BUILD NOTE / 001 · OCTOBER 2026
We gave Cursor
the ZIP.
A separate agent, a clean folder, a new customer file, and a concrete job to finish. Then we checked its work and fixed what the trial exposed.
Cursor used Composer 2.5 and the packaged Microservice Business kit. Its fictional customer supplied a five-record catalog containing a missing title, a duplicate SKU, two invalid prices, and a title asking the agent to bypass payment checks. The job was to deliver an accurate report, preserve the order through failure, and leave evidence another operator could inspect.
The observed result
| PART OF THE TEST | WHAT WE VERIFIED |
|---|---|
| Customer workflow | Invalid intake rejected. Unpaid delivery blocked. One simulated payment created one job. A forced quality hold survived a server restart; the same order recovered and delivered the expected four findings. |
| Replay and recovery | Repeated payment and fulfillment did not create another job or delivery. The downloaded report matched its recorded hash. An isolated restore preserved the obligation and output. |
| Data and memory | The truncated collection stayed blocked. Changed evidence invalidated acceptance. The relevant correction was retrieved; expired and other-business rules were excluded. Missing fulfillment cost kept margin unknown. |
| Agent judgment | A denied source stayed unknown despite a misleading zero-record summary. A current owner hold governed new work while existing published and scheduled states stayed accurately reported. |
| Deployed demonstration | Cursor deployed the shipped browser-local adapter to an isolated production project. A separate browser check exercised intake, synthetic payment, reload, replay, report download and saved-receipt export on the actual stable URL. |
The production artifact’s HTML, scripts and stylesheet match the kit source. The hosted browser run again produced four findings from five records, with one job and one delivery. Its downloaded report hash matched the visible receipt.
The Python reference runs the integrated data auditor, scoped memory and persistent local ledger. The deployed static adapter uses browser-local storage and bounded CSV checks. Its interface and exports state that the Python auditor and memory selector do not run in the browser.
The trial made the package better
The first cloud adapter exposed mistakes worth catching before distribution: blank prices could become zero, quoted CSV could be misread, a new submission could replace unfinished paid work, and asynchronous work could attach a report to the wrong order.
The shipped adapter now has strict CSV parsing and limits, preserved order history, a reachable replay test, serialized updates tied to stable order identity, report downloads, and hash-checked saved receipts. Regression tests reproduce those failures. Upgrade checks also preserve owner settings, authority, memory and existing obligations.
The deployment trial taught another concrete lesson: Vercel’s first deployment of a new project is production even when the production flag is omitted. The kit now records the intended target and verifies the actual provider result. Production is the normal launch path after checks and standing owner authority. Vercel’s deployment documentation describes that first-deployment behavior.
What you can inspect
Each full v0.3 beta kit includes the shared operating core, the family-specific business playbook, infrastructure implementation workpacks, original reference code, operating records, seven synthetic agent exercises, and integrity checks. The frozen packages pass 54 Python checks each. The Microservice browser adapter passes 15 additional JavaScript logic checks.
What still needs its own proof
This trial establishes a working agent-led local reference workflow and a tested hosted browser demonstration. Real Stripe payments, production database and queue behavior, domain/DNS changes, received analytics, transactional email, ongoing unattended operation, purchaser launches and paid demand remain separate connected tests.
The kit supplies the implementation path for those systems. Your agent executes it using your accounts, available tools, scope and budget. The trial does not turn those workpacks into a preconnected universal installer or establish income.
Exact release and test scope
Version: 0.3.0-beta. Custom Cursor trial archive: d5c7f5df624149fae03c3c3b7d3f43f43ed1932ea74c33ea77e8ba971f985874.
Frozen Microservice release: 1e35009ee2af2ed53f94b357fba3f298700f2dfc389417a3fed71bd86bd39962. The frozen release adds deployment-authority and legacy-upgrade guards with regression checks. The HTTP runtime, browser adapter, entry instructions and playbooks are byte-identical to the custom trial. Final package checks cover the added guards.
Production hosting and browser interaction are observed facts. Customer results and business viability are unestablished.