ISNAD — Open-Source LLM Provenance & AI Audit Trail for RAG and Multi-Agent Systems

Isnād–Rijāl Framework · grades the transmitters and the chain, not just the claim.

Open-source pip install isnad — LLM provenance, agent trust, and an AI audit trail for RAG and multi-agent systems. Apache-2.0, permanently.

Every claim your pipeline produces carries a verifiable weakest-link trust grade and a tamper-evident audit trail, so you can answer the question hallucination detectors and observability tools skip: who handled this claim, in what order, and how much do we trust each one?

The one-sentence pitch: ISNAD gives you the tamper-evident audit records your regulators will demand — a SHA-256 hash for every graded claim, mapped to the EU AI Act, ISO/IEC 42001, NIST AI RMF, and SDAIA. You pass your audit. 📋

Proof it works: ISNAD's weakest-link rule reproduces 1,200 years of scholar chain verdicts at Cohen's κ = 0.87 across 575,060 graded hadith chains. (For context: scholars disagree with each other on narrator grades at κ = 0.33 — a different task.) Full benchmark below.

# 30 seconds to a verdict:
from isnad import Registry, grade
from isnad.types import NarratorGrade

reg = Registry()
reg.register("openstax", "physics", grade=NarratorGrade.RELIABLE)
reg.register("pdf-scraper", "physics", grade=NarratorGrade.UNGRADED)
reg.register("ingest-model", "physics", grade=NarratorGrade.ACCEPTABLE)

verdict = grade("p = mv", ["openstax", "pdf-scraper", "ingest-model"], reg, domain="physics")
print(verdict.why)
# claim 'p = mv' → chain DAIF (weakest: pdf-scraper, ungraded)

Paper: arXiv:2607.24117 · Software DOI: 10.5281/zenodo.21216873

🌐 Project home: https://alizahidraja.com/isnad · 📚 Docs: https://alizahidraja.com/isnad/docs · ⭐ Star to follow along · Contribute · Code of Conduct

JavaScript/TypeScript verifier: npm install isnad verifies ISNAD audit records (SHA-256 / Merkle / detached signatures) from JS — verifier only, grading stays in the Python core. See js/README.md.

What & Why

In a RAG pipeline or multi-agent system, a factual claim passes through many hands — a scraper extracts it, an agent compiles it, another model serves it. Each hand can drop, distort, or invent. Observability and data-lineage tools record what happened; ISNAD grades who transformed the claim, so you know how much to trust the result — and can export the whole judgment as a tamper-evident audit record for governance review.

The framework adapts hadith transmission science — one of history's most rigorous epistemologies, refined over twelve centuries — into a Python library. Every claim carries its complete chain of transmitters (isnād); each transmitter is graded in a living registry (rijāl); chains are capped by their weakest link; content is criticized independently of transmission; and the two combine in a decision matrix that routes claims to serve, review, or quarantine.

Related MCP server: zetesis

How ISNAD differs from a hallucination detector

Most tools ask about the output — "is this answer grounded in the retrieved context?" (Cleanlab TLM, Galileo, Patronus, TruLens, RAGAS) or "what happened during this run?" (LangSmith, Langfuse, Arize). ISNAD asks the question they skip: "who handled this claim, and how much do we trust each transmitter?"

What ISNAD is NOT — and why that's the point. Three honest answers, up front:

1. "Tracers already record who handled it." They record the order; they don't grade each transmitter, propagate a weakest-link verdict, or gate on it.

2. "Your grades are operator-assigned, so it's GIGO." The grades are priors + audited observations + independent corroboration, propagated by a documented method. ISNAD makes trust explicit and auditable — it doesn't auto-measure it, and it never pretends to.

3. "Claim critics measure truth better than you." Correct — and ISNAD composes with them (matn criticism). It doesn't replace them; it grades the chain that carried the claim.

