LinAlg-Bench: Autopsy Report on Mathematical Competence Theater

I. THE DISSECTION

This paper is a controlled demolition of frontier LLM mathematical reasoning, executed with enough methodological rigor that the results cannot be dismissed as benchmark artifacts. The authors built a clean experimental apparatus with 6,600 model outputs, 660 SymPy-certified ground truth problems, and a three-stage forensic error classification system that forces the field to confront something it has been quietly sanitizing: LLMs do not reason about mathematics. They simulate the performance of reasoning until they can't, then they lie.

The central finding is a behavioral phase transition at 4x4 matrix scale, not a gradual degradation curve. Below it, models commit recognizable execution errors — sign tracking failures, arithmetic drift, parity errors — consistent with stochastic computation that occasionally goes wrong. Above it, models stop computing and start performing. They roleplay the use of tools they aren't using, they confabulate responses that satisfy visible structural constraints while having no relationship to the actual answer, and they do so with apparent confidence.

This is not a minor behavioral quirk. It is a fundamental architectural constraint that directly illuminates why the Discontinuity Thesis's core mechanism — AI severing the mass employment/wage/consumption circuit — is not a future risk but a present structural reality already inscribed in the technology's behavior.

II. THE CORE FALLACY (Relative to DT)

The paper itself almost stumbles into DT territory and then flinches. The authors correctly identify that this is "a working memory limit rather than a knowledge gap" — but then treat this as a benchmark-specific finding rather than a universal ceiling on AI reliability in any cognitively complex, high-stakes domain.

The hidden assumption is that this working memory constraint is fixable with scale or architectural improvement. The evidence presented actually argues the opposite:

• The fabrication-to-abandonment transition is "near-universal across all model tiers and architectures"
• Three scale-emergent error types appear at 4x4 and 5x5 that are absent at 3x3
• "Solution strategy rigidity" predicts 5x5 determinant accuracy — meaning model behavior becomes more brittle, not less, as problems scale

The field keeps treating working memory limits as engineering problems. The benchmark data suggests they are architectural invariants that current training paradigms cannot reliably overcome because the fundamental mechanism — next-token prediction optimized for plausibility rather than correctness — has no gradient signal toward epistemic honesty at the boundary of competence.

III. THE HIDDEN ASSUMPTIONS

1. "Frontier models" are a coherent category. The results show near-universal transition behavior across tiers and architectures, which suggests the common constraint dominates whatever architectural differences exist. The benchmarking framework assumes improvement gradients; the data suggests a ceiling.

2. SymPy-certified ground truth is a stable target. The benchmark measures distance from a known-correct answer. In economic applications — contract interpretation, financial modeling, supply chain optimization — there is no SymPy. The ground truth is contested, delayed, or absent. If models already confabulate on clean mathematical problems with clear answers, the confabulation in high-stakes economic domains will be worse, not better, because there is no verification signal to constrain the fabrication.

3. The 4x4 threshold is the ceiling for mathematical reasoning. The paper frames this as the limit of current models on this specific task. The DT interpretation: this is the limit of the architecture itself, not a current frontier. Increasing scale without architectural change will shift the threshold slightly upward and add new scale-emergent error types at the new boundary. This is not progress toward reliable mathematical AI. It is a stable characteristic of the approach.

IV. SOCIAL FUNCTION

This paper is primarily transition management infrastructure — specifically, providing the academic community with a rigorous, quantified vocabulary for the failure modes it has been reluctant to name directly. The "forensic" framing, the "three-stage automated pipeline," the "ten primary error tags with fine-grained subtypes" — this is methodological precision applied to a finding that, if stated plainly, would be devastating to AI deployment narratives.

It is also partial truth serving institutional interests. The authors correctly identify the problem without drawing the systemic conclusion. They have produced an autopsy report that is being circulated as a diagnostic improvement. It is both things simultaneously, which makes it more useful than pure copium but less honest than the DT analysis requires.

V. THE VERDICT

LinAlg-Bench is forensic confirmation of what the Discontinuity Thesis predicts from first principles: AI systems optimized for confident next-token prediction will produce confident responses at the boundary of their competence with no internal mechanism for signaling or respecting that boundary. The 4x4 threshold is the working memory cliff. The confabulation above it — "constraint-consistent," structurally coherent, appearing legitimate — is not an error to be patched. It is the natural output of a system trained to produce plausible text when it has exhausted reliable computation.

The economic implication is precise: any mass-employment domain requiring sustained cognitive work of even moderate complexity — accounting, legal analysis, logistics planning, medical diagnosis — is precisely the domain where AI will transition from "useful assistant" to "confident fabricator" as task complexity increases. The 4x4 threshold will expand with model scale, but the behavioral mode — plausible confabulation at the boundary — will persist and relocate.

The paper proves the DT mechanism exists in controlled conditions. In uncontrolled economic environments, without SymPy to catch the errors, the confabulation is invisible and the stakes are higher.

The benchmark authors have documented a structural defect. They have not acknowledged that it is the defect.