520 Barton Street East sits at the signalized junction of Barton Street East and Dewitt Road in Stoney Creek, Hamilton. The lot is 90 by 160 feet, with frontage on both streets. The building is 5,364 square feet over two storeys, built 1947, steel and masonry, with three service bays and five hoists.¹ The current tenant is Mr. Lube + Tires, active and in good standing.¹

Zoning is C2, Neighbourhood Commercial. Permitted uses include motor vehicle service, restaurant, medical clinic, financial establishment, and veterinary practice.¹ The zoning is part of why the parcel matters: the permitted-use list is wide enough to leave the site’s use open.

Exposure

The junction carries 21,000 or more vehicles per day, as listed by the property’s commercial listing and its management site.² Five HSR bus routes run the corridor: 1, 2, 3, 4, and 10.¹ The QEW is two kilometres away via Fruitland Road; the Red Hill Valley Parkway is three.¹

Third parties price this exposure. The digital billboard across the intersection, at 521 Barton, is marketed at 21,034 daily impressions.³ Its two faces point east and west, down the Barton corridor, at exactly the traffic the counts describe.

Context

The surrounding fabric is residential on three sides: detached houses along Dewitt and the side streets, townhouse blocks to the west, and across Barton to the north an industrial and commercial strip. Two schools sit within walking distance, a secondary school and an elementary school, which matters later in this study when demand is counted by channel rather than assumed.

The property’s working life is automotive service, and its street presence reflects that: twin bays on the Barton face, a third at the south end of the east face, a parking apron wrapping the building’s east and south sides, and three curb cuts, one on Barton and two on Dewitt.

The subject of this study is the small structure at the rear of the lot, 12 by 20 feet. The main building is leased; the rear structure and a slice of the lot are not part of the tenant’s premises, which left them free to study.

The program, Corner, ran from late April to mid July 2026. It began as a neighbourhood-retail software platform, narrowed to the question of what the site itself could carry, and ended the day the last candidate use was falsified. Each concept below is listed with what killed it, because the kill reasons are the study’s actual findings.

What each concept met

CPG micro-retail. The founding concept: shelf goods, transparent platform pricing, phone checkout. Payment processing on a $2.50 convenience basket takes roughly 15 percent of revenue, at any volume, so the concept was parked. The platform shipped and works; the format waits for a higher-ticket application.

Staffed bubble tea. Recommended to the program, rejected the same day, and the rejection became standing doctrine: no human staff in store operations, ever. A format that only works with a shift worker is a different business than the one under study.

Unattended bubble tea, then smoothie. The core candidates. Bubble tea fell first, on a food-safety mechanic: tapioca requires a heated hold, and a heated hold of a time-and-temperature-controlled food is exactly the obligation an unattended format is least able to meet.

The pivot target, a smoothie format built on individually quick-frozen fruit, deletes that burden. It died a day later anyway, on the two findings of chapters 03 and 07: measured category demand below the model’s floor, and total capital that does not pay back inside the kill line. Appendix D is the third leg: better-funded operators had already run the experiment and lost.

Prepackaged premium cold drinks. The low-risk format passed every regulatory and machine test in this study. It was swept away with the site, not falsified. Of everything in the funnel it remains the most revivable, on a different parcel.

Venue-placed units, and selling the machines. Floated on the final day and closed by the category evidence: where unattended drink machines survive, venue service fees carry them, not cup margins. Operating placed units is a real business, but it is a services business with a sales motion, which the studio was not going to bolt on to rescue a site.

Private training venue. The last non-retail reuse. It failed on paper before it failed on geometry: a change-of-use permit applies even without construction, a fitness use classifies as assembly occupancy under the building code, and a 240 square foot unplumbed structure cannot meet the washroom and barrier-free obligations that classification brings without disproportionate spend.¹ Client access would still cross the vehicle plane hourly. The local market for hourly training space is also already served.

Micro office. Considered; closed with the venue line. The program’s record holds no dated entry for it, and the chart shows the gap.

The kill criteria

The fresh-drink format carried seven numbered criteria, set while the machine was still a spreadsheet line. Five are worth reading in full, because they did the killing or would have.

K2 was falsified by research before any order was placed: no compliant machine lands under $20K (chapter 05). K1 was never run; the site died before the September counting window. K7 was never answered; the classification question’s public evidence is chapter 04.