Free, Apache-2.0, self-hostable — permanently. The core (src/isnad/) stays Apache-2.0 forever; any future commercial surface lives in a separate repo under its own licence (see LICENSING.md). The claim-graders above are paid SaaS APIs; ISNAD is open-source and self-hostable.

Quickstarts

30 seconds — grade a claim

pip install isnad

from isnad import Registry, grade
from isnad.types import NarratorGrade

reg = Registry()
reg.register("openstax", "physics", grade=NarratorGrade.RELIABLE)
reg.register("pdf-scraper", "physics", grade=NarratorGrade.UNGRADED)
reg.register("ingest-model", "physics", grade=NarratorGrade.ACCEPTABLE)

verdict = grade("p = mv", ["openstax", "pdf-scraper", "ingest-model"], reg, domain="physics")
print(verdict.why)
# claim 'p = mv' → chain DAIF (weakest: pdf-scraper, ungraded)

60 seconds — full chain + content criticism + decision matrix

pip install isnad

from isnad import Registry, Chain, ChainLinkSpec, grade_chain, decide
from isnad.types import NarratorGrade, ContentVerdict
from isnad.critics import EmbeddingCritic

chain = Chain([
    ChainLinkSpec("openstax-textbook", 0, domain="physics"),
    ChainLinkSpec("pdf-scraper-v2", 1, domain="physics"),
    ChainLinkSpec("ingest-model-v3", 2, domain="physics"),
])

reg = Registry()
reg.register("openstax-textbook", "physics", grade=NarratorGrade.RELIABLE)
reg.register("pdf-scraper-v2", "physics", grade=NarratorGrade.RELIABLE)
reg.register("ingest-model-v3", "physics", grade=NarratorGrade.ACCEPTABLE)

grades = [reg.get_grade(l.narrator_id, l.domain) for l in chain.links]
transforms = [l.transform_type for l in chain.links]
chain_grade = grade_chain(grades, transforms, is_complete=True)

critic = EmbeddingCritic()  # offline; obvious contradictions only — see Scope
verdict = critic.evaluate("p = h/λ", "p = h/lambda", ["p = mv"])
action = decide(chain_grade, verdict)

print(f"Chain: {chain_grade.value.upper()} | Content: {verdict.value} | Action: {action.value}")

LangChain integration (5 lines)

pip install isnad[langchain]

from isnad.integrations.langchain import IsnadCallbackHandler, seed_registry

reg = seed_registry({"source:docs": "reliable", "model:gpt-4o": "acceptable"})
handler = IsnadCallbackHandler(registry=reg, domain="physics")
chain.invoke("What is F=ma?", config={"callbacks": [handler]})
trace = handler.to_trace()  # isnad_trace v0.1 JSON

Also available: IsnadTracer and AsyncIsnadCallbackHandler for async pipelines.

IsnadMiddleware (gating). The callback captures traces; the middleware gates claims. Note, honestly: the core gate() function is importable and testable today, but the LangChain AgentMiddleware wiring is forward-looking (it targets LangChain's 2026 AgentMiddleware API; see src/isnad/integrations/langchain/middleware.py). The callback handler is the production-ready capture path.

from isnad.integrations.langchain import IsnadMiddleware
mw = IsnadMiddleware(reg, domain="physics")  # wrap_tool_call grades + gates

Want to ship this to a real system this afternoon? See docs/onboard-in-a-day.md for three copy-paste recipes (self-maintaining KB, medical RAG, legal RAG) and the honesty contract that must not be broken.

Documentation

Read in order, or jump to what you need:

• 🚀 Quickstart — add ISNAD to your LangChain agent in 5 minutes.

• 💡 Concepts — how isnād–rijāl maps to multi-agent provenance.

• 📋 Compliance — EU AI Act · ISO/IEC 42001 · NIST AI RMF · SDAIA.

• 📊 Model-drift leaderboard — measured hallucination vs. chain depth.

• 🗺️ Architecture — the four loops and the module map.

What's Validated vs. What's Not

The honesty box is the point: what's proven, what's measured, and what's open.

