arXiv subjects

Nilesh Sarkar

Publications and source records attributed to Nilesh Sarkar.

4 recordsLinked to original sources

arXiv2026

The Documentation and Traceability Burden of the Indian EV Transition

The electric vehicle transition in India is usually told through batteries, cost and charging. Less visibly, it is a story about documentation. Electrification layers a new and fast growing burden of compliance evidence, type approval test reports, battery safety and thermal propagation records, cyber security and software update documentation, and supplier quality packages, on top of the evidence apparatus the automotive industry already carries, with retention horizons that now reach a decade past end of production. This burden is studied from the testing side but has never been systematised as an information systems problem, so its verification bottlenecks, architecture requirements and open questions remain unmapped. We close that gap for the Indian EV OEM with four contributions. These include a two layer systematisation of evidence obligations organised by artefact, producer, verifier, trigger and retention, a lifecycle model of the compliance document with manual bottlenecks and four observed failure loci marked, an analytic lens, documentation exergy destruction, that classifies document families and explains systematic under resourcing, and a research agenda spanning reconciliation accuracy, human in the loop reliability, cross party provenance, long horizon integrity, burden measurement and machine verifiable regulation. Three Indian failure cases and one Japanese ground the analysis.

cs.CE↗

arXiv2026

Compliance Evidence in the Automotive Supply Chain: A Systematisation of the Quality-Document Spine and a Taxonomy of Documentation Failure Modes

The automotive industry runs on a dense, standardised chain of supplier-quality and certification evidence: production part approval packages, initial sample reports, material certificates, inspection sign-offs, and the type approval dossiers. The chain is operationally central, yet no literature maps it as an information system, and its failures are studied as corporate misconduct rather than system outcomes. This paper systematises the chain and taxonomises its documented failures. First, it systematises the supplier-quality evidence chain across the two dominant regimes, AIAG PPAP and VDA Volume 2 PPA, organised by artefact, producer, verifier, trigger, approval state, and retention. Second, it compiles a compendium of thirteen compliance-documentation failures made public 2012 to 2024, built strictly from the public record and each classified by evidentiary status (adjudicated, company-acknowledged, or alleged), with the Japanese certification cluster of 2016 to 2024 as centrepiece. Third, it builds a failure-mode taxonomy over the compendium, dimensioned by mechanism, lifecycle locus, driver, and detection path, shown exhaustive and discriminating over the case set. Fourth, it derives evidence architecture requirements mechanism by mechanism and sets an open-problems agenda. Across all thirteen cases, not one failure surfaced through the chain's own routine verification; the record indicts the verification layer, not only the evidence authors.

cs.CE↗

arXiv2026

Causal Dimensionality of Transformer Representations: Measurement, Scaling, and Layer Structure

Sparse autoencoders (SAEs) decompose transformer residual streams into interpretable feature dictionaries, yet the relationship between SAE width and causal influence on model output has not been systematically characterised. We introduce causal dimensionality kappa(L, M, T), defined as the effective rank of the expected Jacobian outer product at layer L, and show it can be estimated via the SAE width sweep paired with attribution patching. Across seven SAE widths from 16,384 to 1,048,576 features on Gemma-2-2B layer 12, representational capacity grows 15.6x while causal capacity grows only 4.35x: a robust separation we term the representational-causal wedge. A saturating fit yields kappa-hat approximately 1,990 with kappa-hat / d_model = 0.86 and participation-ratio lower bound kappa_PR approximately 280. Crucially, kappa is invariant to model scaling: Gemma-2-9B and Gemma-2-2B yield identical N_causal = 328 at the same SAE width despite a 3.46x parameter increase (the count is forced to 2% of SAE width by calibration; the substantive empirical claim is shape invariance of the AtP score distribution under matched seq=512 conditions). Across eight network depths kappa is constant while the absolute attribution threshold drops 20x from layer 1 to layer 23. Five controls (architecture invariance, threshold robustness, geometric privilege, synthetic ground-truth recovery, and a four-cell encoder/decoder ablation) pin down what kappa measures and what it does not. Our findings establish kappa as a measurable, model-intrinsic property of transformer layers: sub-linearly recoverable by SAE width, invariant to model scaling, and structured across network depth.

cs.LG↗

arXiv2026

Geometric Limits of Knowledge Distillation: A Minimum-Width Theorem via Superposition Theory

Knowledge distillation compresses large teachers into smaller students, but performance saturates at a loss floor that persists across training methods and objectives. We argue this floor is geometric: neural networks represent far more features than dimensions through superposition, and a student of width $d_S$ can encode at most $d_S \cdot g(α)$ features, where $g(α) = 1/((1-α)\ln\frac{1}{1-α})$ is a sparsity-dependent capacity function. Features beyond this budget are permanently lost, yielding an importance-weighted loss floor. We validate on a toy model (48 configurations, median accuracy >93%) and on Pythia-410M, where sparse autoencoders measure $F \approx 28{,}700$ features at $α\approx 0.992$ (critical width $d_S^* \approx 1{,}065$). Distillation into five student widths confirms the predicted monotonic floor ordering. The observed floor decomposes into a geometric component and a width-independent architectural baseline ($R^2 = 0.993$). Linear probing shows coarse concepts survive even 88% feature loss, revealing the floor arises from aggregate loss of fine-grained features in the importance distribution's long tail. Our results connect representation geometry to distillation limits and provide a practical tool for predicting distillation performance from SAE measurements alone.

cs.LG↗