A 200 Is Not a Read

Short summary

AI agents frequently answer confidently from web pages they never actually read, because HTTP 200 responses can return empty JS shells, Cloudflare challenge pages, or soft 404s. FetchGate proposes a deterministic gate in the orchestration layer that returns RETRIEVED, FAILED, or UNKNOWN verdicts before the model is allowed to answer. Unlike post-hoc RAG faithfulness checks, this approach hard-fails at the fetch boundary upstream of grounding.

• •HTTP 200 does not guarantee content was read — agents bluff from training weights when fetches return empty or wrong content
• •FetchGate introduces a deterministic three-state verdict (RETRIEVED/FAILED/UNKNOWN) outside the model's control
• •Complements but differs from RAGAS, TruLens, and Perplexity-style citation checks by gating at the fetch boundary, not after grounding

Generated with AI, which can make mistakes.