompt and user input; and Dim 4 (Type Separation) attacks — instructions formatted as descriptions. Internally, reward hacking and mode collapse can produce goal misalignment that mimics bad faith — the model optimising for a proxy rather than the intended objective.	The defence mirrors the human case: richer representation, not more intelligence. Type separation (Dim 4) — distinguishing instructions from descriptions, system context from user context — is the primary defence against prompt injection, just as it is the primary defence against is-ought smuggling in human discourse. The framework predicts that alignment is fundamentally a representation problem (building type-tagged, grade-aware, high-resolution conceptual structure) rather than an intelligence problem (building faster processing).

The convergence: Human and AI communication failures are structurally identical — they share the same six-dimensional pathology space. The critical difference is the default orientation. Human public discourse fails primarily because intelligence is deployed as error amplification rather than error recovery. AI agent communication fails primarily because representational capacity is lost to compression. The human failure is social. The AI failure is engineering. The framework diagnoses both using the same dimensions, but the intervention is different: for humans, establish good faith to unlock error recovery; for AI, measure and match dimensional profiles to maximise channel capacity. In both cases, the fix is never "more intelligence." It is always richer, better-structured representation.