The pipeline, against the infrastructure that runs it

Seven stages. For each: what runs it on a conventional stack and who operates that, what ran it in Muster's September 2026 demo build, and who still sees something. Pick a use case — the pipeline is not one shape. The middle column is that build's dated status, not Muster's scope: all seven stages are Muster's to cover, and today's status, link by link, is in docs/runbooks/explaining-muster.md §4c.

Pipeline · infrastructure · observers
The transaction pipeline against the infrastructure that runs each stage Seven stages — discovery, diligence, negotiation, contracting, ordering, settlement, enforcement — each against what runs it conventionally, what ran it in Muster's September 2026 demo build, and who still observes something. Ordering and enforcement are left unassessed. The middle column is a dated status of that build, not Muster's scope. STAGE, AND WHAT LEAKS CONVENTIONALLY, AND WHO RUNS IT THE DEMO BUILD, SEPT 2026 WHO STILL SEES, AND WHAT 01 Discovery counterparties, intent A platform directory, an order book, a group chat. "Add contact" is a request to a server whose whole business is knowing the edges of that graph — and which can be compelled to produce them. No directory, so no lookup. The address is minted locally and handed over out of band. Nothing is resolved, so nobody is asked. The store node: which topics this client subscribes to, and when. The contact graph, pseudonymously. The bootstrap is unauthenticated. 02 Diligence address history, identity graph Block explorers, and the surveillance firms that graph the chain for a living. The lookup is itself a record — of what you checked, and when, logged against your address at the edge. Nothing. There is no diligence surface in this build. Calling that "closed" because we do no lookups would be the overclaim. Nobody — because the stage does not happen, not because it is protected. An absent feature and a closed leak look identical from outside, and are not the same. 03 Negotiation size, terms, reservation price RPCs, quote infrastructure, hosted chat. For a multisig, a transaction service where the operator — and often anyone with the URL — sees the proposal, the signer set and the timing, all before anything reaches a chain. Inside the encrypted room. No coordination service anywhere in it. Approvals are ordinary messages, authenticated by the room — but not by the chain. That a conversation is active, and when. Not what was proposed, who agreed, or how many. The threshold is enforced by the room, never by the zone. 04 Contracting frontend and signing context A web frontend served fresh on every visit, and blind signing: the signer approves opaque calldata, or a hash they have no way to reconstruct independently. This is the norm, not the failure. A native, local-first client. No web frontend, no ambient leakage. But the signing context was not verifiable — the client did not re-derive what it signed. Nothing new leaves the room here. The gap points inward: you could not check what you were agreeing to. Muster's own client re-derives since (F-4, FS-6); this build never did. 05 Ordering pending order flow A public mempool. Everything pending is visible to anyone who watches, which is what MEV is built on — and the node you submit through learns your network origin at the same moment. Not assessed. Needs the zone's own answer on mempool visibility and sequencer behaviour. Guessing it here would be the failure this figure avoids. Known: no network-origin protection for the transfer. (Since Oct 2026 the mixnet carries room sends only.) The zone learns it happened, and when. The rest is not established. 06 Settlement balances, approvals The public ledger and every indexer reading it, permanently and in the open. This is the stage privacy work concentrates on — shielded amounts, mixers, stealth addresses. Shielded at both ends, by default on the private rail. Four rails are offered; each states on its receipt what it disclosed. The private one takes ~7 minutes. That a transfer happened, and when. That is the floor on every rail, including the private one. On a testnet with two users the anonymity set is as small as it gets. 07 Enforcement identifiable operators Whether a block producer can be named, and therefore leaned on. Anonymous randomised selection is the standard; a permissioned sequencer set is not. Almost nobody argues about this stage at all. Not assessed. The testnet runs against a single named sequencer, which by that standard points one way — but the production model is unconfirmed. Whoever can compel the operator, if the operator is identifiable. Unconfirmed for production, and so left blank rather than filled in with something plausible. MIDDLE COLUMN: THE SEPTEMBER 2026 DEMO BUILD, AS IT RAN. A DATED STATUS, NOT MUSTER'S SCOPE.
inside the room, on your device what an outside party learns a gap, or a refusal to score

All seven, side by side

The column on the right is the one that makes this an architecture diagram rather than a table of opinions. Every stage has somebody positioned to learn something from it, and naming who — per stage, on both stacks — is the only way to compare them honestly. Two cells in the middle column say not assessed, and they will keep saying it until the zone answers.