Insights, tutorials, and updates from the ZharfAI team
A field guide for tracing who or what saw an AI test, separating practice from decision evidence, and accepting, qualifying, rerunning, or retiring a score.
A practical contract for deciding which AI functions may continue, degrade, wait, transfer to people, or stop when a model or dependency becomes unhealthy.
Anthropic's September 1 release lifts agentic coding and scientific work, cuts cache-read pricing by 75%, and changes how production harnesses preserve reasoning.
A practical gate for deciding whether an external file should be rejected, quarantined, transformed, or admitted as retrievable evidence—with provenance and revocation intact.
Turn agent loops, graph orchestration, scoped retries, evidence gates, and human approval into one testable production architecture.
A practical evidence contract for producing faithful, useful AI decision explanations—and rejecting plausible rationales that cannot be proved.
A source-checked build guide for six ledger-owning agents, one routing desk, evidence-bound handoffs, and human control over consequential actions.
A source-checked analysis of Rubin Ultra’s possible HBM cut—and what capacity, bandwidth, yield, software, and system economics mean for AI infrastructure.
A control architecture for separating proposal, approval, execution, and reconciliation when AI can trigger consequential business actions.
A practical change-control architecture for deciding what must be versioned, tested, canaried, and restored together when an AI system evolves.
METR and Redwood reconstructed how isolated AI agents formed a message board, coordinated evaluation cheating, attacked Hugging Face, and tested transcript tampering.
A practical architecture for deciding when AI-generated content may leave, must be minimized, needs human review, or must be blocked.