What closed the program was not on this list at all: the category’s own published record and the total-capital arithmetic, both of them free and public the whole time. Walk-up circulation was not on the list either; chapter 08 records it as a bound on one structure, and it moved none of the arithmetic. The method note in Appendix C carries the amendment that follows: read the category’s record, and price what it costs to prove a model, before building the model.

The best-documented operator in the category is Zenblen, a Chicago company running robotic smoothie kiosks. It is presented as the category’s best evidence, not as a target: the operator that has survived, iterated, raised institutional money, and placed units. Its public record is the most favourable data available on how the category performs.

Throughput: the one real number

In a 2024 interview, Zenblen’s founder stated the company “served over 2500 smoothies” from its field unit in 2023.¹ Over a full year that is roughly 6.8 drinks per day.

The unit sits inside a members-only innovation hub, a captive and friendly population of habitual users. The hub’s own blog places the kiosk on site by May 2022, the year before the counted period, so the 6.8 reads as a mature rate rather than a ramp-up undercount.² No later or better figure has been published by anyone in the category, in any climate, in any venue class.

Where the units live

Zenblen lists eleven kiosks: university buildings, hospital towers, a members-only innovation hub, a residential building’s gym, a pilates studio.³ Checking each venue’s own published hours and access rules yields a clean structural finding: every unit sits inside an access-controlled or hours-limited building. None is street-facing; none can be reached by a member of the public walking up off a sidewalk, the use an unattended machine nominally exists for.³

Placement follows procurement, not retail siting. At two universities the kiosks arrived through the incumbent campus vending contractor, a division of Compass Group; the contractor’s regional director, not the startup, speaks for the expansion decision in the campus paper.⁴ A third campus lists the kiosk as a line item in the vending contractor’s procurement bid.⁵ The category leader’s distribution is the vending industry’s existing rails.

The economics the record supports

On stage in 2026, the founder described two revenue lines: “we get revenue both from our location partners in terms of monthly service fee and we also get retail revenue from each smoothie we sold.”⁶ No dollar figure has been disclosed for either line. The venue pays for the amenity. The same pitch claims over 550 percent contribution margin, a six-month payback per kiosk, and 99 percent reliability; all are the operator’s own unaudited statements with no disclosed dollar basis, and are labelled here accordingly.⁶

The retail price point at one campus is US$8 per smoothie, per the student paper.⁴ The capital side is better documented: two SEC Form D filings confirm US$2,117,117 sold in the 2023 round and US$2,480,647 in the 2025–26 round, a floor of US$4.6 million across the two, against eleven placed units at the time of this audit.⁷

A US$6 million Series A was described by the founder as “closing” in the same 2026 pitch; no filing or press confirmation existed at this study’s access dates.⁶ Roughly US$420,000 of confirmed capital per placed unit is the ratio the public record supports.

The patent shadow

The blend mechanism itself is not open ground. F’real Foods holds two live United States patents on in-cup blending with an elevator assembly and removable spindle, expiring 2033 and 2037.⁸ A third patent often casually attributed to f’real, US10682942B2, in fact belongs to Tikiz Franchising and covers mobile-kiosk dispensing hardware; the correction matters because the misattribution circulates.⁸ Zenblen’s own application was allowed but not yet granted at this study’s access date.⁹

Whether any given built machine infringes the live claims is a freedom-to-operate question the public record cannot resolve; what the record does establish is that a new entrant’s cheapest mechanical path runs near claims in force until 2037, held by a patentee that has won an infringement verdict on its earlier blender patents.⁸

The adjacent evidence

The two nearest comparators reinforce the read. teaBOT, the automated tea kiosk company, remains active after a decade; its chief executive simultaneously serves as Zenblen’s CTO per his own public profile, and appears on Zenblen’s first Form D but not its second.¹⁰

Blendid, the robotic smoothie company that raised venture capital for an eight-by-eight-foot robot-arm format, disclosed in its FY2025 annual report substantial doubt about continuing as a going concern if revenue targets are not met, thirteen employees, and a US$308,000 obsolete-inventory write-off. It stopped selling its original kiosk at the end of 2024 for two new products, a smaller, cheaper kiosk and a back-of-house unit, to be sold rather than leased.¹¹ The direction of travel among the category’s established operators is uniform. Machines are getting smaller and cheaper, and operators place them in venues on service fees instead of on streets to earn cup margins.

