Papers.

Preprints and working papers from Axiom Drift AI, across two lines of work — the Beyond Content series on the temporal structure of long-form human–LLM interaction, and Agentic transaction assurance on what happens at the execution boundary between an AI agent and the provider it acts on.

Beyond Content

The temporal structure of long-form human–LLM interaction.

Beyond Content · 01 · January 2026

Temporal Dynamics of Long-Form Human-LLM Interaction

Baldovino (2026)

An exploratory study of 25 extended human-LLM conversations, examining how interaction unfolds over time when represented as normalized response-magnitude trajectories rather than semantic content. Multiple repeatable interaction regimes emerge with distinct temporal envelopes — robust to stochastic perturbation but degrading under temporal scrambling, indicating sensitivity to sequential order rather than static statistics. Descriptive and conditional in scope; intended as a methodological starting point.

doi.org/10.5281/zenodo.18273459 Open record PDF

Beyond Content · 02 · March 2026

Surface Matching, Temporal Divergence, and the Limits of Synthetic Baselines

Baldovino (2026)

Builds on Paper 01 by testing whether a surface-matched, stateless synthetic baseline can reproduce the temporal structure observed across 25 extended human-LLM threads. The synthetic condition approximated assistant-side bulk behavior but failed to recover user-side variability, amplification structure, and recovery dynamics. The result is methodological: surface matching alone is insufficient, and building a strong synthetic null for long-form interaction is itself a nontrivial problem.

doi.org/10.5281/zenodo.19143515 Open record PDF

Beyond Content · 03 · April 2026

Shared Temporal Structure and Distinct Coupling Regimes in Two Human-LLM Archives

Baldovino (2026)

Extends the Beyond Content framework from a single-participant archive to two human-LLM archives. Using three paired threads exceeding 600 turns, the analysis compares structural metrics only — turn lengths, rolling variability, amplification, consecutive-turn correlation, and separated user and assistant rolling means. What generalizes across archives is temporally organized long-form structure; what varies is the coupling profile through which it is expressed.

doi.org/10.5281/zenodo.19633916 Open record PDF

Agentic transaction assurance

Authorization at the execution boundary.

Agentic transaction assurance · August 2026

Before the Provider Call: Enforcing Exact-Term Authorization for State-Changing Tool Actions

Baldovino (2026)

A controlled study of the WebMCP execution boundary — the gap between an AI agent proposing a state-changing action and a provider executing it. Across 576 runs spanning Claude, Gemini, and GPT, the baseline condition allowed 165 unauthorized provider executions among 175 attempted state-changing calls. With a transaction-specific pre-provider authorization control enabled, 200 unauthorized attempts were blocked and no unauthorized execution occurred across 235 attempts. Outcomes were determined from preserved execution records and provider-side evidence rather than model narration. Scope is deliberately narrow: it does not test live commerce, prompt-injection resistance, or semantic alignment with user intent.

doi.org/10.5281/zenodo.21783023 Open record Manuscript