Dated Siu Inference MCP Server

Touchstone Assay

Touchstone Assay publishes Dated SIU — the benchmark price of AI inference work, measured by actually buying inference against a versioned task basket, not by surveying list prices.

The two hard invariants

1. The print is computed from executed runs only. Published list prices inform exchange rates and are never inputs to the print.
2. The escrow contract is non-custodial. Funds move only to the pre-agreed seller, back to the buyer, or to the fee treasury — no admin path to user money exists in any code path.

How to run a print

Placeholder — wired up once packages/print exists.

Docs

• CLAUDE.md — vocabulary, invariants, build boundaries.
• docs/build1-spec.md — the Build 1 engineering specification.
• docs/plan.md — orientation, dependency order, and open risks.
• docs/architecture.md — one diagram of the whole system.
• docs/demo.md / docs/demo-arc.md — the buyer/seller agent loop, real transcripts, on Base Sepolia and Arc Testnet.

Deployments

Arc Testnet

Chain ID 5042002. Deployed 2026-09-05 from commit 95eea03.

Full record — transaction hashes, block numbers, constructor arguments, compiler settings, and live smoke-test results — is canonical in data/deployments/arc-testnet.json. This README section is a convenience view generated from that file (node scripts/generate-readme-deployments.mjs) and must never be hand-edited to disagree with it.

Not deployed to mainnet. Requirements before it can be:

• treasury MUST be a Safe/multisig, not an EOA. It is immutable with no setter, so a lost key makes fee revenue permanently unrecoverable and the only remedy is redeploying the escrow and migrating every integrator.
• Generate TOUCHSTONE_PUBLISHER_KEY on an air-gapped machine. The testnet key reused here has passed through repository-adjacent files.
• DeployArcTestnet.s.sol is chain-guarded to 5042002; a mainnet deployment needs its own script and its own review.
• Arc mainnet's own chain ID and USDC contract address must be independently re-verified against Circle's/Arc's own documentation at deploy time — never copied from this testnet record, even though Arc's testnet USDC happens to sit at a distinctive, memorable address (0x3600...0000).
• Arc's public mainnet launches 2026-09-16 — see packages/contracts/script/DeployArcMainnet.s.sol, prepared but not run, with its chain-id guard deliberately left as an env-var placeholder rather than a guessed constant.

Base Sepolia

Chain ID 84532. Deployed 2026-08-18 from commit 7b1d6ed.

Full record — transaction hashes, block numbers, constructor arguments, compiler settings, and live smoke-test results — is canonical in data/deployments/base-sepolia.json. This README section is a convenience view generated from that file (node scripts/generate-readme-deployments.mjs) and must never be hand-edited to disagree with it.

Not deployed to mainnet. Requirements before it can be:

• treasury MUST be a Safe/multisig, not an EOA. It is immutable with no setter, so a lost key makes fee revenue permanently unrecoverable and the only remedy is redeploying the escrow and migrating every integrator.
• Generate TOUCHSTONE_PUBLISHER_KEY on an air-gapped machine. The testnet key used here has passed through repository-adjacent files.
• Deploy.s.sol is chain-guarded to 84532; a mainnet deployment needs its own script and its own review.

Verify a print independently

A print's own signature and public_key fields only prove internal consistency — that some key signed this exact body. They cannot prove that key is Touchstone Assay's, because a tampered file could carry a self-consistent signature over a different key entirely. Closing that gap is the entire reason the publisher key is coupled to TouchstoneAttestation: the contract's publisher address is immutable and lives outside the file, so it is a truth a tampered print cannot rewrite.

The loop:

1. Recompute the print's body hash independently (JCS-canonicalise the body minus its signature fields, keccak256 it).
2. Recover the signer's address from the raw {signature, hash} pair — not read from the print's own public_key field. (Recovery yields two address candidates, since the stored signature carries no recovery bit; exactly one matches a real signer.)
3. Compare the recovered address against TouchstoneAttestation.publisher(), read live from chain.
4. Confirm the same body hash is anchored (postedAt(bodyHash) > 0).

BASE_SEPOLIA_RPC_URL=... pnpm --filter @touchstone/print run verify-onchain <print-id> base-sepolia

No TOUCHSTONE_PUBLISHER_KEY is needed — this command only reads. Real output against the worked-example print (docs/siu-worked-example.md, Dated SIU $0.0383) anchored on Base Sepolia:

On-chain publisher():        0x284ff2F8605Ff8AFeDa6959B856Bb7E6d48f845a
Recovered signer candidates: 0x7ab67ceaaf33b21336b0481b3e8a867b70503e35, 0x284ff2f8605ff8afeda6959b856bb7e6d48f845a
  -> MATCH (recovery id 1): this print was signed by the on-chain publisher.

postedAt(bodyHash): 1787035752
  -> ANCHORED at 2026-08-18T06:49:12.000Z

VERIFIED: signature matches the on-chain publisher AND the hash is anchored.

A print signed by any other key — or never anchored at all — reports NOT VERIFIED and exits non-zero (verified live against Base Sepolia with an unrelated key: neither recovered candidate matched publisher(), and postedAt read 0).

Historical event data needs an archive RPC

Public RPC endpoints prune old history. Confirmed live, 2026-09-27: sepolia.base.org now refuses eth_getBlockByNumber/eth_getLogs for anything before block ~46,000,000 — which includes this project's own August 2026 escrow deployment and smoke tests. Two different kinds of read are affected very differently:

• Current on-chain state is unaffected. verifyPrintOnChain above (postedAt(bodyHash), publisher()) reads TouchstoneAttestation's own current storage, not historical logs — it works regardless of how old the anchored print is, and always will, on any node.
• A historical event-log scan needs a real archive endpoint. The console's own event indexer (packages/console) always scans from each contract's fixed deployment block, so it always needs one — set TOUCHSTONE_<CHAIN>_ARCHIVE_RPC (e.g. TOUCHSTONE_BASE_SEPOLIA_ARCHIVE_RPC) to a real archive-capable RPC URL (we use Alchemy; its Base Sepolia endpoints serve full history on every tier — Infura is an equally standard alternative). The public endpoint is kept for current-state reads (feeBps/treasury/current block number) only.
• verify_receipt's own transaction-receipt lookup was checked, not assumed, against this same pruning event: it still succeeds for a real settlement from August 2026 — getTransactionReceipt by hash appears to survive on this provider's public endpoint longer than open-ended block/log range queries do. That is not a guarantee it always will, on this or any provider, so OnChainSettlementReader/verify_receipt also accept the same TOUCHSTONE_<CHAIN>_ARCHIVE_RPC as an optional override for verifying an old settlement, falling back to the public endpoint when unset.

If you are independently verifying an old print or receipt and a historical read comes back empty, this is almost certainly why — the public endpoint's pruning horizon, not a real gap in what was actually recorded. Configure an archive RPC before concluding otherwise.