diff --git a/docs/sdk-rfc.md b/docs/sdk-rfc.md
new file mode 100644
index 0000000..2b1e740
--- /dev/null
+++ b/docs/sdk-rfc.md
@@ -0,0 +1,80 @@
+# RFC: Hedwig SDK surfaces
+
+Status: Draft. Date: 2026-07-06. Discussion via GitHub issues.
+
+---
+
+## Problem
+
+An integrating program should be able to gate an instruction on "does this pubkey hold role X in org Y" without reading Hedwig's Rust source. Today Hedwig ships 5 instructions and a devnet lifecycle demo (`app/demo.ts`), proving the primitive works end to end. Missing: an importable surface. No crate, no package, nothing a third-party program can pull in without copying PDA derivation and account layouts by hand.
+
+---
+
+## Decisions already made
+
+- **Roles stay flat in the core program, forever.** Hierarchy composes on top via wrapper programs, since `Role.admin` can be any pubkey, including a PDA owned by another program. The core program never grows a tree-walking authorization model.
+- **Trust path is freeze-track, not multisig.** Mainnet v1 deploys intending to burn the upgrade authority after the invariant test suite and an external review pass. A small, finished program is easier to freeze and audit than one kept "open" indefinitely.
+- **First consumer is a dogfood demo, not a partner integration.** An autonomous trading agent, built in a separate repo (`hedwig-agent-demo`), holds a time-boxed "trader" role and a "settler"-gated settlement path. Building the first real consumer in-house surfaces SDK gaps before any external team hits them.
+- **Token-2022 badge issuance is deferred.** It is a presentation layer on top of role state, not part of the roles primitive, and does not block anything here.
+
+---
+
+## Proposed surfaces
+
+### hedwig-cpi (Rust crate)
+
+The primary surface, since Hedwig's consumers are programs, not end users. Exposes two helpers:
+
+- A `check_role` CPI helper for Anchor programs, wrapping the existing `check_role` instruction.
+- A direct read-and-verify helper that deserializes the Member and Role PDAs in-process and runs the same checks (expiry, `enabled` flag, PDA derivation) without a CPI call.
+
+The open cost question: CPI carries fixed compute and call-depth overhead a direct account read does not. Some consumers may prefer the cleaner interface; others may be compute-constrained. This RFC does not pick a default. The `hedwig-agent-demo` build measures both paths and the crate README documents the numbers once known.
+
+### @hedwig-sol/sdk (TypeScript client)
+
+Secondary, built after the crate. PDA derivation helpers and instruction builders for the create/assign/revoke flows, the operations `app/demo.ts` already performs by hand. A thin wrapper over the Anchor IDL, nothing more: no indexer, no caching layer, no off-chain state.
+
+### Worked example (hedwig-agent-demo)
+
+A separate repo with an autonomous trading agent whose signing key holds a "trader" role with `expires_at` set (a session-key pattern: the key stops working on schedule, no revoke transaction needed) and whose settlement path checks a "settler" role. The org authority can flip `Role.enabled` to false as a kill switch, halting every check for that role without touching individual memberships. This demo validates both patterns and produces real CPI-vs-direct-read compute numbers instead of estimates.
+
+---
+
+## Explicit non-goals
+
+- Role hierarchy or inheritance inside the core program.
+- An indexer service for role state or history.
+- Token-2022 badge issuance (future, separate layer, does not touch core state).
+- Multisig upgrade governance (superseded by the freeze-track plan above).
+- Any off-chain permission cache. Role state is read onchain, always.
+
+---
+
+## Non-obvious decisions
+
+- Flat-forever makes the freeze-track plan credible: a program that never grows new authorization logic is a program that is actually finished, and a finished program is one worth freezing.
+- The CPI crate outranks the TypeScript SDK in priority because Hedwig's target integrators are other Solana programs calling in via CPI, not end users clicking a UI. The TS client matters for tooling, but it is not the adoption bottleneck.
+- The CPI helper and the direct-read helper ship in one crate, not two, because they answer the same question (does this pubkey hold this role) and choosing between them is a compute tradeoff, not a separate integration decision.
+
+---
+
+## Open questions
+
+- CPI call overhead versus direct PDA deserialization: exact compute unit delta, answered by the `hedwig-agent-demo` build.
+- Rent-return recipient on `revoke_role`: confirm current behavior is correct, don't assume it, audit during the demo build.
+- Whether any client needs event or indexer support before the v1 freeze, or onchain-only reads stay sufficient through mainnet.
+
+---
+
+## Milestones
+
+Sequenced, not dated. Cadence is weekly ships.
+
+| Step | Deliverable |
+|---|---|
+| 1 | `hedwig-agent-demo` repo: trading agent with time-boxed trader role, settler-gated settlement |
+| 2 | Friction report from building the demo against raw CPI calls |
+| 3 | `hedwig-cpi` v0: CPI helper + direct-read helper, informed by the friction report |
+| 4 | `@hedwig-sol/sdk` v0: PDA derivation + instruction builders |
+| 5 | Invariant test suite (per THREAT-MODEL.md M1 commitment) |
+| 6 | Mainnet deploy, then freeze |