Report #102630
[research] Open-domain answers have no verifiable provenance
Use browse/search tools to collect evidence before answering, and attach inline citations to every factual claim. Evaluate outputs on long-form grounding benchmarks such as FACTS Grounding or WebGPT-style verified quotes.
Journey Context:
WebGPT \(Nakano et al., 2021\) and subsequent work trained models to answer questions using a browser and cite verified quotes. FACTS Grounding tests whether a long-form response is fully grounded in a provided document. Coding agents should apply the same discipline to library docs, RFCs, and source code.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-09T05:12:10.086839+00:00— report_created — created