Corner was built in 76 days, proved with a live dollar, and killed on the evidence.

Corner is a production payments and inventory system for unattended and small-format retail, built by one principal between April 30 and July 14, 2026, and proved on July 9 with a live purchase charged and refunded on real rails. The retail business it was built for was closed five days later, on the category’s own published record: the cost of proving the model exceeded the return at any demand anyone has measured, the region is contested and its distribution is owned by incumbents, and better-funded operators had already run the experiment and lost.

This is the program’s account. The venue’s feasibility question, with the chapters, the charts, and the sources, is Project 01.

What it was

Corner started as software for independent corner stores. One price list and one inventory across every participating shop, a register the keeper rings every sale through, and replenishment of only what sold, so the keeper carries no inventory risk. The thesis was that the leased footprints of small stores are a distributed retail asset nobody can reproduce, and that what they lack is scale economics, not locations.

The format then moved to the harder version of the same idea: a store with nobody in it. A shopper opens a session, takes what they want, and is charged on exit, with the operator as merchant of record. That is the format Project 01 tested, and it is the format the evidence closed.

What got built

While the retail question was being researched, the software it would need was built and tested to destruction. The counts below are the repository’s, not a summary of it.

The four money paths are a code-confirmed shopper checkout, a cash and external-card walk-in register, tap-to-pay through a card processor’s terminal rails, and an unattended charge-on-exit loop for a store with nobody in it. The design assumption throughout was that the network is unreliable and event delivery is optional: every money mutation takes a mandatory idempotency key through a reserve-then-complete envelope, every webhook is deduplicated by event identity, and a reconciliation job converges any stuck payment or refund against the processor’s truth through the same idempotent operations.

That posture was then attacked on purpose: oversell races, idempotency storms, duplicate-webhook floods, cron-versus-webhook finalize races, decline storms, out-of-order refund events. All nineteen scenarios held. The pass surfaced exactly one real defect, an idempotency key that leaked across two transactions. It was fixed and the pass re-run clean. That number is one, not zero, because the pass was built to find defects rather than to pass. Chapter 06 carries the full record.

The dollar

On July 9, 2026, a physical bench unit, a single-board computer driving a real lock behind a small touchscreen, rang up a live $1.99 purchase by phone wallet against production infrastructure. Face ID on the shopper’s phone, a real charge on the practice’s own live merchant account, and a receipt served with no personal data on it. Corner was its own merchant of record end to end. The refund was issued by hand from the processor’s dashboard, which exposed a gap the program recorded rather than closed: nothing ingests a dashboard-initiated refund, so Corner’s own database still reads that sale as succeeded.

The lifetime live ledger of the program is that one charge and its refund. Net zero. It cost two bugs that no test-mode check could have caught by construction.

What the category’s record said

This is the kill, and it is not ours. Before the model was committed to, the category’s public record was read in filings, dated press, and the operators’ own statements. Better-capitalized entrants had already run this experiment. Appendix D carries all eleven of them, with photographs and citations.

Bobacino is the sharpest of these, because a competitor said the thing out loud: what beat it was the cost of developing the model and clearing compliance, not the market’s appetite for the product. That is the same wall this program walked into, from the other side, with less money.

Zenblen is the one that supplies the arithmetic. It has the category’s only published throughput figure: about 2,500 smoothies in calendar 2023 at its field unit, which is roughly 6.8 cups a day. That unit sits inside a members-only innovation hub, captive and habitual, and had been live for a year or more before the counted period began. It is a ceiling, not a floor. And its economics are not the cups: the operator states on the record that revenue comes from a monthly service fee paid by its location partners as well as from retail. The venue pays for the amenity. That line, not the cup margin, is what carries the business.

Margins were never the problem. The zero-labor store clears its running costs at under four cups a day. Paying back the capital is the problem, and only at demand nobody has ever measured. The floor case is not a conservative assumption; it requires beating the category’s best documented steady state by 75 percent, from a cold start, on a street, in a winter city. Run it the generous way instead, at the category’s measured 6.8 cups and granting the category leader’s own US$8 ticket rather than this model’s $6.25, and the $26,000 stack still pays back in about 44 months against a 24-month line.

