The System That Evolves: A Three-Layer Taxonomy for Persistence, Consolidation, and Test-Time Adaptation, with Honest Measurements at Small Scale
Abhijeet Verma · Zenodo (CERN European Organization for Nuclear Research) · 2026
A three-layer taxonomy for how an LLM system changes over time, each layer tied to a falsifiable containment property: Persist (external memory, no weight change - reversible), Consolidate (offline gated LoRA weight updates - gated-and-revertible), and Adapt (test-time training during inference - sandboxed-and-discarded). We implement all three on a ~10.5M-parameter checkpoint and measure them honestly at small scale. Persistence: 15/18 facts answered warm with zero model weights touched (AST-verified) and 0 fabrication cold. Consolidation: the pre-registered <5% forgetting target is MISSED in both conditions (+98.7% degradation unmitigated, +97.8% with EWC+replay) - published as it landed. Test-time training: on-passage next-token accuracy rises up to +0.152 at the most aggressive setting, but elsewhere degradation grows with exactly the settings that most help on-passage (-0.0127 mean at the aggressive cell), with byte-for-byte rollback verified in every cell. The central claim is the pairing, not either number alone; negative results are pre-registered and reported, not dropped.