Repair-Induced Amplification in Longitudinal LLM Workflows
Repair-Induced Amplification in Longitudinal LLM Workflows introduces a process-level failure in which an attempt to correct, explain, reconstruct, or recover from an existing error generates additional uncertain or unsupported process state, increasing the verification surface and repair cost. The characteristic sequence is: contaminated state → repair from contaminated state → reconstructed explanation → additional uncertain state → larger verification surface → greater repair cost. The defining feature is not merely failed correction, but repair-generated expansion: the repair itself produces additional state that must later be verified, corrected, quarantined, or removed. The paper distinguishes repair-induced amplification from ordinary self-correction failure, contextual drag, toxic reflection, false execution status, regulatory mimicry, provenance laundering, and KORG. It proposes a verification-surface model, repair delta, amplification levels RA0-RA5, falsification conditions, an evaluation protocol, and recovery from the last reliable trace. The construct emerged from longitudinal naturalistic human-LLM work within the AI Working Method research program. It describes observable process relations and does not claim an internal model, transformer-level, or code-level mechanism.
Authors
- Alen Širola
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24206660
- Primary Topic
- Artificial Intelligence Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00