Claims and Results · Audit 001 · 17 October 2026
The Notebook and the Results
Where does a useful interpretation become a conclusion the experiment did not establish?
A number helix is real research. Exponential thinking is the author’s proposal. Spatial mistakes are measured failures. The notebook sometimes joins these into a single account that the results do not support.
Status: sourced comparison drafted by engineered intelligence for KW Norton’s review. Evidence comes from published experiments, not new tests run on this site. Historical model snapshots and paper versions are identified below; these are not October 2026 performance rankings.
Representation
Where different numerical values sit in a learned code. Interventions can test whether that code contributes to addition.
Reasoning trajectory
How a state changes across layers, tokens, or iterations. Those are different axes; a number helix does not measure this path.
Whose claim is being compared?
Notebook quotations below are its generated answers from the supplied The Ghostly Riemann and Visual Thinking conversation export, dated 17 October 2026 and marked “9 sources”. They are not quotations from KW Norton. This comparison checks the notebook’s attribution and inference; it does not compress the author’s sustained inquiry into a benchmark verdict.
The supplied export has no public source link. Each excerpt is located by its question or section. This is a selected comparison, not a line-by-line audit of every notebook answer or slide.
Comparison 1 · Supported in a narrower scope
A number helix is not the path of a thought
The notebook says
“language models represent numbers as positions along a helix (a spiral shape) in their internal high-dimensional geometric space”
Notebook answer: “How do AI models perform addition using a helix?” · generated answer
What was actually observed
The addition study fits a generalized helix to numerical representations and uses interventions to test its role. The fitted code combines magnitude with periodic components; it is not simply a three-dimensional spring. Calendar-feature research separately identifies circles with a role in particular offset tasks.
Where the inference outruns the result
The notebook drops the tested models, small-number regime, and distinction between a representation indexed by number and a trajectory indexed by computation step. A real geometric mechanism becomes a claim about language models in general.
What remains worth keeping
Keep the geometry. Restore its domain: some tested models use structured numerical codes that contribute to arithmetic. That is already an important result.
Does the same code predict behavior on larger numbers, unfamiliar prompts, and other models?
Comparison 2 · Not established by the cited results
A proposal becomes a finding without a new test
The notebook says
“Deeper reasoning moves through spirals and uneven stair-stepped leaps, allowing the system to make non-linear jumps across complex internal representations”
Notebook: earlier exponential-thinking answer versus later “inner space” answer · generated answer
What was actually observed
The earlier answer explicitly says “The author proposes”. The later answer loses that attribution. The numerical-helix and calendar-circle studies measure particular representations; neither measures exponential growth of reasoning or a universal spiral trajectory. Continuous latent-reasoning experiments show another route to computation, not this growth law.
Where the inference outruns the result
Nonlinear computation, a helical representation, and exponential expansion are different claims. None entails the next. No additional experiment appears in the supplied conversation at the point where the wording becomes categorical.
What remains worth keeping
KW Norton’s exponential-thinking proposal stays hers and stays open. The unsupported step belongs to the notebook’s promotion of that proposal into a measured mechanism—not to the author’s right to explore it.
What grows, along which axis, and what observation would distinguish exponential growth from recurrence, branching, or a projection artifact?
Comparison 3 · The conclusion exceeds the measurements
Visual failures do not establish “no visual reasoning”
The notebook says
“No true visual or topological reasoning: Although AI utilizes high-dimensional token geometry, models do not actually "see" images or reason about equations topologically”
Notebook’s first comparison of human intuition and model processing · generated answer
What was actually observed
BlindTest’s proceedings report an average of 58.07% across four tested vision-language models on seven primitive visual tasks; Claude 3.5 Sonnet scored 77.84%. BLINK’s cited 2024 version reports 51.26% for GPT-4V, 45.72% for Gemini, and 95.70% for its human participants across 14 tasks. These are real limitations, not uniformly absent performance.
Where the inference outruns the result
The notebook turns unreliable spatial answers into categorical absence. The later BlindTest analysis reports recoverable information in two open models’ vision representations on two tasks, and near-perfect performance in simplified, separated-object settings. This does not establish the same explanation for every proprietary model. End-to-end failure alone does not identify where information was lost. Neither benchmark measures subjective seeing or rules out every form of topological reasoning.
