Abstract:
Large Language Models (LLMs) can produce fluent explanations while returning unreliable
exact answers when mathematical prompts impose large representations, repeated operations,
or restrictive output budgets. This dissertation examines that observable weakness using
TokenPath, a trace-capture instrument, and the Token-Stress-Test (TST), a procedurally
generated collection of mathematical stress tasks. Finite-Space Constraints (FSC) are treated
as a theoretical motivation; an FSC-consistent stress outcome is defined operationally as an
evaluable incorrect, missing, or truncated answer to a registered TST item under a recorded
finite output budget. This rule identifies outcomes under stress but does not, by itself, establish
an internal causal mechanism.
Reconciliation of the retained records reconstructed the historical evidence from 46 selected
source files. The canonical ledger contains 9 000 eligible baseline observations across six
models, six non-leaking task families, five output budgets, and 50 observations per model–
family–budget cell. Exact task-specific scoring produced 499 correct answers, an accuracy of
5.54% (Wilson 95% confidence interval 5.09–6.04%). Performance varied substantially by
task family: sequence prediction reached 21.53%, symbolic expansion 9.07%, large-number
arithmetic 2.53%, repetitive sequences 0.13%, and base conversion and combinatorics 0%. The
last three families therefore have severe floor effects and require calibrated tiered validation before
they can support comparative claims. Token length was available for all six models, whereas
entropy and effective-vocabulary measures were comparable only for the three open-weight
models whose logits were retained.
Because correct outcomes were rare and several predictors were separated, the primary
multivariable analysis used Firth-penalised logistic regression. It estimated associations between
correctness and output budget, realised response length, retrospective within-family difficulty,
model, and family; these estimates are descriptive and not causal, and realised length is postgeneration.
The historical tool-ablation archive contained 140 declared groups, but only 70
matched the same item and ground truth. After excluding 10 leaking chain-reasoning pairs, 60
valid pairs remained; neither condition was correct, giving a sample recovery efficacy of zero
without establishing a population-level absence of tool benefit. A separate tool-event funnel
records requests, executions, successes, and verifications without treating them as paired effects.
The study implements TokenPath and TST and reports comparative mathematical outcomes.
The intended mechanism taxonomy and low-overhead mitigation claims are not fully supported
by the retained evidence. These limits qualify the achievement of the original four research
objectives.