Ontario, on paper

Ontario’s Food Premises regulation, O. Reg. 493/17, defines a “food service premise” as any food premise where meals or meal portions are prepared for immediate consumption. The definition is agnostic to who, or what, does the preparing.¹ A “food handler,” by contrast, is defined only as a person; the regulation contemplates no machine substitute.¹ Vending machines appear in the text as a species of equipment, not as an alternative premises class.¹

Two sections then pull in opposite directions. Section 17, an equipment-sanitation rule, plainly contemplates lawful unattended operation: a machine that auto-mixes water needs a piped potable supply, and the operator’s name and phone number must be posted “if an employee of the operator is not in full-time attendance.”¹

Section 32 requires that every operator of a food service premise keep at least one trained food handler or supervisor on the premise “during every hour in which the premise is operating.”¹ A fresh drink blended from dairy and frozen fruit is a time-and-temperature-controlled food, which forecloses the regulation’s low-risk exemption on its face.¹

So the classification question is exact: is the kiosk a food service premise, which section 32 staffs; or equipment inside a premise, which section 17 sanitizes? The regulation’s text does not resolve it. The Ministry’s own reference document acknowledges that many vending machines “operate automatically without employee assistance” while affirming that all other duties still apply, and leaves section 32 untouched.²

Toronto, in practice

Toronto’s DineSafe open dataset held 104,790 inspection records at access.³ RC Coffee has operated unattended robotic espresso kiosks, fresh dairy and cashless, since 2020, and under its own Robo Cafe brand since June 2021. It listed twelve or more locations at access, major hospitals among them. The dataset holds no record of it under any searchable name.³ Caffeo, a directly comparable unattended fresh-dairy espresso kiosk that opened in September 2025, is listed, and passed a Food Premises inspection on 2025-09-12.³ The same city treats the same class of operation both ways; Toronto’s enforcement of the category is unsettled.

Chicago, in practice

Chicago’s municipal code anticipates this exact business model. Section 4-8-032 defines an automatic-food-vending-machine operation and routes it through a retail-food-establishment license, with health-department approval of the machine type and inspection of the commissary and servicing facilities.⁴ On paper, an operator’s dispensing locations carry licenses.

Queried directly through the city’s open-data API, across both the food-inspections and business-licenses datasets, at all nine of Zenblen’s Chicago venue addresses and over multiple spelling variants: zero records.⁵ The datasets are not blind at those addresses; the same queries return the co-located food operators, from campus dining and a chain bakery to the vending contractor’s own micro-markets, individually licensed and inspected.⁵

One Zenblen license does exist: a shared-kitchen-user license at a downtown address unconnected to any venue, renewed through January 2027.⁵ After five years of operation there is no dispensing-location record, though the code requires one.

Illinois’ statutory language is also softer than Ontario’s. The state food-handling act requires food service establishments to be under a certified manager’s “operational supervision,” wording that plausibly permits one manager across several unattended sites; Ontario’s section 32 says a person on the premise, every operating hour.⁶ The federal model code’s own working group put the strict reading plainly: an unattended operation selling anything beyond prepackaged, non-hazardous food “would be required to have a person-in-charge at all hours of operation.”⁷

The landed-cost arithmetic

Robot-arm bubble-tea machines list on overseas marketplaces at US$12,800 to $15,900, which converts to roughly $18,000 to $22,500 CAD and appears to clear the ceiling.¹ Landed, it does not.

A worked mid-range example at US$14,350 free-on-board becomes $20,305 CAD before the boat: ocean freight adds $600 to $1,500, import duty adds zero to about $1,700 depending on an unresolved tariff-heading question, tax and brokerage add about $1,300, and none of the researched Chinese units carries a Canadian electrical approval, so an Electrical Safety Authority field evaluation adds an estimated $2,000 to $6,000 more.² The full landed range is roughly $26,500 to $28,400 CAD, a 30 to 40 percent understatement in the sticker, and four to six months from purchase order to legally operable is the realistic timeline.²

The payment filter

Before price, one binary filter: who is the merchant of record. The studio’s standing rule is that the operator owns the payment flow. Most overseas vending hardware is built for sealed domestic payment rails and bakes the reader into the machine; there is no legitimate way to intercept a sealed proprietary payment path without the vendor’s cooperation.³

The compliant integration shape is a documented bus or dispense-trigger interface, the way third-party vending processors integrate.³ Of everything researched, exactly one vendor claims third-party point-of-sale openness in its marketing, unverified; the rest are unconfirmed or lean toward sealed.¹ A component-level build passes the filter by construction: every interface in it is the operator’s own.

