Govern the knowledge assistant, not just the answer.

1 · What this governs

The assistant answers. The protocol governs the answer as an act.

KYE Protocol™ binds the assistant to a purpose, checks admissibility, and proves the cited basis.

The kind of assistant this governs is a document-QA or retrieval-augmented system: a clinician asks about dosing, an analyst asks about a filing, and the assistant pulls passages from an approved corpus and replies with citations. Products in this category sit close to high-stakes decisions, so the question is not only "is the answer good?" but "was the assistant allowed to answer, and can the basis be replayed?"

KYE Protocol™ answers the second question. It binds the assistant to a purpose and scope, decides admissibility at the boundary, pins each retrieval call, packs the cited sources as evidence, and seals a replay-proof. The result is an act a regulator can re-derive, not a log they must trust.

3 · The envelopes

Six existing schemas, assembled into one chain.

Each envelope validates against the protocol schema it reuses. No new schema was minted.

• Governance binding — the assistant declares purpose, read-only scope, and its single retriever.
• Admissibility decision — twelve per-check verdicts answer whether retrieval is allowed.
• Tool-call pin — the retrieval call, fixed by version and input and output hashes.
• Evidence pack — cited guideline and formulary passages as signed signals.
• Replay-proof — the decision re-runs to a byte-equal result.
• Completion ledger — the actual file reads and tool call, checked against the declared scope.

5 · Verify it yourself

Provable from public keys alone.

Every signed envelope in the chain verifies against the published key set at trust/self-audit-jwks.json — the same Ed25519 verification surface the protocol uses to audit itself. No credentials, no vendor portal: parse the JSON, canonicalise the payload, and check the signature. The self-audit page walks the verifier step by step.

Drop-in onramp

Wrap your RAG answer in a dozen lines.

Keep your retriever and your model. Add KYE Protocol™ at the answer boundary.

Ask decide() whether the answer may proceed and under whose authority, then seal the cited basis as an Evidence Pack™ — verifiable from public keys alone. The free tier runs client-side. The canonical snippets are kye:snippet:govern-rag-ts and kye:snippet:govern-rag-py.

// TypeScript — npm install @kye/sdk
import { KyeClient } from "@kye/sdk";

const kye = new KyeClient({ apiKey: process.env.KYE_API_KEY });

// Your RAG pipeline stays as-is; KYE governs the answer as an act.
export async function answer(question: string, principal: string) {
  const sources = await retrieve(question);
  const decision = await kye.decide({
    purpose: "knowledge.assistant.answer",
    principal,
    inputs: { question, source_ids: sources.map((s) => s.id) },
  });
  if (decision.effect !== "permit") return decision.refusal;

  const draft = await generate(question, sources);
  await kye.evidence.pack(decision.context_seal, { answer: draft, sources });
  return draft;
}

Python parity ships as kye:snippet:govern-rag-py via pip install kye-sdk. KYE Protocol™ never sees your corpus or your model output beyond the source_ids you pass it — it governs whether the answer may proceed, not whether it is true (§63 / §31).