Files
chains/mpcvm/README.md
T
zeekay c0d4022c71 Cite the specs by the numbers they carry
LP-134 was renumbered LP-1340 (lp-1340-lux-chain-topology); 21 citations
across 16 files still named the old number, which resolves to nothing.

Corona rests on Module-LWE, not Ring-LWE — LP-4805 says so in three places
and b17e086 already corrected one file, leaving evm/main.go and
schain/SECURITY_review.md claiming the wrong hardness assumption. The
two-family argument in LP-4900 depends on which lattice problem this is.

P3Q at 0x012205 is kind-byte signature dispatch; STARK/FRI is starkfri at
0x012220 (LP-7050). Naming P3Q for a scheme at another slot describes a
precompile that is not the one being imported.

Carries the substance of 2f22fa2, the one commit the parallel line held that
this history did not.
2026-09-01 01:40:34 -07:00

5.4 KiB
Raw Permalink Blame History

mpcvm — M-Chain

M-Chain is the MPC threshold-custody chain. chains/mpcvm is its VM: a plugin-loaded chain.ChainVM with genesis, validators, block production and replicated state.

It exists to hold the keys that custody bridged assets on external chains, and to do that on-chain instead of in an off-chain signer cluster.

  • vmID mpcvm = constants.MPCVMID, CB58 qCURact1n41FcoNBch8iMVBwc9AWie48D118ZNJ5tBdWrvryS. The plugin binary's filename must be that CB58 — that is how the node resolves a CreateChainTx's vmID to an implementation. A vmID is an immutable one-way door once a chain is created with it; TestVMID_IsCanonicalAndStable pins it and lists every declaration that must move together.
  • Genesis chain. M-Chain is in the P-Chain chain set at height 0 on every network (genesis/configs/*/mchain.json), tracked by validators from boot — not created later by a CreateChainTx someone has to remember to submit.
  • Per LP-1340 / LP-7050 there is no T-Chain and no teleportvm: teleport IS bridgevm (B-Chain, LP-6000), and the FHE half of the retired ThresholdVM is fhevm (F-Chain, LP-8200). Any identifier still naming "T-Chain" or "ThresholdVM-as-a-chain" is stale.

The quorum, and why it is spelled "3-of-5"

A threshold policy has two numbers that differ by one, and confusing them is the defining failure mode of this domain:

meaning where it appears
K signers required operator language, runbooks, genesis policy
t polynomial degree = K−1 cmp.Keygen, frost.Keygen, stored configs

Passing an operator's 3 where a library wants t builds a 4-of-5 key. Nothing errors; three custodians simply cannot sign, and for a key that already holds funds the only fix is a resharing ceremony with parties that may no longer exist.

So M-Chain never carries a bare threshold number. Genesis states "policy": "3-of-5", it decodes to a quorum.Policy, and the degree is obtained only by calling Policy.Degree(). luxfi/threshold/pkg/quorum is the one place K and t are allowed to meet.

M-Chain additionally requires 2K > N (types.HasUniqueQuorum). CGGMP21 is unforgeable against up to N−1 corruptions, so 2-of-5 is cryptographically sound; but with two disjoint quorums, two halves of the committee could authorise contradictory releases of the same funds.

The policy must also fit the network's validator set — the committee IS the validator set, so N cannot exceed the validators that exist to hold shares. TestMChainPolicyIsSatisfiableByTheGenesisValidatorSet in the node repo enforces it.

What is on-chain and what is not

replicated in the state root contents
c/ consensus yes yes key registry, ceremony log, blocks, height index
n/ node-private no no this validator's secret key shares

The registry holds only public values — policy, participants, group public key, custody address. A share never enters consensus state. State is written the moment each fact becomes true, so a crash cannot lose the record of who holds the funds, and Initialize resumes the accepted tip rather than recomputing genesis.

Blocks are verified, not trusted

  • A sign operation carries the signature; Verify re-checks it against the group public key already in the registry. A proposer cannot fabricate one without forging ECDSA, and a validator that sat out the ceremony still verifies the result.
  • A keygen operation carries a proof of possession by the new group key over its own registration. That rules out registering a key you do not control. The declared degree is bound separately, by participants cross-checking the record against their own share — one honest participant is enough to reject a mis-declared key.
  • Every block carries its post-state root. Verify recomputes it, so two validators that would diverge cannot both accept.

Leaderless and permissionless

Ceremony ids are derived from the task — H(tag ‖ keyID ‖ digest ‖ sorted signers) — so every validator converges on the same ceremony with no announce round and no coordinator. The signing quorum is likewise a pure function of the task, so all nodes agree on which K sign without an election.

The committee is the chain's own validator set. Joining the signing ring is joining the validator set: no allowlist, no operator registry, no admin-gated key ceremony.

Proof

TestBridgeCustody_ThreeOfFive runs five VMs over the real gossip transport: a real CGGMP21 DKG at degree 2, exactly three of five sign, two decline, the signature verifies under the group key, and a block carrying it is verified and accepted by all five — including the two that never touched the ceremony — with every node ending on the same state root.

Layout

path role
vm.go chain.ChainVM: Initialize/resume, BuildBlock, custody API
custody.go ceremony lifecycle — DKG, signing, quorum selection, staging
state.go persisted state: registry, ceremony log, roots, share store
block.go the verified state transition
wire.go native-ZAP struct-is-wire encodings
transport.go gossip router (production) + in-process mesh (tests)
executor.go drives luxfi/threshold protocol handlers
types/ shared ceremony state machine and the 2K > N custody floor
fhe/ TFHE helpers retained for the F-Chain handoff