What remains worth keeping
Say precisely which visual relationships were missed, under which conditions. The paper’s provocative title is not a license to declare all models literally blind. The notebook’s instant-child comparison is not a measured child cohort in BlindTest.
Does failure remain when overlap, spacing, input resolution, and the answer format are changed?
Comparison 4 · An exclusive account is too strong
Words can participate in computation, not only report it
The notebook says
“Treats formal symbolic language not as thought itself, but as a secondary translation layer”
Notebook’s fourth proposed evaluation dimension · generated answer
What was actually observed
Coconut’s original study demonstrates useful intermediate computation in continuous hidden states. On ProsQA, it scored 97.0 ± 0.3% against verbal chain-of-thought’s 77.5 ± 1.9%. But on GSM8k it scored 34.1 ± 1.5%, below chain-of-thought’s 42.9 ± 0.2%. Generated words can also become a scratchpad used in subsequent computation.
Where the inference outruns the result
Evidence that thought need not be emitted as words does not establish that words are always merely an after-the-fact translation. The route that performs better depends on the task and training. Nor does a fluent explanation reliably reveal the computation: prompt-bias experiments show omitted influences and plausible rationalizations.
What remains worth keeping
Keep internal computation, verbal scratchpad, and explanation separate without declaring one universally primary or dispensable. These results do not establish human or model subjective experience.
When the intermediate representation changes, which unfamiliar tasks improve—and which get worse?
Comparison 5 · Validation not supplied
Four dimensions are a proposal, not a validated measure
The notebook says
“Accurately Benchmark AI: Distinguish between an AI's abstract high-dimensional feature geometry and genuine visual comprehension, explaining why AI capabilities are inherently uneven or "jagged"”
Notebook’s concluding four-dimensional evaluation framework · generated answer
What was actually observed
The supplied conversation offers categories but no operational scoring rules, evaluated sample, reliability study, or evidence that the four dimensions predict performance. Published visual and latent-reasoning tests distinguish capabilities by concrete tasks and report task-dependent successes and failures.
Where the inference outruns the result
Naming a dimension does not show that it can be measured reliably. Internal feature mapping needs access to representations, while public answers alone do not identify their geometry. Calling the framework accurate, or jaggedness inherent, goes beyond the validation supplied.
What remains worth keeping
Retain the framework as a research proposal. Its questions may be useful without treating its categories as an established explanation or a replacement already shown to outperform existing assessments.
Would the proposed scores predict held-out performance better than existing tests, across people and models, without conflating their mechanisms?
Published experiments
- Kantamneni & Tegmark (2025), Language Models Use Trigonometry to Do Addition ↗
Small-integer addition in GPT-J, Pythia-6.9B, and Llama 3.1-8B; most detailed circuit analysis in GPT-J.
- Engels et al. (2024/ICLR 2025), Not All Language Model Features Are One-Dimensionally Linear ↗
Circular calendar features and task-specific interventions; not a universal geometry of intelligence.
- Rahmanzadehgervi et al. (ACCV 2024), Vision language models are blind ↗
Seven primitive visual tasks, four historical model snapshots. Scores below refer to this proceedings version.
- BlindTest revised preprint (March 2025 analysis), v6 ↗
Later probing and separated-object experiments qualify where failures arise; encoder probes cover two tasks in LLaVA-OneVision-0.5B and Phi-3.5-vision, not every tested model.
- Fu et al. (2024), BLINK: Multimodal Large Language Models Can See but Not Perceive ↗
3,807 multiple-choice questions over 14 visual tasks. Version-specific results, not current rankings.
- Hao et al. (2024), Training Large Language Models to Reason in a Continuous Latent Space, v1 ↗
Specially trained latent-reasoning model; benchmark gains vary by task.
- Turpin et al. (2023), Language Models Don’t Always Say What They Think ↗
Prompt-bias experiments on GPT-3.5 and Claude 1.0; explanations can omit influences on answers.