Isnād–Rijāl Framework

Grade the narrators, not just log them. Claim-level provenance for multi-agent knowledge systems — adapted from classical hadith transmission science.

Paper: arXiv:2607.24117 | Paper DOI: 10.48550/arXiv.2607.24117 | Software DOI: 10.5281/zenodo.21216873

🌐 Full project home: https://alizahidraja.com/isnad

What & Why

In modern AI pipelines, a factual claim passes through many hands — a scraper extracts it, a model compiles it, another serves it. Each hand can drop, distort, or invent. Existing tools record what happened. ISNAD grades who transformed the claim, so it can tell you how much to trust the result.

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

60-Second Quickstart

pip install isnad

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

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

# Seed-grade known narrators (required for coverage — see §8 experiment)
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)

# Grade the chain
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)

# Content criticism (now functional — embedding-based)
critic = EmbeddingCritic()
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 (older flat-list reporter with built-in report()) and AsyncIsnadCallbackHandler for async pipelines.

Citation

If you use ISNAD in your research, please cite:

@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},
}

@software{raja2026isnad,
  author  = {Ali Zahid Raja},
  title   = {Isnād–Rijāl Framework: Reference Implementation},
  year    = 2026,
  doi     = {10.5281/zenodo.21216873},
  orcid   = {0009-0003-7875-4590},
}

GitHub users: click "Cite this repository" on the repo sidebar (powered by CITATION.cff).

What's Validated vs. What's Not

The honesty box is a feature. We tell you exactly what works, what's limited, and where you need to supply your own components.

Endpoint identity (model version drift)

For model narrators, grades are keyed by alias@version when a chain link supplies a resolved version (e.g. ingest-model-v3@2.0). Deploying a new model behind the same service name creates a new registry identity — track record does not carry forward. That reset is intentional (paper §4.2).

• Register with model_version via the API or Registry.register_versioned()
• Pass version on each chain link (or use @version in seed keys for LangChain)
• Legacy alias-only registrations still work when version is omitted, unknown, or a non-resolved tag (latest, dev, canary)

Demo: uv run python examples/endpoint_identity_drift_demo.py

Concept → Module Map

🗺️ Full architecture diagram: docs/ARCHITECTURE.drawio — 3 tabs: System Architecture, Claim Lifecycle (data flow), and Validation Matrix (what's proven vs what's not). Open in draw.io or VS Code Draw.io extension.

The Decision Matrix

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 EmbeddingCritic, LLMCritic

critic = EmbeddingCritic()                            # offline, fast
critic = LLMCritic(api_key="sk-...")                  # LLM-backed, higher quality

Good first issues:

• Implement an alternative critic (sentence-transformers embedding, CrewAI integration)
• Seed-grade bootstrapper from published benchmark data
• Extend semantic corroboration to multi-source corpora (ArXiv, textbooks, news)

Trace Capture & Chain Viewer

ISNAD now ships a capture → schema → viewer pipeline that makes transmission chains visible. The contract is isnad_trace v0.1 — a versioned JSON schema aligned with W3C PROV-DM and PROV-AGENT (arXiv 2508.02866).

Capture (LangChain callback)

from isnad.integrations.langchain import IsnadCallbackHandler, seed_registry

reg = seed_registry({"source:my-docs": "reliable", "model:gpt-4o": "acceptable"})
handler = IsnadCallbackHandler(registry=reg, domain="physics")

# Attach to any LangChain/LangGraph pipeline
chain.invoke("What is F=ma?", config={"callbacks": [handler]})

trace = handler.to_trace()
print(trace.model_dump_json(indent=2))  # isnad_trace v0.1

Key properties:

• Tree from run_id/parent_run_id — LangChain already provides the tree; the handler reconstructs it without timestamps or heuristics.
• Input provenance — every retrieved document is recorded with source, doc_id, and content hash. Full content is redacted by default.
• Model version captured — resolved model version (e.g. gpt-4o-2024-08-06) from response metadata, not the endpoint alias. Records null explicitly when unavailable.
• Shared ancestry detection — overlapping retrieval sets, shared model families, or shared upstream sources are flagged as shared_ancestry_detected.
• Never breaks your pipeline — every callback is wrapped in try/except.
• Async support — AsyncIsnadCallbackHandler for async LangChain runs.

