Skip to content
View karesansui-u's full-sized avatar

Block or report karesansui-u

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
karesansui-u/README.md

AI doesn't sleep. That's why it breaks.

LLMs degrade in long conversations. Everyone assumes it's a context length problem. It's not. It's contradiction accumulation.

GPT-4o-mini: 100% → 10%. Gemini: 100% → 0%. Google's 1M-token window? Still -47.8pp. Removing contradictions restores accuracy. Expanding the window doesn't.

I built cognitive sleep for AI — a metabolism that resolves contradictions during idle time, like human sleep consolidates memory.

Result: +52.2pp accuracy improvement across 8 models (p<0.001). 18 patents filed. 8 concepts with zero prior art worldwide.


🧬 Projects

delta-prune — Scan & clean contradictions before sending to any LLM API. 3 lines of code. PyPI

from delta_prune import DeltaPrune
prune = DeltaPrune(llm=OpenAILLM())
result = prune(messages)  # contradictions resolved

DeltaZero — Full metabolic architecture. 4-layer memory, temporal integration, survival equation monitoring. The research system behind the papers.

DeltaLint — Structural contradiction scanner for codebases. Finds where one module's assumptions contradict another's behavior.


📄 Papers

"Context rot is not a length problem. It's a contradiction problem."

# Title Key Result DOI
1 Structural Collapse as Information Loss S = μ × e^{-δ}: contradiction causes exponential decay. Validated on SAT + 11 LLM models. Formally verified in Lean 4. Zenodo
2 Predicting Computational Cost from δ Same δ governs both solution existence and computational cost. Sensitivity exponent is solver-dependent. Zenodo
3 Cognitive Sleep for LLMs External metabolism prevents context rot. ON vs OFF: +52.2pp (p<0.001, d=8.80). Unexpected: ON exceeds contradiction-free baseline. Zenodo

🔬 Lean 4 formal proofs — 160 propositions, sorry = 0, axiom = 0

📦 OSF Project — All papers, data, and code in one place


Software engineer, Japan

@karesansui-u's activity is private