diff --git a/docs/INTEGRATION.md b/docs/INTEGRATION.md
new file mode 100644
index 0000000..f21ab13
--- /dev/null
+++ b/docs/INTEGRATION.md
@@ -0,0 +1,115 @@
+# Building on provework — the integration guide
+
+prowork is a **base**: a working verification substrate that other
+projects build applications on. This document is for the builder who
+wants to do exactly that. It maps the interfaces you touch, the
+guarantees you inherit, and the limits you accept.
+
+---
+
+## The one-paragraph model
+
+A **job** is a deterministic integer-only program (RISC-V, or V2:
+SP1-native) plus an input blob. **Workers** execute it; a
+**coordinator** dispatches, verifies, and keeps a ledger. Every
+execution reduces to a journal, and every verification tier consumes
+the same commitment. You write the program; you choose the trust
+tier; the substrate handles dispatch, transport, verification,
+fraud handling, and evidence.
+
+## The four interfaces you touch
+
+### 1. The job ABI (write programs against this)
+
+- **Legacy (`rv-abi`):** bare-metal RV64IMC. Input at `0x1000_0000`
+  (u64 length, then bytes), output at `0x2000_0000` (u64 length,
+  then bytes), image base `0x8000_0000`, stack top `0x83F0_0000`,
+  halt via `ebreak`. Pinned ISA: `rv64imc` — no atomics, no floats,
+  no syscalls. Scaffold: `jobkit new`.
+- **V2 (`sp1-v2`):** an SP1-native guest — read input via
+  `sp1_zkvm::io::read`, commit `blake3(input)` then your output via
+  `sp1_zkvm::io::commit`. Runs directly on the zkVM (no emulator,
+  ~94× cheaper to prove), executable only by SP1 fleets. Scaffold:
+  `jobkit new <name> --v2`.
+
+Determinism is the contract: the same input must always produce the
+same output on any host. Anything that breaks that (clocks, threads,
+floats, randomness) is outside the ABI by design.
+
+### 2. The journal contract (what consensus and receipts commit to)
+
+- Legacy: `journal = u64_le(instructions) || output`; commitment =
+  `SHA-256(journal)`. The instruction count is architectural state
+  (rvcore's count) and IS bound.
+- V2: `journal = u64_le(0) || output`; commitment = `SHA-256(journal)`.
+  SP1 cycle counts are a compiler/SDK artifact and are deliberately
+  NOT bound. V2 receipts commit `blake3(input)` as the job binding.
+- Workers sign the full result (ed25519 over all fields); the
+  coordinator's judge reproduces the identical journal bytes.
+
+### 3. The wire protocol (build your own client or coordinator)
+
+Length-prefixed JSON frames over TCP or TLS 1.3
+(`crates/wire`). Client roles authenticate by Ed25519 nonce
+signature plus per-connection admission proof-of-work. Messages:
+`Hello` (with role worker/submitter), `Nonce`, `NonceSignature`,
+`JobAssignment`, `JobResult`, `BlobRequest`/`Blob` (peer-to-peer),
+`BlobUpload`/`BlobAck` (cross-machine submissions),
+`JobSubmission`/`SubmissionAck`/`JobOutcome` (submitters),
+`ReceiptClaim` (zk tier). TLS: the coordinator generates a
+self-signed certificate; clients pin its BLAKE3 fingerprint
+(`wire::tls::client_stream_pinned`).
+
+### 4. The trust-tier dial (what you choose per submission)
+
+| Tier | How | When |
+|---|---|---|
+| Quorum (N=3→5) | default dispatch | cheap jobs, fraud tolerated |
+| Optimistic + dispute | escalation/reserves | 1× cost, judged on challenge |
+| zk receipt | `submit --require-zk` | high assurance; proof replaces consensus |
+
+The receipt queue deduplicates by
+`BLAKE3(manifest_id ‖ elf_id ‖ input_id)` — identical content proves
+once. Receipts verify offline (`jobkit verify`), no coordinator
+required.
+
+## Guaranteed vs. your responsibility
+
+**Inherited from the substrate:** dispatch and quorum anchoring,
+identity + admission PoW, TLS with fingerprint pinning,
+content-addressed blob integrity, fraud detection + bonded slashing,
+replay/zk judging, evidence bundles, receipt verification.
+
+**Your responsibility:** program determinism (the ABI cannot enforce
+it for arbitrary logic — pin your dependencies, avoid nondeterminism),
+input and output *semantics* (the substrate verifies the execution
+happened, not that your program computes what you meant), operating a
+coordinator (it is the job owner's agent), and any multi-tenant or
+payment layer above the ledger.
+
+## Measured envelopes (see docs/PREPRINT.md, Tables 1–4)
+
+- Legacy job through the emulator-in-guest: 322–362× cycle
+  multiplier; CPU-provable to ≈28–31K job instructions on a 32 GB
+  host; 168.9M-cycle job proves in 34.2 s on a 32 GB Blackwell GPU.
+- V2 native: 1.79M cycles for the same 32 KiB job (94× fewer);
+  2.8 s GPU / 80.5 s CPU; ≈1.4M job instructions CPU-provable on
+  32 GB.
+- Interpreter throughput: up to 57.7M instructions/s (2 MiB job).
+
+## Declared limits (do not build on these yet)
+
+- Single coordinator (no decentralized coordination).
+- No NAT traversal — fleets over LAN or reachable addresses.
+- Native V2 jobs require an SP1 fleet (Linux workers); legacy jobs
+  run everywhere.
+- Receipts bind the exact SP1 toolchain version; upgrades re-key the
+  verifying set.
+- The ledger's economics (+10/−100, PoW dial) are calibrated
+  placeholders, not a market.
+
+## License
+
+MIT OR Apache-2.0 for the entire tree (see
+docs/LICENSE-STRATEGY.md). Job code you write is yours to license
+however you want — the scaffolds impose nothing.