ISNAD: Provenance and Trust Framework for Multi-Agent Systems

alizahidrajaa · reddit · 2026-07-08

The author open-sourced ISNAD, a declarative provenance framework designed for multi-agent and compiled knowledge pipelines, inspired by the classical Islamic science of hadith transmission. In multi-agent pipelines, a claim passes through multiple "narrators" (crawler → ingestion model → synthesis model). ISNAD maintains a tiered trust registry for each narrator by domain, attaches a transmission chain to every claim, and caps trust using the "weakest link" principle. Claims are then routed via a "chain-level scoring × content critique" decision matrix (serve/alert/review/quarantine). In real-world corpus experiments with 20,000 claims, the weakest-link scoring successfully isolated untrustworthy narrators, while confidence-based gating provided no extra benefit. Unverified aspects include cross-source corroboration (due to insufficient triggers in real data).

Related event: Open-Source Multi-Agent Trust Framework Inspired by Hadith Science(2 posts)→

Original post →

More from coding & agent

coding & agent channel →