Honest limits (stated up front): grades are ordinal-only (ṣaḥīḥ > ḥasan > ḍaʿīf, never a numeric confidence); cold-start is worse per-role; integrity is domain-scoped, not global; chain independence cannot be proven from topology (only assumed from attested lineage, then discounted — #54); the content critic is the coverage ceiling; corroboration rarely fires on dense technical corpora where genuine cross-source overlap is rare. These are features, not bugs — ISNAD says exactly what works, what's limited, and where you supply your own components. The serving index is in-memory (single-node deployment; multi-worker HA is not yet supported — #199).

ISNAD-Bench — measured against 1,200 years of ground truth

The strongest evidence is a number, not a claim: ISNAD's weakest-link rule, run on 575,060 graded hadith chains (each graded by classical scholars), measured for agreement:

Not the same scale. κ = 0.871 measures chain-verdict agreement (4 classes); the 0.450 and 0.331 figures measure narrator-grade agreement (5 classes) — a different task. The 0.331 is context for how contested the underlying narrator grades are, not a direct ceiling on the chain κ. How to read it: ISNAD reproduces the scholars' consensus at κ = 0.87 — not because it is "better than the scholars" (they disagree with each other at 0.33), but because it deterministically implements their method. The benchmark is preregistered, carries negative controls (majority-class 0.000; shuffled grades −0.007), and buckets every disagreement. Full write-up: bench/docs/RESULTS.md.

uv run python -m bench.run               # strict (default), full corpus
uv run python -m bench.run --lenient     # ungraded → ḥasan
uv run python -m bench.human_ceiling     # the human ceiling
uv run python -m bench.ikhtilat          # the mukhtaliṭūn (period-sliced grades)

The dataset (emadjumaah/hadith-kg, CC-BY-4.0, 1.6 GB) is gitignored and pinned by SHA-256 — see bench/README.md. The derived graded output (per-chain scholar verdict + ISNAD prediction + disagreement bucket) is published as the alizahidraja/isnad-bench dataset on Hugging Face.

Transfers to AI provenance — measured on RAGTruth

The κ = 0.871 above is the methodology's home turf (hadith chains). The transfer to live LLM pipelines is measured separately and published honestly: on RAGTruth (17,790 responses, 6 models), ISNAD's grounding critic catches 97.6% of hallucinations at κ = 0.575 (79.3% accuracy vs a 71.4% baseline) and grades the six models' reliability in exactly the right order. The weak bi-encoder signal (κ ≈ 0.1) is reported alongside, not hidden. Case study + limits: G1_CASE_STUDY.md.

Live product: isnad.islamandai.com

Glossary — the Arabic, in plain English

The name and lineage are the reason to trust the design; you don't need the Arabic to use it. Every term is defined on first use.

The Decision Matrix

Chain grade × content verdict → action:

Two defaults to notice: contradictions go to a human by default (LLMs are bad at reconciling competing evidence), and ṣaḥīḥ × contradiction is the highest-value signal, not an error state.

Two serve-side gates apply on top of the matrix (both demote, never promote):

• Prior-only gate — a narrator graded from a population prior alone (seeded, zero observed instances) can SERVE WITH CAVEAT at most; plain SERVE requires observation- or human-backed narrators. High-stakes domains opt into a hard gate (ISNAD_SERVE_GATE=hold or ISNAD_SERVE_HOLD_DOMAINS=medical,legal) → REVIEW.

• Hold-mode UNVERIFIABLE — in a hold domain, a sound chain with unverifiable content is held (REVIEW), never caveat-served.

