writing/agent graphs ·
Dozen is a twelve-node marketplace between brands and 1k–50k TikTok and Instagram creators. The model writes captions. It does not accept a seat, post a disclosure, or release a payout. This is what that constraint actually forced.
Every number below comes from production-imported tests in the private product repository, measured 19 August 2026. The public artifact is the case study. This page does not claim live Stripe volume, Connect transfers, or GMV.
01 / the hole
Micro-creators with 1k–50k followers already make $50–$200 posts. They find that work in DMs, in Discord, or not at all. Brands who want authentic reach either overpay one macro or hire a coordinator to wrangle a dozen micros by hand.
GRIN and Aspire sell the coordinator an $11k–$100k/year cockpit. Collabstr and #paid sell a directory and leave the DMs to you. The hole is not discovery. The hole is brief, draft, FTC, revise, post, pay — the work that happens after a name is on a list.
That is a graph, not a chat. A one-pass agent demo papers over it by marking accept, post, and pay passed so a single handler can look finished. I built that version first. It took about forty-five minutes and it was a product lie. The captions were fine. The money edges were fiction.
02 / why the obvious designs fail
Accept spends a creator’s reputation. Post spends a disclosure. Pay spends someone else’s money. If the model owns those nodes, the demo finishes and the product is a caption generator with a progress bar. A failed verifier has to be a stop, not a suggestion the model can talk past.
Enterprise influencer SaaS sells seats to the coordinator. That is a reasonable business if you are selling software to agencies. It is the wrong shape if the product is supposed to be the coordinator. Dozen takes 18% when a swarm pays (15% on priority or on a $2,500+ budget). No seat license. The platform is paid when money moves.
A public /demo that looks paid is the fastest way to lie. Dozen’s demo is labeled a closed sample against the seeded bench. Views are derived from baseline engagement, not a TikTok API. The live graph leaves accept, post, and pay open. I would rather a visitor see an unfinished node than a fabricated payout.
03 / the graph
Draft is the only node whose closer is the model. Everything else is a script, a policy, or a human. Irreversible nodes — safety, accept, post, pay — cannot be owned by the model. A companion graph linter fails the topology if they are.
| Node | Closer | Done when |
|---|---|---|
| intake | script | Budget ≥ $200, niche + platform present |
| safety· irrev. | policy | Fail-closed banned verticals |
| assemble | script | 6–16 micros, fee math, fake-risk < 70 bps |
| offer | script | Rate band + brand escrow funded |
| accept· irrev. | human | ≥60% accept or explicit bench refill |
| draft | model | Drafts exist only for accepted seats |
| verify | script | FTC disclosure + must-say + avoid |
| revise | script | ≤2 selective retries on this subgraph only |
| review | human | Brand sees only the batch |
| post· irrev. | human | Disclosure still present on every accepted slot |
| measure | script | ER not anomalous vs baseline |
| pay· irrev. | policy | Escrow funded + posted + no open dispute |
The inner loop is draft → verify → revise, max two attempts, and it does not restart intake or assemble. The outer loop is accept, which stays running until 60% of offered seats accept and at least six seats fill — or the brand explicitly refills from the bench. Rejected seats stay in the offered denominator. Six accepts of sixteen is not enough.
04 / escrow
A marketplace that lets creators accept before the brand’s money has landed is asking them to spend reputation on a promise. The first version did that. Offers went out the moment assemble passed. That was the second product lie.
Launch now stops in funding. Offers are written. Accept is idle. The brand opens Stripe Checkout for the full budget. Confirm reads the session back from Stripe: payment_status must be paid, the metadata must match the brand and the campaign, and amount_total must be at least the budget. One cent short stays fail-closed. A missing secret key never creates a session.
Only then does accept open. Pay has a third gate on top of “every accepted seat posted” and “measure is clean”: escrow must already be funded. Anomalous engagement — more than 3.2× the creator’s baseline ER, or under 0.15× — holds that seat and holds the escrow. The campaign does not flip to paid so the graph can look finished.
What pay does write, on a clean run, is a platform ledger row. The detail field says the quiet part: creator bank transfers wait on Stripe Connect. Connect is not onboarded. Queuing a transfer I cannot send is honest. Marking the creator paid in the world is not.
05 / results
The tests import the live matching, loop, verifier, escrow, and Stripe modules. They are not a second implementation. As of 19 August 2026: 43 product tests and 5 graph-linter tests, 0 failures.
| Property | Measured | Target |
|---|---|---|
| Production-imported tests | 43 pass | Import live modules, not stubs |
| Graph linter | 5 pass | Model cannot close an irreversible edge |
| Accept without funded escrow | closed | Fail-closed until Checkout is paid |
| Paid amount below budget | rejected | Underpay cannot open accept |
| Missing Stripe secret key | throws | Checkout and priority stay off |
| Anomalous ER after post | that seat held | Measure can fail without closing pay |
| Clean measure + funded escrow | ledger released | Bank transfers wait on Connect |
| Production GMV | not claimed | No live money, no live posts |
A launch essay is usually a traffic event. This one is a constraint dump. There is no live campaign, no creator who has been paid, and no brand budget that has moved through Stripe in production. The 43 tests prove the graph refuses to pretend otherwise. That is the result.
06 / what is still scaffolding
The product talks to Stripe’s Checkout Sessions API the same way Cite already does: form-urlencoded REST, no SDK, a secret key or a thrown error. If the key is missing, priority stays off and escrow stays unfunded. That is the correct failure. It is not the same thing as a live marketplace moving brand money.
When pay passes, Dozen writes a dozen_payouts row with status queued and the sentence “Connect transfers are not live.” The 18% take is computed. The creator’s rate is recorded. No bank account is credited. Claiming payouts here would be the same class of lie as auto-closing accept.
Post re-checks #ad fail-closed. It does not fetch a TikTok or Instagram URL, and it does not confirm the live caption. The public /demo is labeled a closed sample against the seed bench. Views on that page are derived from each creator’s baseline ER, not a platform API.
The public artifacts are this essay and the case study. Shipping a public repo before Connect and live post confirmation would invite people to treat a graph that is honest about its holes as a finished marketplace.
07 / the actual product
Most agent demos optimize for a complete-looking trace. The trace is the wrong object. The object is the set of edges a model is forbidden to close — accept, post, pay, publish, wire money — and the verifier that owns each one when the model wants to keep going.
Dozen is that constraint applied to a two-sided market I actually want to exist: brands in a city, micros who already post, a coordinator no one wants to hire. The captions were never the hard part. The hard part is refusing to mark the swarm paid until the money is real.