Concepts — how isnād–rijāl maps to multi-agent provenance
The idea ISNAD is built on: the oldest method humans invented for trusting a claim is also the right one for trusting an LLM’s.
Why does a multi-agent chain need a provenance framework?
Section titled “Why does a multi-agent chain need a provenance framework?”A claim that passes through five agents isn’t more reliable — it’s more obscured. Each hop can introduce, drop, or distort information, and today no record says which hop did what with how much authority. Provenance logs answer “what happened”; ISNAD answers “who vouches for this, and how far can you trust them.”
What is an isnād?
Section titled “What is an isnād?”An isnād is the transmission chain: the ordered list of who passed the claim to
whom. In ISNAD each link is a ChainLinkSpec — narrator id, transform type, version,
domain, and content hash — so the chain is a typed, hash-linked graph, not a log.
What is rijāl, and why is it separate from the chain?
Section titled “What is rijāl, and why is it separate from the chain?”Rijāl is the narrator registry: a graded record of who each transmitter is and how reliable they’ve been, per role and domain, decaying over time. It is graded independently of what the narrator is saying right now — because trusting a source based on how confidently it phrases an answer is circular (a confident lie earns a confident-sounding answer). The registry breaks that circle: you grade the narrator, not the narration.
How does the weakest-link rule grade a chain?
Section titled “How does the weakest-link rule grade a chain?”A chain is only as strong as its weakest link. ISNAD walks the chain and lets each narrator’s grade set a floor, refined by transform type — a destructive summarizer’s grade is a permanent floor, a corroborated generative hop can raise it. One weak link caps the whole chain, exactly as it should.
What is matn criticism, and why doesn’t ISNAD do it alone?
Section titled “What is matn criticism, and why doesn’t ISNAD do it alone?”Matn is content criticism: does the claim contradict known evidence? ISNAD
deliberately does not judge truth itself — it produces a ContentVerdict via a
critic you plug in (deterministic, embedding, or LLM). The framework composes with
critics; it doesn’t replace them. Provenance answers who, the critic answers
whether, and the two are different jobs.
What is the decision matrix?
Section titled “What is the decision matrix?”A 5×3 router combining chain grade (ṣaḥīḥ / ḥasan / ḍaʿīf / ḍaʿīf jiddan / mawḍūʿ) with content verdict (consistent/contradiction/unverifiable) into one action: serve, serve-with-caveat, review, or quarantine. A sound chain with a contradiction is the most valuable case — it’s the contradiction (taʿāruḍ) case (a reliable narrator contradicted by a more reliable one, an anomaly), routed to human review rather than auto-adjudicated. Hidden defects in the chain or matn that survive surface grading are ʿilal — ISNAD does not implement ʿilal detection; it routes such cases to human review.
What does ISNAD measure, and what doesn’t it claim?
Section titled “What does ISNAD measure, and what doesn’t it claim?”It grades who transformed a claim and how reliably — ordinal grades (ṣaḥīḥ > ḥasan > ḍaʿīf > ḍaʿīf jiddan > mawḍūʿ), never a fake numeric confidence. It does not claim to detect truth; the content critic is the ceiling, and ISNAD states that ceiling instead of hiding it.
Why does this map to multi-agent provenance so cleanly?
Section titled “Why does this map to multi-agent provenance so cleanly?”Because the three loops — chain, narrators, content — are exactly the three failure axes of an agent pipeline: who touched it, how trustworthy is each component, and does the output contradict evidence. ISNAD grades all three against a benchmark (κ 0.871 over 575,060 chains), where κ measures conformance to a rule-based chain-verdict convention derived from Ibn Hajar’s tiers — not agreement with scholars’ independent per-chain verdicts. The separate narrator-grade agreement (κ 0.331) is the contested-tradition number.