Dev Proposes "Evidentiality Framework": Provenance Markers for Every LLM Claim
jzesbaugh · reddit · 2026-09-25
A developer shared an Evidentiality Framework on Reddit that attaches lightweight provenance markers to individual LLM claims:
- (u) supplied by the user
- (m) checked against an external source or tool
- (g) model-generated inference or unsupported output
The author stresses markers don't mean "true" — sources, users and models can all be wrong — the goal is making each claim's provenance visible in the text itself, so readers or downstream tools don't have to guess.
Open questions include what counts as a claim, how provenance propagates through summarization, how to handle mixed sentences, and whether the simple notation should be backed by a machine-readable structure. The framework is intentionally lightweight; the author invites developers in RAG/citations/provenance to implement or break it.
More from coding & agent
- Practical guide to shipping MCPs users want: match your dashboard, skip CodeMode, add docs search — teropa · 2026-09-25
- SkillsBench accepted at NeurIPS 2026: 87-task benchmark tests whether agent skills actually help — HanchungLee · 2026-09-25
- Neo Roundtable: a tiny shared conversation layer so ChatGPT, Claude and coding agents stop re-explaining your project — gambit2054 · 2026-09-25
- Dev builds local AI GTM workflow, argues the next platform entry point is hardware-bound — dotey · 2026-09-25
- TypeSafe AI launches Jev, a 'System One' model for bounded decisions in agent runs — hardimanjames · 2026-09-25
- "Astra, get me 99 fishing": one-line prompt sends an AI agent grinding RuneScape — justalexoki · 2026-09-25