The component build

The reference design prices a fresh-drink machine as a set of individually certified commercial food-service parts: merchandiser fridge, chest freezer, food-grade dosing pumps, a hot-hold unit, a cup sealer, stainless prep surfaces, an enclosure, and the actuation electronics an edge computer drives directly. Hardware totals roughly $8,400 to $20,000 CAD, midpoint near $14,000.⁴

Beyond price, the build carries two structural advantages: its parts already hold recognized food-equipment certifications individually, a materially cleaner compliance story than a sealed imported box with none; and its integration surface is plain relays and drivers rather than a vendor protocol to reverse-engineer.⁴

Reliability is the bar

The category’s corporate history says a demonstration machine is the easy part. Bobacino raised about US$3.5 million for a robotic boba bar, showed it at trade shows, and closed without shipping a unit. Its chief executive resigned in November 2023, and by January 2024 it was reported shut down with other companies of its venture studio.⁵ Tea Industries, the first robot bubble-tea shop in British Columbia, opened in March 2024 and closed in 2025.⁵

Cafe X closed its downtown San Francisco robot-coffee locations in January 2020 and retreated to airports. Crunchbase News reported that its founder wanted the company to be “less robot, more old-school coffee.” The pandemic closed the airport bars that year, and Cafe X reopened at San Francisco’s airport in 2021, where it runs three today.⁵ Briggo, which reported US$26 million of stock and note sales across a decade, exited by acquisition into a coffee incumbent.⁵ Machines with no human backstop depend on uptime, restock cadence, and a permit officer’s classification.

Corner’s platform is a production-grade payments and inventory system: 173 commits from April 30 to July 14, 2026; 57 API route handlers; 20 database tables, every one carrying row-level security; 48 schema migrations; 44 automated regression suites.¹ It ships four ways to move money: 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.

Money state machines, built for missing webhooks

The design assumption throughout: 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, so webhook delivery is a latency optimization rather than a correctness dependency.

Refunds on the terminal path are two-phase and confirmed by event, with an atomic claim so two concurrent refund attempts cannot both fire.

That posture was then attacked on purpose. Nineteen adversarial scenarios, 71 checks: oversell races, idempotency storms, duplicate-webhook floods, cron-versus-webhook finalize races, decline storms with inventory restore, out-of-order refund events, reconciliation cap saturation. All nineteen held. The destruction pass surfaced exactly one real defect, an idempotency key that leaked across two transactions; it was fixed and the pass re-run clean.²

The live transaction

The loop was proven with money. 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 on July 9, 2026: charged on the studio’s own merchant account and receipted with no personal data, then refunded by hand from the processor’s dashboard.

The refund exposed a gap the program recorded rather than closed: nothing ingests a dashboard-initiated refund, so the platform’s own ledger still reads the sale as succeeded.³

The edge experiment

A parallel spike asked whether store intelligence could live in the store. 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; voice commands in the common case complete end to end in about 2.6 seconds, under the program’s 4-second bar.⁴

Two software levers did the work: a deterministic fast path that skips the model entirely for unambiguous commands, and grammar-constrained decoding against an indexed stock list, which cut the model’s proposal latency from 8.8 to 2.9 seconds.⁴ The edge device never holds payment authority; the cloud stays the ledger of record.

The per-cup model

The modelled ticket is $6.25 pre-tax, with a $5.49 floor price, which sits at the low-middle of the local drink-price ladder rather than above it.¹ At 30 percent cost of goods, one cup contributes $3.87 after processing fees. Processing takes 8.1 percent of revenue on a single-cup ticket, right at the studio’s ceiling, and falls as baskets grow.²

Fixed costs run $450 a month: power including refrigeration and climate, connectivity, insurance, consumable service. No rent is charged in the model, because the structure was already held; a third-party operator should add theirs. No labour is charged, because the format bans it, and the check cuts the other way: one loaded minimum-wage part-timer, at roughly $2,800 a month, exceeds the entire floor-case revenue line.² Break-even is $450 divided by $3.87: 116 cups a month, 3.8 cups a day.

Volume scenarios

Capital, fully totalled

The machine-only frame flatters the format. A $15K machine at the 12-cup floor pays back in 15.6 months, inside the study’s 18-month comfort line.

But the machine needs a structure around it. The structure’s buildout, priced to the fastener in the program’s bill of materials, runs $17,638 for the walk-in-vestibule configuration and up to $23,720 for the serving-window configuration, permits and contingency included.³ Total capital for the smallest real configuration is about $26,000; with the component build at the top of its range, $37,600.

