Field notes on agent infrastructure, retrieval, and evaluation—what shipped, how it held up, and the evidence behind the decision.
PolicyStrata is public as a paper PDF, reproducible artifact kit, and GitHub release for cross-layer policy drift testing in LLM data-agent stacks.
Evidence: Benchmark: 1720/1720 non-cleanBenchmark: 1720/1720 non-clean | Clean controls: 80, 0 false positives
A local-first evidence workflow for agents with zero-dollar budget enforcement, open discovery, provider escalation, and inspectable artifacts.
Evidence: Release: DocPull 5.0.0Release: DocPull 5.0.0 | Boundary: --budget 0
Portable OKF source bundles let agents reuse scraped knowledge without turning every crawl into another custom format.
Evidence: Export spec: OKF v0.1Export spec: OKF v0.1 | Smoke-tested: 5 bundles
Grammar constraints reduce SQL failure modes on adversarial schema-drift prompts over live data without claiming a blanket quality boost.
Evidence: Dataset: 20M NYC Taxi rowsDataset: 20M NYC Taxi rows | SQL grammar: 43 Lark rules
A routing policy for web retrieval: crawl directly when the URL is known, then escalate for discovery, freshness, or archived surfaces.
Evidence: Targets: 8 web surfacesTargets: 8 web surfaces | Runs: 120 provider runs