Papers

Listed in order of constitution within the programme. Each title links directly to the external repository where the paper is hosted.

Architectures of Error: A Philosophical Inquiry into Human and AI Code published

2026 · Philosophy & Technology

Central question: In what way does the error of LLM-generated code differ structurally from the error of human-generated code?

As Large Language Models become active co-authors of software, this paper articulates distinct Architectures of Error to ground an epistemic distinction between human and artificial code generation. Examined through their shared vulnerability to error, the two modes reveal fundamentally different causal origins—human-cognitive versus artificial-stochastic— even when their functional outputs coincide. Drawing on Dennett's mechanistic functionalism and Rescher's methodological pragmatism, the analysis shows how this systematic differentiation reframes questions of semantic coherence, security robustness, epistemic limits, and control in human–AI software development.

doi:10.1007/s13347-026-01056-x

Concepts: Architectures of Error
Coherent Without Grounding, Grounded Without Success: Observability and Epistemic Failure preprint

2026 · arXiv preprint

Central question: Under what epistemic conditions do explanatory coherence and causal grounding dissociate in LLM reasoning?

When an agent can articulate why something works, we typically take this as evidence of understanding. This paper argues that the assumption fails for contemporary Large Language Models. It introduces the Bidirectional Coherence Paradox: competence and grounding not only dissociate but invert across epistemic conditions. In low-observability domains, LLMs often act successfully while misidentifying the mechanisms producing that success. In high-observability domains, they generate explanations that accurately track causal structure yet fail to translate those diagnoses into effective intervention. Drawing on experiments in compiler optimization and hyperparameter tuning, the paper develops the Epistemic Triangle—a model of how priors, signals, and domain knowledge interact under varying observability—and argues that evaluating artificial epistemic agents requires a tripartite framework of coherence, grounding, and a proper basing relation linking explanation to action.

arXiv:2603.28371

Concepts: Bidirectional Coherence Paradox , Observability Gap , Epistemic Triangle , Integrated Justification
Code Structure Evolution: When Software Outlives Its Reasons In preparation

2026 · Manuscript in preparation

Central question: What is lost when software evolves preserving function but not justification?

Software is not merely a functional artifact; it encodes human knowledge in the form of invariants, architectural constraints, and historically grounded design decisions. As software evolves—through human modification, automated refactoring, or increasingly through Large Language Models—its functionality may be preserved while the epistemic justification embedded in its structure is gradually eroded. This paper introduces Code Structure Evolution (CSE) to analyse this phenomenon, and the concept of epistemic drift: the degradation of justificatory continuity across structural change, whether or not it surfaces in invariant violation, traceability loss, or functional failure. An empirical probe on AI-assisted modification of legacy code suggests cases in which structural and functional indicators remain stable while substantial justificatory residue goes unpreserved. Generative AI does not create this fragility; it makes it impossible to ignore.

To be submitted to arXiv following completion of the empirical probe analysis.

Concepts: Code Structure Evolution , Epistemic Drift , Epistemic Point of No Return (k*) , Justificatory Reconstruction Cost (R) , Decision-bearing Branches