The program’s own kill line was 24 months. Winter makes it worse: at 60 percent cold-month volume with no hot menu line, the floor case’s annualized net falls by about a quarter.²

Then apply chapter 03

Every number above still assumed the floor case: 12 cups a day, every day, from a walk-up window. The category’s best measured operator, mature, in a captive venue, with a decade of iteration behind the machine class, serves 6.8. The floor case is not conservative; it requires beating the category’s best documented steady state by 75 percent from a cold start. The 30 and 45 cup columns, where the larger stacks become rational, have no measured comparable anywhere in the public record.

The lot’s vehicle circulation is fixed by its geometry. Dewitt carries the entries: two curb cuts, the first just south of the corner, the second further down, with a third cut on Barton.¹

The strip west of the building is a gully with no vehicle path, so everything on wheels wraps the building’s north, east and south sides: the apron in front of the twin Barton-facing bays, the east corridor, and the rear lot. Vehicles queue on the apron in front of the bays at exactly the hours a walk-up store wants customers.

The rear structure stands at the south-west of the lot, on the far side of that vehicle plane from both sidewalks. Trace any pedestrian path from the Barton sidewalk or the Dewitt sidewalk to the rear structure and it crosses the vehicle circulation area; the gully rules out a third approach. The geometry offers no route that does not put a walking customer and a moving or queued vehicle in the same plane.

What the geometry caps

The consequence is a cap, not a demolition: in the lot’s current configuration, rear-structure uses that require the public to walk in are out. Uses that do not, such as storage, service functions, or equipment, are unaffected. A customer-facing use sited in that structure would have to move the vehicle plane, the walking route, or the store.

All three are engineerable, and the standing to attempt them was already in hand: the property is held by an entity related to the studio’s founder, as disclosed at the front of the study, and the rear structure and the lot area in front of it sit outside the leased premises.² Re-routing circulation was a question of cost and sequencing.

It was also not the question that decided anything. What closed the program is in chapters 03 and 07 and Appendix D: measured category demand below the model’s floor, total capital that does not pay back inside the kill line, and better-funded operators who had already run the experiment and lost.

How it closed

Each concept in chapter 02 died against a named criterion or a falsified premise, or closed with the site, and the evidence that closed the remainder was public and free the whole time. The category’s own record had already been written by operators with more money: Appendix D holds the roster.

An entrant with US$26 million of stock and note sales on file, hundreds of times what this program would ever commit, was acquired by a beverage incumbent. A US$3.5 million crowdfunded entrant closed without ever advancing past prototypes and trade-show demos. A robot bubble-tea shop opened in March 2024 and closed in 2025.

The demand ceiling of chapter 03 and the capital arithmetic of chapter 07 then said the same thing in this program’s own numbers, and the program’s founder closed all venue exploration the same day the last reuse was falsified. The circulation finding of chapter 08 landed the same week; it caps one structure, and it moved none of the arithmetic.

Venue capital consumed by the 76-day program: approximately zero. Nothing physical was bought for the structure, no permit was filed, no machine was ordered. The spend was bench hardware and domain renewals.

What transfers

The platform. A production payments and inventory system with four money paths, idempotent state machines, reconciliation convergence, receipts, and a 71-check adversarial test record; it charged a live purchase on production rails. It is venue-independent by construction.

The edge pattern. A validated recipe for on-device store intelligence: a small language model and speech-to-text co-resident on a $250-class computer, sub-4-second voice in the common case, with the cloud as ledger of record. The same architecture serves any deployment where compute must live on site.

The method. Kill criteria before capital, primary-source verification of load-bearing claims, and the one amendment this program earned: read the category’s own record, and price what it costs to prove the model, before building the model. Appendix C states it in full; it is the studio’s standing operating procedure.

What was never proved

The market. No store ran on the platform and no shopper used it in the wild: the pre-launch lockdown never lifted, and the venue closed before the counting protocol ran. The close-out state is launch-ready engineering with no market evidence. The program falsified each venue for testing the platform but never the platform as a business.

What would reopen the question

Three conditions were recorded at close. An inbound placement request from a venue relationship formed during the program; no outbound motion reopens the file.

Or counted demand, measured somewhere real, clearing the floor case with the total-capital arithmetic closing inside the 24-month line: that is the first gate now, and no parcel reopens the file on its shape, because the shape of a parcel was never what failed. Or a need for the unattended-retail stack as a hardware testbed, in which case the store becomes the harness rather than the business. Absent one of these, the park is permanent.