Your AI-built selling page should have to prove it works — before it goes live.
-22251f){kind=icon}
Anyone can now generate a website from a prompt. Nobody checks whether a real buyer can actually finish the purchase on it. AI-built pages look done and silently break at checkout — wrong quantity, a dead CTA, a confirmation that never arrives. The site ships; the sale doesn't.
ProofGate is a team of agents that replaces a launch operations function — and refuses to approve its own work.
A merchant sends a voice note and photos. An agent crew builds the selling page and compiles the buyer's intention — "book two seats on mobile and receive a confirmation" — into an executable contract. An independent verifier attempts that exact journey. Any failure becomes a permanent regression test, a failure-specific specialist is spawned to repair the page, and only deterministic policy code promotes it to production — after a real booking/payment event witnesses the path. The public Proof Passport stays revocable: cron replays the contracts and flips it amber/red on regression, then repairs or rolls back.
The builder cannot approve its own work. The verifier cannot change the page. Only deterministic policy code turns a passport green.
A prompt to Claude/GPT/v0/Lovable emits a candidate. ProofGate owns the release:
| A prompt | ProofGate |
|---|---|
| Reads and guesses it looks right | Executes the buyer journey — the interpreter, not the model, says pass/fail |
| One model, one shot | Capability-separated crew: builder ≠ verifier ≠ release authority |
| No durable state | Versioned specs, contracts, evals, incidents, provenance |
| Certifies on vibes | Only deterministic code promotes — no model call turns a passport green |
| Forgets the failure | Every failure becomes a permanent regression test it can never regress on |
The defensible unit isn't generated code — it's the growing set of real buyer failures the business can never ship again.
voice brief ─► Cloudflare CANARY ─► buyer contract compiled
│
▼
independent verifier (no deploy/pay keys)
│ FAIL
┌─────────────────┴─────────────────┐
│ failure → permanent eval │
│ specialist spawned → SiteSpec patch│
│ new immutable version │
└─────────────────┬─────────────────┘
│ EXACT replay ─► PASS
▼
real external event (booking / payment)
│
┌── deterministic release authority ──┐
acknowledged? │ false → BLOCK · amber │
│ true → PROMOTE · green │
└──────────────────┬─────────────────┘
▼
PRODUCTION + green Proof Passport
▲
Hermes cron replays contracts ─► regression ─► amber/red ─► repair/rollback
npm install
npm run demo # the full proof loop, in-process, deterministic — prints the timeline below
npm test # 32 green: 3 legacy + 29 unit (release gate, oracle, verifier separation, repair)
npm start # serve the live merchant page + Proof Passport at http://localhost:4173[03] (c) verifier runs contract "book 2 seats" → FAILED (QUANTITY_UNSUPPORTED, page supports 1)
[04] (d) failure captured as eval + repair specialist spawned (role absent at kickoff)
[05] (e) specialist patches a Zod-validated SiteSpec → new immutable version (qty_max=2)
[06] (f) EXACT same contract re-runs → PASSED
[07] (g) release gate, identical facts:
acknowledged=false → BLOCK · passport amber (EXTERNAL_ACKNOWLEDGMENT_PENDING)
acknowledged=true → PROMOTE· passport GREEN (EXACT_EXTERNAL_ACKNOWLEDGMENT)
[08] (h) GREEN — certified version repaired-qty2-v2 · promoted by deterministic policy, not a model
Monorepo — the agency's organs, each capability-scoped:
| Path | Role |
|---|---|
packages/domain |
SiteSpec (Zod) + immutable, hash-verified site versions |
packages/renderer |
the one constrained, production renderer — agents only touch validated specs |
packages/contract-runner |
compiles buyer intentions into executable contracts |
packages/release-policy |
release authority (deterministic promote gate), external oracle, quantity repair |
packages/hermes-io |
Hermes/Telegram boundary |
apps/verifier-runner |
capability-separated verifier (holds no deploy/payment/promotion keys) |
apps/edge-runtime |
Cloudflare Worker — canary/production sites + Proof Passport |
apps/spike-b-dispatcher |
external-event witness path |
convex/ |
durable state: specs, contracts, runs, events, evals, incidents, passport |
scripts/demo-loop.mjs |
the end-to-end proof loop (npm run demo) |
Capability separation (enforced, not named): the builder can write specs but not promote; the verifier can run journeys but not write; the release authority can promote but not generate. A mentor asking "can the builder approve itself?" is shown the guard, not a promise.
Runtime: TypeScript · Node · Hono · Zod · Vitest · Playwright Power-ups (AI as Agency): Convex (all product state) · Cloudflare (deploys + passport) · Linkup (claim verification) · ElevenLabs (voice intake + confirmations) · Wispr Flow (build) Runs on: Hermes — gateway intake, subagent delegation, cron guardian, persistent memory.
| Parameter | How ProofGate earns it |
|---|---|
| Real output (20×) | Real merchant page + real external event; launches and guardian incidents count as tasks |
| Agent org (5×) | Manager plans per-brief; specialist spawned from a live failure |
| Observability (7×) | Per-run trace + cost; red-vs-green diff; regression alert = passport flip |
| Evals (5×) | Every failure auto-becomes a versioned, release-blocking contract |
| Memory (2×) | Now (task) + history (this merchant's failures) + policy (rules) across handoffs |
| Area | State |
|---|---|
| Live merchant page + Proof Passport (touchable) | 🟢 live |
End-to-end proof loop (npm run demo) |
🟢 deterministic, tested |
| Capability separation + deterministic gate | 🟢 enforced + covered by tests |
| Spike A (real Cloudflare Worker render) · Spike B (booking oracle green on consented merchant) | 🟢 passed |
| Telegram voice intake → live generation · production Cloudflare deploy | 🟡 target |
| Dodo live payments · ElevenLabs · Linkup | 🟡 pending credentials — not claimed until wired |
ProofGate never presents a mock, sandbox event, teammate payment, or prerecorded audio as live proof. Failures are preserved, never deleted to inflate a success rate. Real offers map to real deliverables; merchant and buyer consent is logged. Every scoring claim points to an inspectable artifact — see EVIDENCE.md.
- PROOFGATE_BUILD_BIBLE.md — full spec & execution brief (30 sections)
- docs/SYSTEM_ARCHITECTURE.md — system architecture, pipelines, trust boundaries
- DEMO_RUNBOOK.md — judging demo, beat by beat
- EVIDENCE.md — the claim ledger
- docs/privacy-and-consent.md · docs/provider-evidence.md