Where these numbers come from

Cybercab Fleet Operator answers one question: if you bought a self-driving taxi and ran it as a business in your town for ninety days, what would happen to your bank account? It is a planning tool, not a prediction, and this page is the whole answer to "says who?"

Every figure below was read out of the numbers the game actually runs on, or out of the live research behind each city. None of it was typed in by hand, which is the only way a page like this stays true after the thing it describes changes.

Built 2026-09-20 from the register in this copy of the game and from the live city research. The page re-checks both when you open it, and says so below if either has moved on.

The honest split

The model runs on 184 numbers. Here is exactly how many of them anybody else can vouch for.

Three quarters of it is ours. That is not a confession, it is the shape of the subject. Nobody has ever run a privately owned robotaxi, so there is no trade body publishing what one earns, what it costs to insure or how often it gets stuck. The alternative to an estimate here is not a fact — it is silence.

What we can promise instead is that every estimate is labeled as one, carries its reasoning, and can be replaced by yours. 144 of the 184 are editable inside the game, and changing one changes the model immediately.

A season in a metro runs on 114 of these; a season in a small town runs on all 184, because a town’s own geography — its radius, how far a fare goes, how fast the traffic moves — is on the list too. The register audits itself every time it loads: a published figure that carries no dated document, arithmetic with no working shown, or an estimate with no reasoning is a failure. It currently reports 0 failures.

What the 184 numbers are about

The same figures as the bar above, grouped by what they describe.

The car itself is the best-documented thing in the subject and none of it is ours: battery, efficiency, motor, weight and seats come off one EPA certificate. The permit rules come off one Nevada Transportation Authority order. Those are also the two groups you cannot edit, because a certificate is not local knowledge.

How a town gets researched

29 towns and cities have been researched so far, and the same machine does every one of them. It runs in three passes, and the order matters.

Anything the United States government already publishes for every city in the country is taken straight out of the national file. Nobody is asked, nothing is searched for, and there is no page that could be misread.

• Census Bureau — Gazetteer place files, subcounty and county population estimates, and the American Community Survey summary files. 32 files.
• FAA — commercial-service enplanements, and the national airport list. 2 files.
• EIA — the Electric Power Monthly state price tables and Form 861. 3 files.
• NOAA — 1991-2020 monthly climate normals, one file per weather station. 107 files.
• BLS — the May 2024 Occupational Employment and Wage Statistics national and area files for vehicle cleaners, and the area definitions that decide which OES area a town falls in. 3 files.

The town is matched to its row by a code, never by its name — a Census place or county FIPS number, an FAA airport identifier, or a coordinate. Names are how you end up with the wrong Springfield. Across the 29 towns, 1,212 of the 2,287 researched numbers came out of these files, from 147 distinct downloads shared between every town in a run.

What is still researched is the LOCAL STORY, and it is researched because it is real, published and findable: the utility's time-of-use split (a filed tariff, not a table), public fast-charge prices, the visitor economy, the state's minimum liability statute, the city's own parking schedule, and the named demand generators -- hospitals with their bed counts, colleges with their enrolments, the venue that fills the town. That is the half of a pack no federal file will ever carry, and it is the half worth having: Dover, Tennessee is a town of 1,918 people in the federal files and nothing more, and what makes it interesting is that 84% of its rides are visitors to Fort Donelson National Battlefield.

Every record the model returned as 'sourced' had its cited URL fetched. The page had to load, and the figure had to be in its text within 400 characters of one of the record's own keywords. Anything that failed was demoted to 'estimated' with the failure written into its math.

Across the 29 towns 422 citations were checked. 301 passed. 121 failed and were demoted to labeled estimates, which is 29% of everything the model claimed to have a source for. A figure that loses its citation does not quietly stay a fact.

Where a demoted figure was feeding another one, the second is demoted too and says why. That happened 14 times. 74 holes are written down across the six packs — things the research could not establish, recorded rather than papered over.

Every research claim about charging stations, population or airports is checked against a federal bulk file the pipeline already holds on disk: chargers against the DOE/AFDC station snapshot (open, public sites within 10 straight-line miles of the Census town center), population against the Census place and county estimates, airports against the FAA enplanement file. Confirmed means the federal file agrees within tolerance (a stall count within 25% or one stall; a figure within 5%). Disputed means the federal file holds a different figure -- both are printed and the pack keeps its own citation. Unconfirmed means the federal file has no record of the thing claimed, which is never treated as false: the locator is voluntary submission and grant-funded sites lag it. Hospital claims are not checked because no federal hospital file is cached. A planned or permitted site never counts as present.

Across 29 of the 29 towns the rule found 308 claims on 233 records and notes: 191 confirmed, 6 disputed (both figures printed, the citation kept), 5 unconfirmed (the federal file has no record; never called false) and 106 hospital claims skipped because no federal hospital file is cached. The worked example is Quincy, IL: its research note counts seven fast-charging stalls at John Wood Community College; the DOE locator lists no open public fast charger within 10 miles, and a cited 2024 state grant funded two 2-stall sites there that the locator has not caught up with. Unconfirmed, not false — and the charger sim prints exactly that.

