← Back to Entropic
Provenance

Why you can trust the Monday Brief

The brief tells you which client to call. This page tells you how it knows — and what it does when it isn't sure.

The problem with AI that reports on your business

Most AI tools give you an answer. They don't give you the route the answer travelled.

By the time a line reaches your Monday Brief, it has passed through several hands: a source system, an extraction step, a reasoning step, a summarisation step. Each hand can drop something, distort it, or invent it. And a fluent, confident model at the end of that chain cannot repair a bad read at the start — it can only restate it more convincingly.

That's tolerable when an AI is drafting copy. It isn't tolerable when it's telling you a $3k/mo client is about to churn.

What Entropic does instead

1. Chain, not just output
Every claim in your brief carries the chain it came through. You can open the underlying email, campaign, or task.
2. Graded transmitters
Each step in the pipeline carries a reliability grade for the kind of work it's doing. Grades are per-domain — good at reading campaign data isn't the same as good at reading client sentiment.
3. Weakest link caps the chain
A chain is only as trustworthy as its worst step. No amount of downstream confidence raises it.
4. Corroboration, gated
Two independent sources agreeing raises trust. Two restatements of the same source do not — and the system knows the difference.

What happens when the system isn't sure

Chain
Content
What Entropic does
Sound
Consistent
Appears in your brief
Sound
Sources contradict
Flagged for your judgment — the most useful signal in the brief
Weak
Consistent
Held; the system looks for corroboration first
Weak
Sources contradict
Quarantined. Never served.

A sound chain carrying contradicted content isn't a failure. It's the single most informative thing the system can hand you.

The research

This architecture isn't improvised. It comes from ISNAD, an open-source framework published by Entropic co-founder Ali Zahid Raja: Grading the Narrators: An Isnād–Rijāl Framework for Claim-Level Provenance in Multi-Agent Knowledge Systems (arXiv:2607.24117, July 2026).

The framework adapts a method that already solved this problem once. For twelve centuries, classical Islamic hadith scholarship worked on a structurally identical question: how do you decide whether to trust a report that reached you through a chain of human narrators? The answer they built was a discipline — every report carries its chain, every narrator carries a documented reliability grade, the chain is only as strong as its weakest transmitter, and the content of a report is criticised separately from the route it took.

Multi-agent AI has exactly that problem, with models in the place of narrators. ISNAD transfers the method.

The rigour belongs to twelve centuries of scholarship. The transfer is the contribution.

The five ISNAD principles
  1. Every claim carries its transmission chain (isnād)
  2. Every transmitter keeps a living, per-domain grade (rijāl)
  3. The weakest link caps the chain's trust
  4. Independent corroboration upgrades — capped and gated (mutābaʿāt)
  5. Content is criticised separately from the chain (matn)
Paper →
arXiv:2607.24117
Paper DOI →
10.48550/arXiv.2607.24117
Source code →
Apache-2.0 on GitHub
Package →
pip install isnad
Software DOI →
Archived on Zenodo
Project page →
Full framework write-up

What this means for you

You don't have to take the brief's word for anything. Every red flag is a link back to the thing that caused it — so the first thing you do on a Monday isn't verifying the AI, it's making the call.

See it on your own stack.

15 minutes. I'll look at your tools live and tell you what the first brief would surface.

Book Your Free Fit Call →
← Back to Website