Aly Sawft · Founder & Engineer, Sawftware LLC · · 7 min read
A compliance officer asks her firm's contract-review agent: "Does Section 4.2 allow unilateral termination?"
The agent answers confidently: "Yes — the agreement permits either party to terminate with 30 days' notice."
Three weeks later, opposing counsel produces the actual PDF. Section 4.2 says nothing about unilateral termination. Nobody can show what the agent actually read. There is no bundle ID, no manifest hash, no signature tying the answer to a specific byte sequence.
This is not only a model-quality problem. It is a provenance problem. The agent gave an answer. It did not give a citation anyone can verify.
Retrieval-augmented generation (RAG) improved agents dramatically. But retrieval answers a different question than verification:
| What RAG gives you | What RAG cannot prove |
|---|---|
| Relevant text chunks | That chunks match a specific file version |
| A plausible synthesis | That the synthesis quotes the source faithfully |
| Lower hallucination rate | Cryptographic proof of what was read |
| Fast semantic search | Offline audit without trusting your stack |
When a regulator, judge, or enterprise security team asks "prove what the model saw," an embedding similarity score is not enough. They need evidence that survives dispute.
A useful citation for agent workflows includes:
With those pieces, a third party can download the evidence bundle, recompute hashes, verify the signature, and confirm the agent's answer referenced real content — offline, without ongoing API access.
Standards: SHA-256 (NIST FIPS 180-4) for artifact hashing; EIP-191 for manifest signatures.
That is qualitatively different from "we retrieved chunk 47 from Pinecone."
If you ship document-heavy agents today, ask:
If the answer is no, you have a recall system, not a verifiable memory layer. For compliance, finance, and legal — that is not enough.
Agents need citations that carry proof: signed bundles, not just retrieved text. See our manifest hash checker and evidence bundles pillar for the capture model.