Architecture.
The topology, revised — the stack holds, the system grew.
Revised 25 October 2026. The six layers have not changed since they were drawn. What runs underneath them has: the gate is deployed, the chain is sealed with a signature — and the door in front of the model now runs as the gate behind it does: intention, verified and signed at intake, before entry — shipped 25 October 2026, its attestations and revocations sealed in the chain beside the verdicts. Every claim on this page is scoped to what the system can actually prove.
The stack
Six layers, in order of authority
The sovereign writes the rules; the machine proposes inside them; the gate judges what comes out. Nothing below a layer can overrule it. There is a gate in front of the model — intention, verified and signed at the door — and a gate afterward: the output, judged before it lands.
Sovereign
The human authority. Author of governance, holder of the override, owner of the parameters. Wrong parameters are the sovereign’s fault — the guarantee is scoped to them.
Governance
The “should we” layer. Policy, consequence classes, and the terms under which any proposal may be heard at all.
Pacing
The doorman — and the door is a checkpoint, not a courtesy. Pacing hosts the front gate: forms the proposal, throttles generation, and renders the intention verdict here, the attestation signed and sealed into the chain before any proposal travels on. Unverified never gets in; the default is closed.
Proposal
A structured, inspectable act — actor, action, target, consequence class, amount — never an opaque command passed straight through.
Transformer
The imagination. It generates; it does not decide. Its output is a proposal, and a proposal is not an event.
Gate
Post-compute verification. Deterministic rules judge the output in parallel; a veto is eradication, not a warning. Check the input and the output — is it good?
The live system
What is actually deployed
One site, one gate, one ledger — and no model calls anywhere in the gate’s path. This is the map as of this revision.
The public site: the papers, the sector pages, this map. Served as static files; it holds no keys and makes no decisions.
LiveThe public sandbox. Paste what an agent handed you, or point the checker at a real website. Runs in the visitor’s browser; unlisted by design while it is a sandbox.
LiveThe gate. POST /api/v1/verdict judges a proposal and seals the record; POST /api/v1/attest seals a caller’s standing intent at the door, before any verdict; GET /api/v1/chain reads the ledger back, metadata only; GET /api/v1/chain/pubkey serves the signing key openly.
Deterministic, and few enough to audit: destructive acts are vetoed; money is judged against consequence-priced ceilings; velocity is capped; named targets and delete-language escalate. Same proposal, same verdict — every time.
LiveEvery verdict is written to a hash-chained ledger: chain_hash = SHA-256(prev_hash | seq | ts | proposal_hash | verdict). A record seals its input and its verdict to the record before it, so history is replayed and verified — not reconstructed.
Each record is also signed, Ed25519, over the same payload the hash seals. Live since 23 October 2026. The private key exists only as a secret held by the gate’s operator; the public half is published by the API itself.
LiveGET /api/v1/fetch-check lets the workspace read a page the visitor’s browser cannot — SSRF-guarded, redirect hops re-checked, size- and time-capped, and unauthenticated by design because it is a public sandbox tool.
Intention verification and behavior watch at the door Pacing keeps (Tools 03–04). A caller seals a signed attestation of standing intent before any verdict; each proposal is checked against that signature (D3), the sealed trajectory is tallied for drift and probing (D4), and a revoked attestation refuses everything downstream (D5). Shipped 25 October 2026 — the attestation, an intention escalation, and a revocation are records 3–6 of the production chain.
LiveProof, scoped
What the signature proves — and what it doesn’t
Provenance, not quality. The signature makes a record attributable; the gate’s rules are what make a verdict worth attributing. Both halves, stated plainly.
It proves
- Origin — the gate itself sealed this record, verifiable by anyone holding the public key, with no cooperation from Verity.
- Integrity — the record has not changed since it was sealed; recompute the chain and check the signature, and any edit shows.
- Order — each record is bound to its predecessor. History cannot be silently re-sequenced.
It does not prove
- When — a record’s timestamp is the gate’s own assertion. External time-anchoring is planned, not claimed.
- Rightness — a signature certifies custody, not correctness. The verdict stands or falls on the published rules.
- Contents — verification needs the hash and the signature, never the proposal itself. Proposals stay inside the ledger.
The data rule & what’s next
Mined. Never published.
The chain stores every proposal it judges. That record is mined — failure patterns, rule refinement, evidence — and it is not published. Caller data does not leave. The dashboard reads metadata only: the chain proves the record; it does not publish it.
RFC 3161 anchoring of the chain head on a schedule, so the clock belongs to a third party. Designed; rides with the first production engagement.
PlannedVerdicts against retrieved sources — supported, contradicted, not checked — beyond the consistency checks the sandbox runs today.
PlannedSealing other operators’ records under the Verity key, under agreement. The root key is never sold; issuance is the product.
Planned