Re-running the same research used to come back with materially different answers, and those answers drive the whole demand estimate. 5 of them were taken off the model altogether and are now computed from a pinned national survey and measured Census ratios. What judgment remains is asked 5 times and reported as the middle answer with its spread. Across all 87 of those figures the typical gap between the highest and lowest answer is 1.30 times, and the worst is 3.40 times. That is the honest width of the part of this that is judgment.

One pack, three lenses: the Town Sims

Researching a town once is the expensive step. Every simulator after the first reads the same pack and asks a different question of it, so a second sim costs nothing to research and a third costs nothing more. Each is a simulator, not a game: a budget for one site, worked out from the pack’s own records plus a short register of stated assumptions, every one of them tagged with why and editable in the reader’s browser.

⚡ The fast charger (/fleet/charger/, model 0.1.0, built 2026-09-18): would one two-stall DC fast-charging site pay for itself here? It adds three federal files the packs do not carry — EV registrations by state, the DOE station locator, the Census town center. Under its defaults seven of 29 towns pay back within seven years, two are marginal and 20 do not pay.

🧺 The coin laundromat (/fleet/laundromat/, model 0.1.0, built 2026-09-18): would one attended coin laundromat — about twenty washers, someone on the floor twelve hours a day — pay for itself here? Its demand is the pack’s own no-car figure turned back into households: the occupied, renter and no-vehicle household counts are read out of the cached Census table (ACS B25044) the pack’s no_vehicle_share was computed from, matched on the same GEO_ID, and the build refuses to run if the share it recomputes disagrees with the pack. Rent is the pack’s commercial-bay figure scaled to the floor area, the attendant is paid the pack’s BLS local wage for vehicle cleaners (a proxy, and said so), electricity is the pack’s commercial rate. What the packs do not hold — a laundromat count, a wash price, a water rate — is a typed assumption in the register with its reason, and the limits are printed. Under its defaults one of 29 towns pays back within seven years, nine are marginal and 19 do not; the store is full in 19 towns, where the verdict is the wage and the rent rather than demand, and run unattended 19 pay back. The counts here are read from each sim’s results.json when this page is built, not typed.

The pattern for a fourth: a coverage census first. A sim whose inputs the pack does not already cover completely does not get built — the airport-run van, for instance, would need a traffic count no pack carries.

The cross-check: where we can be caught out, we let ourselves be

The demand estimate is the number everything else hangs off. Where somebody publishes a real trip count for a town, we build ours without looking at theirs and then put the two side by side. Where nobody publishes one, we say so and name the five places we looked.

Of the 29 towns researched: 1 had a published figure we landed close to, 2 had one we did not, 25 have no published figure at all, and 1 is a gap in our own work.

What we know is weak

A methodology page that only lists strengths is marketing. These are the four places where, if this tool is wrong for your town, it is most likely wrong here.

Every number is editable. Change these first.

Open Every number behind this game inside the game, tap any figure, and type your own. It goes into the model straight away and stays in your browser. Nothing is sent anywhere — and if you then put a season on the board, it goes up marked as run on custom numbers, because a season played on different figures is not comparable with one that was not.

These 5 are the ones whose own reasoning admits that a figure you know beats ours:

That list is not hand-picked. It is every editable record whose reasoning contains "most worth replacing", "better than ours", "change it here" or "if you know" — matched against the register when this page was built.

What one town costs to research

Between $0.00 and $5.30, measured on the run that actually produced each pack. All 29 together came to $57.30.

262 model calls and 1,270 web searches across the 29, and about 746 minutes of wall clock across the 28 that recorded it. The federal files are free and are downloaded once and shared, so most of that money buys the things no national file carries: the local electricity tariff, what a detailer charges, what a bay rents for, and who fills the town on a Saturday.

measured when the research actually ran on 2026-09-09; rebuilding the pack from the saved answers costs nothing

What this page is not claiming

The game is not a forecast of Las Vegas, Quincy or anywhere else. The fares, the network’s cut, the settlement lag, the loan terms, the insurance, the cleaning prices, the depot rents, the wear per mile, the penalties, how often things go wrong, and every curve describing how a district answers heat or rain — those are ours. They are tuned to make the decisions interesting. Do not quote them at anybody.

Tesla has published no Robotaxi fare table, no owner-operator revenue share and no Cybercab price. The only price on the record is Musk’s October 2024 "below $30,000". The widely repeated "$25,000" is not sourced to Tesla and this game does not use it. Nobody is running a privately owned Cybercab anywhere today; the premise is fiction built on a real statement.

Inside the game, 23 constants are tagged sourced and name their document, 4 are arithmetic on those, 3 are stated as approximate geography, and the rest are tagged ours in the source code itself. The simulation is version 1.5.0.