Research records.
Questions that remain open.

Persistent AI systems may retain factual summaries while losing reasoning context, rejected hypotheses, and constraints as sessions change. The Unified Cognitive Substrate preprint proposes a way to represent and preserve selected signals across those transitions.

The authors report an eleven-experiment investigation, including a 1,563× change in context-sensitive routing differentiation under a specific controlled benchmark. That is not a claim that an AI system became 1,563 times more intelligent, accurate, or reliable.

The versioned paper and companion code are publicly available for inspection through Zenodo. The report should be read with its methods, version history, assumptions, and limitations; independent replication is not established here.

A model can produce an apparently good decision without establishing that the same method will hold under later conditions. These linked essays examine judgment formation, historical reasoning context, and the difference between fluent answers and dependable operation.

The February 2026 posts are publicly accessible essays. They include author interpretations and a prior-art search whose completeness has not been independently established here. They should not be mistaken for peer-reviewed confirmation of novelty or a demonstrated system-wide reliability gain.

Related engineering specifications can be linked when their authoritative public texts and evidence are available.

This proposal asks how evidence, contradictory facts, and constraints can remain traceable as an AI system forms a conclusion.

The proposed ratchet approach is intended to preserve validated constraints across stages of analysis. Whether a particular implementation achieves that property requires tests of source quality, adversarial inputs, temporal consistency, and the actual execution path.

A longer private reference document has been described, but the full authoritative text, public license, and independent benchmark record have not been verified for this page. Treat the mechanism as a research proposal rather than a proven legal or professional decision system.