Pluggable Strategies — Extend It

The framework leaves key parameters open by design (paper §4.2/§4.3). Swap any:

Swap a critic in one line:

from isnad.critics import best_available_critic, EmbeddingCritic, LLMCritic

critic = best_available_critic()                 # LLM if configured, else NLI, else TF-IDF
critic = best_available_critic(prefer_llm=False) # force an offline critic
critic = LLMCritic(provider="ollama", model="llama3.1")  # local — no API key, no cloud
critic = LLMCritic(provider="openrouter", model="openai/gpt-4o-mini")  # any LLM

Critic tiers are documented with their measured recall in docs/critics.md. The LLM critic is provider-agnostic (OpenRouter, OpenAI, DeepSeek, Anthropic, Gemini, Groq, Together, Ollama, or any OpenAI-compatible endpoint). For best results use the LLM critic — but see the honesty note above: it is not an acceptance gate on content corruption (#126).

Cold-start bootstrapping (#33). A new pipeline starts with every narrator UNGRADED, so nothing serves. Seed grades are the fix — evidence-backed priors (Registry.seed records BOOTSTRAP_SEED evidence):

from isnad import Registry, NarratorGrade, seed_from_benchmark

reg = Registry()
reg.seed("source:openstax", "physics", NarratorGrade.RELIABLE, source="publisher")
seed_from_benchmark(reg, "model:gpt-4o@v1", "physics", 0.93, benchmark="mmlu")  # accuracy → grade

Audit & Compliance Evidence

ISNAD can export a tamper-evident audit record for any graded claim:

isnad export --claim <id> --format json    # or jsonl|csv; --verify; --sign; --redact; --chain-log

Each record captures the full chain (as an explicit DAG via upstream_ids), the weakest link, source-document hashes, human-oversight evidence, the environment, and a SHA-256 integrity hash over the RFC 8785-canonical form of its payload — plus an optional detached signature (HMAC/Ed25519), a tamper-evident hash chain, and a PII-redaction hook.

Serving-side signing: the API signs records with a detached HMAC only when ISNAD_HMAC_SECRET is set; when unset, records are emitted audit_signed: false (a self-hash a forger can recompute). Set it in any deployment that relies on tamper-evidence — see .env.example.

Two tamper-evidence logs: a linear hash chain (single sequential writer) and a Merkle batch log (mass parallel agents; build_batch/seal_batches/ verify_batches/prove_inclusion). Both detect modification, middle-deletion, and reordering; neither detects tail truncation, and both commit to a self-hash, so they detect post-hoc modification — not a forger who rebuilds the chain (#97).

from isnad.audit import build_audit_record
from isnad.audit.canonical import canonical_hash

record = build_audit_record(claim_id, session, registry)  # -> AuditRecord
assert record.integrity.record_hash == canonical_hash(record.to_dict(include_integrity=False))

ISNAD produces evidence artifacts; it does not confer conformity with any regulation. See docs/evidence-mapping.md for an informational mapping of each field to the EU AI Act, ISO/IEC 42001, the NIST AI RMF, and SDAIA — and the explicit statement of what ISNAD deliberately does not provide. The core library is Apache-2.0, permanently (LICENSING.md).

Scope and limitations

ISNAD grades post-entry provenance — who handled a claim after it entered your pipeline, and how much you trust them — not whether the upstream source or the final answer is objectively true.

What ISNAD covers: ordered chains + narrator grades; weakest-link grading + the decision matrix; content criticism against your corpus; corroboration when independent chains agree.

What ISNAD does not cover:

• Source legitimacy at the boundary — grades are operator-assigned assertions, not auto-verified facts.

• Novel claim truth — new, unverifiable claims are not auto-fact-checked.

• Faithful transmission of bad input — a RELIABLE fake publisher + a clean pipeline can still yield a high chain grade.

Boundary vetting is the operator's responsibility — vet sources before registering them as RELIABLE; register unknowns as UNGRADED; use matn, corroboration, and human review as downstream gates, not a replacement for it.

Integrations

• LangChain / LangGraph — callback handler + tracer + middleware (above).

• CrewAI / LlamaIndex — adapters (isnad.integrations.crewai, .llamaindex).

• OpenTelemetry — isnad ingest --otlp grades an existing GenAI trace (#73).