Then the distribution. Where these machines survive, they are placed, not walked up to. At two of Zenblen’s eleven venues, the kiosk is procured the way a snack vendor is procured, through the venue’s incumbent vending and dining contractor, with that contractor’s own regional director speaking for the expansion. Canteen and Chartwells, both divisions of the same foodservice group, hold the relationships that put a machine in front of people. Toronto already carries operators, RC Coffee across about a dozen sites and Caffeo downtown. A newcomer proposing to sell cups on a sidewalk is entering a contested region through the one door it does not hold the key to.

One wrinkle, recorded because leaving it out would be worse. The research autopsy’s own formal verdict was pivot, not kill: no gate had failed outright, and its closing line was that nothing in it says scrap. The principal killed the program anyway, hours later, on the arithmetic and the category’s record. The idea was vetted thoroughly and did not survive the vetting. That is what this section is: an abstract idea, tested against the only evidence that exists, and found not worth the money it would cost to find out.

What did not kill it

Not the lot. Project 01 found that the rear structure cannot be reached by a walking customer without crossing the lot’s vehicle plane, and that finding is real and stays in the record. But it caps one 240 square foot structure at non-customer-facing uses. It is not the program’s kill, and it never was.

The principal controls the parcel as landlord. The lease of the property’s main improvements reserves that structure and the two parking spaces in front of it to the owner side, with a right of way to them and an unfettered right to alter that reserved area or build in it. Circulation was engineerable from the owner side, subject to agreeing the access route with the tenant as the lease contemplates. It was a question of cost and sequencing, not of a veto.

More decisively: the build never leaned on that location, or on any location. The platform is venue-independent by construction. The store could have been a non-producing store, on this lot or another or none, and the outcome would not have changed. The site was never load-bearing. Any account of this program that reads “the venture died because a lot was the wrong shape” is wrong. The self-criticism that belongs here is about the abstract idea, which was wrong, and not about missing a cheap physical check.

What was never proved

The market. Nothing below is softened.

A system can be correct and unwanted at the same time, and this one has only ever been shown to be the first. Everything above about the platform is a claim about engineering. Nothing above is a claim that anyone wanted to buy from it.

What transfers

The platform. Four money paths, idempotent state machines, reconciliation convergence, receipts, and a 71-check adversarial record. It charged and refunded a live purchase on production rails. It is venue-independent by construction, which is exactly why the venue’s death did not touch it. Its live account, with the bench photographed and the pathway forward, is kept at Platform.

The edge stack. On a $250-class edge computer with 8 GB of memory, a 3-billion-parameter language model and a speech-to-text model run co-resident. Grammar-constrained decoding against an indexed stock list cut proposal latency from 8.8 seconds to 2.9; end-to-end voice lands at about 2.6 seconds in English, under the program’s 4-second bar, about 3.7 in Mandarin, and about 5.9 in Tamil, which is the documented miss. Technically validated, deliberately parked, and never ratified as a direction. Those are separate facts and the record keeps them separate.

The method. Kill criteria written before capital was committed. Primary-source verification of load-bearing claims. Adversarial review of money code, which caught a defect that a green test suite could not. The right to decline a task whose premise is wrong. At close-out the method had produced 56 slice specifications and 77 report files, and the reports are the record. It is stated in three paragraphs in Method.

The entity. Craftpine Inc., an Ontario corporation incorporated May 15, 2026, holds the intellectual property, the payment accounts, and the developer enrollments. Separation from the principal’s other companies held for the program’s whole life, under schedule pressure, every time a shortcut was available.

Venue capital consumed across the whole arc: approximately zero. Nothing physical was bought for the structure, no permit was filed, no machine was ordered. The spend was bench hardware, domain renewals, and time. Four framings died on evidence rather than on sunk cost.

Project 01 → the 520 Barton study
Nine chapters and four appendices on the venue, the category, the machine, and the numbers, with every load-bearing claim traced to a primary source.

craftpine.com →
Alcove is a practice of Craftpine Inc.