The system maintains two distinct personas (“Fujiwara”, a stoic Edo-period ronin, and “James”, a formal British analyst) across 220 dialogue turns in stable equilibrium. This shows that cognitive coherence and tone consistency can be controlled at runtime rather than in model weights.

Unlike LangChain or RAG frameworks that orchestrate prompts, Sigma Runtime treats the model as a dynamic field with measurable drift and equilibrium parameters. It applies real-time feedback — injecting entropy or coherence corrections when needed — to maintain identity and prevent both drift and crystallization. The effect is similar to RLHF-style fine-tuning, but done externally and vendor-agnostic.

This decouples application logic from any specific LLM provider. The same runtime behavior has been validated on GPT-5.2 and Gemini-3, with Claude tests planned next.

We use narrative identities like “Fujiwara” or “James” because their linguistic styles make stability easy to verify by eye. If the runtime can hold these for 100+ turns, it can maintain any structured identity or agent tone.

Runtime versions ≥ v0.4 are proprietary, but the architecture is open under the Sigma Runtime Standard (SRS): https://github.com/sigmastratum/documentation/tree/main/srs

A reproducible early version (SR-EI-037) is available here: https://github.com/sigmastratum/documentation/tree/bf473712a...

Regulated under DOI: 10.5281/zenodo.18085782 — non-commercial implementations are fully open.

HN discussion focus: – Runtime-level vs weight-level control – Model-agnostic identity stability – Feedback-based anti-crystallization – Can cognitive coherence be standardized?