Skip to content

Proof for your AI's answers

When your AI gives an answer,
we keep the proof.

A locked record of where each AI answer came from.

If your team uses AI to write, price, underwrite, or advise customers, the EU will soon require you to prove what the AI did and where its answers came from — stored for at least six months, starting December 2027. ISNAD keeps that proof automatically: a locked, un-editable record of every answer — where it came from, who or what handled it, and how much you can trust each step. When a regulator or auditor asks, you hand over a folder instead of panicking.

The EU's new AI record-keeping rules start 2 December 2027 — and six months of records can't be backfilled. Start now.

1 · The claim

A claim passes through 5 hops — source, retriever, model, tool, agent.

2 · The grades

ISNAD labels each hop — Reliable → Caution → Weak → Do not use (or ungraded).

3 · The record

You get one signed record: who handled it, in what order, and how much to trust each step.

Apache-2.0 · arXiv:2607.24117 · 1,000+ tests · 575,060 graded chains

The proof, on one pageLOCKED

Answer to: "Is this client eligible for the loan?"

Source: client file + policy document

Handled by: AI model · reviewed by Sarah

Trust: Caution — one step was flagged as weak

Record locked: yes — if anyone changes it, it shows

This is the folder you hand to an auditor. Not a certificate — the proof.

Does this apply to you?

Three questions. Three yeses = it likely does.

Do you use AI to write, price, or advise customers?

Is a wrong answer consequential — a loan, a claim, advice someone acts on?

Do you operate in or serve the EU?

If all three are yes, you're likely an Annex III high-risk use case — think credit scoring, insurance, employment, biometrics, or any decision that affects someone's rights. Confirm with your counsel. Separately, if your AI is a safety component of a regulated product — a medical device, machinery, or toy — you may be high-risk under Art 6(1) even if you're not on the Annex III list.

Questions to ask your counsel: "Does Art 12 apply to us, and by when?" · "What must we retain, and for how long?" · "What evidence will the regulator expect?" — our plain-English DPO notes →

Verifiable, not guesswork — every figure traces to a primary source

Measured on a live LLM stack

Routes to a human reviewer

On a 1,800-response stratified sample of RAGTruth (17,790 total · 6 models), ISNAD flags 96.0% of hallucinations at 60.4% precision — 70.3% accuracy vs a 55.8% baseline — with the 49.9% false-positive cost disclosed. The point: anything uncertain routes to a human — never silently passed.

Recall ≠ precision — the precision / false-positive sheet is on request. Read the benchmark report ↗

arXiv:2607.24117

cs.AI · 25pp

1,000+ tests

public CI, every commit

pip install isnad

v3.2.0 · Python 3.11+

LangChain

community middleware

Adapters

LangGraph·CrewAI·LlamaIndex·MCP·OTel

Zenodo DOI

10.5281/zenodo.21216872

In plain language: we tested the grading on 575,060 graded chains and it matches the documented rule-based convention (Ibn Hajar's 12 tiers) 87.1% of the time — a benchmark of the method, not a truth score. On live AI answers it flags 96% of bad ones, with the false-positive cost stated honestly.

See it on your own pipeline.

The 2-week assessment maps your stack, grades every hop, and hands you the signed evidence package.

Key takeaways

The short version.

  • 1.ISNAD tags everything that handled an answer — the source, the search, the model, the tool, the agent — and the least-trusted step sets the caution level to one of five actions: serve, caveat, review, quarantine, or reject (and quarantine the narrator).
  • 2.Every claim gets a signed, tamper-detecting audit record — SHA-256 self-hash, HMAC/Ed25519 signature, append-only Merkle log — that an auditor can inspect.
  • 3.Open source, Apache-2.0 (self-host, $0). Validated on 575,060 graded chains and on live AI answers.
  • 4.It grades who handled a claim, not whether it's true — and issues no compliance certificates. It produces evidence supporting EU AI Act Art 12 logging and Art 19 retention (≥6 months).

The problem

Your AI makes claims. You can't prove who handled them.