Runnable demo (no API keys required):

python examples/isnad_langchain_demo.py

Viewer

Open viewer/index.html in a browser. Three hand-built fixtures demonstrate the framework's key signals:

The viewer renders fixture 3 by default — it is the most important case.

What the viewer shows — and what it doesn't

Validated signals (mechanisms confirmed empirically or structurally):

Indicative signals (displayed honestly, not validated):

What is NOT shown:

• No numeric confidence (never 87.3). Grades are ordinal bands: ṣaḥīḥ/ḥasan/ḍaʿīf/mawḍūʿ.
• No colour-alone confidence encoding. The viewer respects prefers-reduced-motion and visible keyboard focus.
• No collapsed axes. Chain integrity and origin strength are always separate.
• No implicit corroboration. unverified independence is not rendered as agreement.

Experimental Validation — Semantic Corroboration (§8)

Status: Empirically validated on real data.

The framework's most distinctive contribution — independent-chain corroboration (mutābaʿāt) — has been validated on two corpora of increasing difficulty:

Results

Combined: 603 + 104 = 707 claim pairs tested across both corpora. 100% corroboration fire rate. 8/8 negative controls. Zero false positives. Source URLs for every claim.

Key Findings

1. Corroboration fires on semantically-matched cross-source data — different text, same meaning, genuinely independent sources
2. Independence detector correctly identifies madār — chains with shared model families or upstream sources are blocked from upgrade
3. HASAN cap enforced — corroboration never falsely reaches SAHIH
4. Information-theoretic weights calibrated — effective weight scales with chain quality (1.5 for single HASAN corroborator, 2.5+ for multiple)

Reproduce

cd experiments/corroboration_v2
pip install isnad sentence-transformers scikit-learn requests
python run.py               # Fetch Wikipedia + semantic match + corroborate
python negative_controls.py  # Verify all 8 gates
python analyze.py            # Statistical analysis

Full methodology, results, negative controls, and paper gap analysis in:

• experiments/corroboration_v2/README.md
• experiments/PAPER_GAP_ANALYSIS.md

Ecosystem

• 🌐 Site: https://alizahidraja.com/isnad
• 📄 Paper (arXiv): https://arxiv.org/abs/2607.24117
• 📄 Paper (DOI): https://doi.org/10.48550/arXiv.2607.24117
• 💾 Software (DOI): https://doi.org/10.5281/zenodo.21216873
• 📦 PyPI: https://pypi.org/project/isnad/
• 📝 Companion gist: https://gist.github.com/alizahidraja/56beaadf493976182f38aa602b8958e2
• 🧪 §8 Experiment & results: experiments/s8_gated_vs_ungated/
• 🔬 Semantic Corroboration v2 (Wikipedia): experiments/corroboration_v2/
• 📚 Corroboration v3 (Physics Textbooks): experiments/corroboration_v3/ — 104/104 on s8 corpus
• 🗺️ Architecture Diagram: docs/ARCHITECTURE.drawio — 3 tabs: System, Data Flow, Validation Matrix
• 🔌 LangChain integration: src/isnad/integrations/langchain/
• 🔗 Trace schema spec: docs/trace-schema.md — v0.1 with PROV/PROV-AGENT mapping
• 👁️ Chain viewer: viewer/index.html — open in browser, renders all 3 fixtures
• 🧪 Trace capture demo: examples/isnad_langchain_demo.py — runnable without API keys
• 📊 Critic evaluation: src/isnad/critics/CRITIC_EVAL.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},
}

@software{raja2026isnad,
  author  = {Ali Zahid Raja},
  title   = {Isnād–Rijāl Framework: Reference Implementation},
  year    = 2026,
  doi     = {10.5281/zenodo.21216873},
  orcid   = {0009-0003-7875-4590},
}

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. Built in public — collaborators welcome.

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