Frontier-scale large language models (LLMs) can become unavailable due to various factors like pricing changes or geopolitical constraints. The concept of an "Evidence-Carrying Cognitive Mesh on DePIN" offers a solution to maintain AI capabilities even when these models are inaccessible. It focuses on using small local models alongside verifiable evidence artifacts that can be shared across decentralized compute networks. Instead of relying on a central authority for validation, nodes in this system use observable proof objects—like receipts and signatures—that can be easily verified by others. The approach shifts the perspective on intelligence from depending on large models to viewing it as a property of the network itself. The cognitive mesh acts like a decentralized supply chain for knowledge. Each node generates outputs, accompanied by evidence detailing the processes used to arrive at those outputs. This evidence is then verified by other nodes, creating a graph of claims with clear provenance. The system emphasizes "observable-only" governance, meaning decisions are based on verifiable artifacts rather than authority figures, which is crucial for decentralized environments. Performance metrics are redefined in terms of integrity, capability, and accessibility. Integrity ensures that outputs are linked to verifiable evidence, capability refers to the system's ability to produce useful results despite using smaller models, and accessibility guarantees that these results remain available to all users without gatekeeping. The cognitive mesh improves the power of small models by leveraging shared, verified memories and processes, enabling a robust AI framework that doesn't rely on the largest models.