01

Every AI answer has no record of who handled it

An LLM gives you an answer. Who handled it, in what order, and how much do you trust each one? Observability logs what happened — not who transformed the claim.

02

Multi-agent chains obscure, they don't reassure

A claim that survives five hand-offs isn't more reliable — it's more obscured. Correlated agents that share one bad source amplify the error.

03

The deadline can't be retrofitted

Art 12 logging and Art 19 retention duties bind high-risk systems from 2 Dec 2027. Six months of logs can't be backfilled. Coverage begins at install.

How it works

Three steps to the proof.

Step 1

Tag every AI answer

Every answer gets tagged with where it came from and everything that touched it — the source document, the search, the AI model, the tool.

Step 2

Grade each step simply

Each step gets a grade you can read — Reliable → Caution → Weak → Do not use. The weakest step sets the caution level for the whole answer.

Step 3

Seal it in a locked record

It's sealed in a locked record. If anyone tries to change it later, it shows — and it's kept for as long as the law requires (at least six months).

Fail-closed by default: ambiguous or ungradeable chains are flagged for review, never silently passed. That's the 3.2% scored fail-closed.
The hard part is done: tested on 575,060 graded chains + the sealed evidence layer.
Works on your own computers: we handle the setup during the assessment.

Why ISNAD

Not another hallucination detector.

Most tools ask about the output. ISNAD asks: "who handled this claim, and how much do we trust each step?"

Cleanlab TLM

Grades the trustworthiness of the LLM response — doesn't record the chain or grade who handled it.

Galileo · Patronus · TruLens · RAGAS

Measure output faithfulness vs context — skip the chain and who handled it.

LangSmith · Langfuse · Arize

Record traces (what happened) — don't grade who handled the claim.

OpenLineage · Marquez · DataHub

Track data lineage — don't grade trust or gate on it.

ISNAD

Grades who handled it and the chain — the least-trusted step sets the caution level, duplicate sources are discounted, and the record is locked. One of the few that does all three.

Complementary, not a replacement — ISNAD composes with these tools.

Compliance & security

Evidence infrastructure. Not a certification.

ISNAD issues no certificates. Certification belongs to the auditor; producing the record belongs to us. ISNAD produces evidence artefacts, not compliance. Framework references are informational — any "maps onto / supports" statement requires your counsel's sign-off.

EU AI Act

Art. 12 → logging capability (no retention) · Art. 13 → the chain · Art. 14 → fail-closed events · Annex IV → inputs for technical documentation. Art. 19: logs kept ≥6 months. Retention set by your policy — self-host keeps it on your infrastructure.

ISO/IEC 42001

Versioned registry, per-link hashes, signed records — the controlled, documented evidence an audit collects.

NIST AI RMF

Govern/Map provenance & traceability — every output traces back through its chain to its source.

SDAIA

Transparency, accountability and traceability — a direct implementation of the principles.

Key custody

HMAC-SHA256 (default) / Ed25519 (opt-in) · operator-held key · keys never stored in chain data.

Compromise playbook

Revocation is a signed event inside the log — not a rewrite. Tamper-detecting audit + admin-action audit.

Published threat model

Sleeper narrators, Sybil, grade-import, tail-truncation, forger limits — documented in the repo ↗.

Signed = developer-held key, not third-party audited. We state exactly what's done vs in progress.

What it looks like in your world

One locked record per answer — no jargon, no Greek, no hashes. The folder you hand the auditor.

A loan decision

"This client is eligible." — from the client file + policy doc, handled by the AI model, reviewed by Sarah. Trust: Caution. Record locked.

An insurance claim

"This claim is approved." — from the policy + photos, handled by the AI model. Trust: Reliable. Record locked.

Legal advice

"This clause is enforceable." — from the contract + precedent, handled by the AI, reviewed by counsel. Trust: Caution. Record locked.

Accounting advice

"This expense is deductible." — from the invoice + tax code, handled by the AI model. Trust: Reliable. Record locked.

The fail-closed path

"This client is low-risk." — but one source was flagged untrusted. Trust: Do not use → quarantined, not served. Record locked.