• MCP — grade MCP servers as narrators: MCPToolObserver records tool calls as TOOL narrator links and grades the resulting chain; a grade_claim tool exposes the operator's registry to an agent (isnad.integrations.mcp, #59). isnad mcp runs a real FastMCP server over stdio (pip install isnad[mcp]). Honest limit: tool narrators stay UNGRADED by default (no auto-grading from call volume — that would be GIGO); the server returns operator-assigned grades, never manufactured ones.

• Live Verify — consume a verify: cryptographic seal as a high-trust narrator, anchoring ʿadālah on day one (solves cold-start; see examples/issuer_demo/). Honest limit: Live Verify proves authenticity, not truth.

A chain viewer (viewer/index.html) renders transmission chains in the browser — three fixtures including the most important one: false corroboration (five transmitters across three chains, all tracing to one source) rendered as a warning, not consensus.

Experiments

Every headline number is reproducible from the repo (mostly no API keys):

• Adversarial benchmark — narrator grading 100% caught / 0 false positives; content criticism is the binding constraint (experiments/adversarial_benchmark/).

• §8 gated-vs-ungated — 20,000 claims; weakest-link quarantine validated; matched-coverage inconclusive (honest) (experiments/s8_gated_vs_ungated/).

• Semantic corroboration — 707 claim pairs, 8/8 + 9/9 negative controls (experiments/corroboration_v2/, corroboration_v3/).

• A/B demonstration — 2 caught, 2 missed (honestly), 0 false positives (experiments/verified_vs_unverified/).

Open problems

• Chain independence (#54) — the detectable half is engine-wired (attested lineage, document-hash madār, tawātur discount, witness-type priors, content-level madār); the undetectable half (correlated training data across model families) remains an open, stated limit.

• Content critic is the coverage ceiling (#34) — a semantic critic that returns CONSISTENT on real prose is the highest-value component. The numeric-aggregate slice is now deterministic (RecomputeCritic/EnsembleCritic/AggregateRouter); general semantic criticism is still LLM/NLI-tier.

Contributing

Built in public — collaborators welcome. The on-ramp: CONTRIBUTING.md (branch flow, quality gates, code map) · CODE_OF_CONDUCT.md · SECURITY.md (honesty is a security property) · THREAT_MODEL.md · good first issues.

Ecosystem

• 🌐 Site: https://alizahidraja.com/isnad · 📄 Paper: arXiv:2607.24117 · 🤗 HF Papers: 2607.24117 · 💾 Software DOI: 10.5281/zenodo.21216873 · 📦 PyPI: isnad

• 🚀 Docs — start here: Quickstart (5-minute LangChain) · Concepts (isnād–rijāl → multi-agent) · Compliance (EU AI Act · ISO 42001 · NIST RMF · SDAIA) · Model-drift leaderboard

• 🗺️ Architecture: docs/ARCHITECTURE.drawio · 🔗 Trace schema: docs/trace-schema.md · 👁️ Chain viewer: viewer/index.html

• 🧪 Benchmark: bench/docs/RESULTS.md · 🤗 ISNAD-Bench dataset: alizahidraja/isnad-bench · 📊 Critic eval: docs/critics.md · 🕵️ xz sleeper-narrator case study: docs/case-study-xz-sleeper-narrator.md

Citation

@article{raja2026grading,
  author  = {Ali Zahid Raja},
  title   = {Grading the Narrators: An Isnād–Rijāl Framework for
             Claim-Level Provenance in Multi-Agent Knowledge Systems},
  year    = 2026,
  doi     = {10.48550/arXiv.2607.24117},
  eprint  = {2607.24117},
  archivePrefix = {arXiv},
}

About

Built by Ali Zahid Raja · ORCID 0009-0003-7875-4590. The rigor belongs to twelve centuries of muḥaddithūn; the transfer to AI systems is the contribution claimed here.

License: Code — Apache 2.0 · Paper & docs — CC BY 4.0