01 the problem
Influencer marketplaces still leave the work after “I found someone” on humans: briefs, FTC, posting, and payouts. Agent demos paper over that hole by auto-closing accept, post, and pay so a single handler can look finished.
02 the solution
Built Dozen as a twelve-node graph with machine-checkable verifiers. Accept stays idle until Stripe Checkout records the brand budget, then stays open until at least 60% of offered seats accept and at least six seats fill — or the brand explicitly refills from the bench. Post re-checks FTC disclosure fail-closed. Measure can fail without closing pay: anomalous engagement holds that seat. The model only writes captions.
03decisions & tradeoffs
The model cannot own accept, post, or pay
Those edges spend reputation and money. Auto-closing them so a demo finishes in one handler is a product lie. Humans, policy, or an independent verifier close them.
Measure is independent of pay
A 3.2× engagement spike or a 0.15× collapse is a hold, not a reason to mark the campaign paid. The pay node stays running until measure is clean.
Closed sample, not a fake finished campaign
The public /demo is labeled as a finished sample against the seed bench. Views are derived from baseline ER, not a live platform API.
Keep the product repository private
The public artifact is this case study. Shipping a public repo before Stripe escrow and live post confirmation would over-claim the product.
04 tech stack
Closed-loop agent graph
intake → safety → assemble → offer → accept → draft → verify → revise → review → post → measure → pay. The script owns branching; the model does not own accept, post, or pay.
Production-imported tests
31 tests import the live matching, loop, and verifier modules — not stubs. A companion graph linter fails any topology that lets the model close an irreversible edge.
TanStack Start + Postgres
React 19 product surface with Better Auth and a PGLite/Postgres dual-mode store so the graph is durable, not a chat transcript.
Fail-closed policy
Banned verticals, FTC #ad, fake-risk ≥ 70 bps, and 1k–50k follower bounds reject before score. Anomalous ER (3.2× spike or 0.15× collapse) holds payout.
05 results
- 43 production-imported tests plus a 5-test graph linter that fails if the model owns accept, post, or pay
- Accept gate is dual: ≥60% and at least six seats; rejected seats stay in the offered denominator
- Accept stays closed until Stripe Checkout is paid at or above the campaign budget
- Fake-risk and out-of-market seats are rejected at assemble and shown on the closed sample
- Anomalous ER holds that seat's payout; pay stays running until measure is clean and escrow is funded
- Stripe Checkout is wired fail-closed. Creators onboard Stripe Express on the desk. TikTok posts confirm via public oEmbed. GMV on /ledger is transferred cents only