A Write Barrier That Blocks Structural Collapse in LLM Reasoning

In multi-step tasks, LLM outputs may replace structured intermediate forms with scalar results.

Example: (2 + 3) * 4 → 20

This removes structural decomposition and makes downstream reasoning dependent on a scalar result.

I built a minimal prototype that enforces admissibility before persistence using a write barrier. The write barrier separates model output from persistent state.

Core mechanisms: * Append-only lineage (no in-place mutation) * Explicit proposal → invariant check → commit cycle * Whitelisted structural transforms * Deterministic invariant checks before commit

If a transform attempts to collapse structure (e.g., replacing a decomposed expression with a scalar), the proposal is rejected and never enters the lineage.

Important distinction: Invalid states may still be representable as data, but they are uncommittable under the governed commit path.

This does not modify the model. It constrains persistence architecturally around model outputs.

Arithmetic is used purely as a stress-test domain to isolate one narrow claim: Certain structural collapses can be made impossible to persist.

Limitations: * Domain-specific invariants * Not a symbolic solver * Does not improve model accuracy * In-memory prototype storage

GitHub: https://github.com/PersistentVlad/persistent-reasoning-archi...