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 under contradiction. Removing contradictions restores accuracy. Expanding the window doesn't.
The framework generates testable predictions. The multi-attractor extension predicts Gemini's accuracy follows P = 1/(1+K/L) — derived from 3 data points, confirmed at 5 context lengths (32K–512K) with MAE 2%, two out-of-sample predictions within 1%. Full chain: axioms → Lean 4 proof → prediction → experiment.
delta-prune — Scan & clean contradictions before sending to any LLM API. 3 lines of code.
from delta_prune import DeltaPrune
prune = DeltaPrune(llm=OpenAILLM())
result = prune(messages) # contradictions resolvedDeltaLint — Structural contradiction scanner for codebases. Finds where one module's assumptions contradict another's behavior.
"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 → exponential decay. Multi-attractor prediction confirmed (K=19.5±3.4, MAE 2%, 5 context lengths). Lean 4 verified (sorry=0). | Zenodo |
| 2 | Predicting Computational Cost from δ | Same δ governs both solution existence and computational cost. Sensitivity exponent is solver-dependent. | Zenodo |
| 3 | Contradiction Metabolism for LLMs | External metabolism prevents context rot. +52.2pp (n=3, p=0.027; 8-model sign test p=0.0107). Metabolized system exceeds contradiction-free baseline. | Zenodo |
📊 delta-survival-papers — All papers (TeX + PDF), raw experiment data (1,000+ trials across 11 models), reproduction scripts, and Lean 4 proofs. See DATA.md for direct download links.
🔬 Lean 4 formal proofs — 160 propositions, sorry = 0, axiom = 0
📦 OSF Project — All papers, data, and code in one place
Software engineer, Japan