Who it's for

Built for the person who answers the auditor.

Compliance / DPO

"What did the AI do, and how do you know?" — you hand over the folder.

You own the EU AI Act clock: Art 12 logging and Art 19's six-month retention can't be backfilled, and 2 Dec 2027 is coming. ISNAD produces the signed, locked evidence bundle and the Art 12/19 gap map — every record mapped to the article it supports, verifiable offline. Your counsel decides compliance; we make the evidence that lets them.

Owner / ops manager

Not sure this applies to you?

Start with the plain-English check above: does the EU AI Act touch your workflow, and what's the cost of not having the folder on 2 Dec 2027? No jargon — just the answer you need to decide.

CISO

You own the blast radius.

Operator-held keys, an append-only signed log, a compromise playbook, and a published threat model. Full technical detail →

Developers: pip install isnad — Apache-2.0, self-host, $0. See the repo →

Pricing

Start free. Pay for the evidence you need.

Today, by bank transfer — email contact@isnadhq.com for details and an invoice.

Self-host

$0

Apache-2.0, forever

  • pip install isnad
  • Your infra, your data
  • Verify offline, leave anytime
  • Community support
Run the open source ↗

Trust Assessment

$5,000

2 weeks · one pipeline

  • Pipeline map — every step named
  • Graded registry — per-domain trust grades
  • Measured accuracy report
  • Evidence bundle + Art 12/19 gap map
  • Credited to your annual contract
Email to book ↗

Enterprise support

Talk to us

self-host · annual

  • Self-host on your VPC / on-prem
  • Priority support + SLA
  • Evidence-program support (12 months)
  • Annual evidence program — from $12,000/year
Email to talk ↗

The Trust Assessment is a technical readiness review — not an Art. 43 conformity assessment, not a notified body.

FAQ

Honest answers to the hard questions.

What is ISNAD, in one sentence?

ISNAD puts a signed, tamper-detecting label on everything your AI says — who handled it and how much to trust each step — so you can show an auditor where every answer came from.

Is ISNAD "EU AI Act compliant"?

No — and we won't claim it is. ISNAD produces the signed, hash-chained evidence supporting Art 12 logging and Art 19 retention. Whether your deployment is compliant is your counsel's call.

Does ISNAD tell me whether a claim is true?

No. ISNAD grades who handled a claim, not whether it is true. Grades are operator-assigned priors. Refusing to fake a confidence number is the brand.

Where do the benchmark numbers come from?

0.871 is agreement with the documented rule-based convention (Ibn Hajar's 12 tiers) across 575,060 graded chains — not ground truth. The number that matters for your stack is κ = 0.4345 on RAGTruth (96.0% recall; 3.2% fail-closed, 49.9% false-positive cost disclosed).

Can I self-host? What if you disappear?

Yes — the Apache-2.0 core runs on your infra. Your evidence exports in an open format (JSON + HMAC + Merkle proof) you can verify offline. That continuity is part of the design.

Who is behind this?

ISNAD is built in public (Apache-2.0, arXiv:2607.24117, 1,000+ tests). We lead with that, not hide it: published threat model, failure modes, benchmark controls. Contact: contact@isnadhq.com.

Plain words

The terms, in one line each.

ISNAD — Arabic for "chain of narration." The graded chain your claim traveled.
Step — anything that handled the answer: a document, a search, a model, a tool, or a person.
κ (kappa) — agreement score: 1 = perfect, 0 = chance. Reported with limits.
DAIF — Arabic for "weak." A chain grade, not a typo.
HMAC / Ed25519 — the cryptographic signatures that seal each record.
Merkle log — an append-only ledger: if anyone edits history, the hashes break.
Weakest link — the least-trusted step caps the whole chain.
Retriever — the component that fetches source documents for a RAG pipeline.
Tamper-detecting — the record detects tampering; it does not certify that a claim is true.

Prove who handled it.

Attach a signed, hash-chained custody record to every claim your AI makes — before your auditor asks.

contact@isnadhq.com