Research program

FIXER Sim — a sim-to-real research program for the Bay Area film industry

Real GMV today: $0.

  • The Sim is a hypothesis generator, not a forecast, a measurement or traction.
  • All people in the simulation are synthetic.
  • No production data is used: nothing here comes from real FIXER members, orders or payments.

We simulate a film-production marketplace to decide how it should work before real orders exist, then test the simulation against the first 50–200 live orders. This page shows what is modelled, what is guessed, and what would prove us wrong.

Three questions the simulation answers

  1. How many active runners per zone per hour fill an order within 30 minutes?

    Open — not answered yet

  2. Under what demand, cancellations and no-shows do runners' hourly earnings stay above a floor?

    Open — not answered yet

  3. Where does the marketplace break — at what cancellation, no-show or supply level do orders stop filling?

    Open — not answered yet

Answers are ranges from many simulated days with different assumptions, not forecasts. Today none is answered yet.

What FIXER Sim is

A simulated shoot day, a way to learn its rules, a way to check them

Twin — what we digitise

A simulated Bay Area shoot day: productions, runners, orders, routes and payouts. 120 product rules come from FIXER's own product spec. FIXER spec (internal) Routes come from OSRM, a routing model, not live traffic. Sourced Money is conserved: a run whose ledger does not net to $0.00 is refused. The 89 behaviour rates (cancellations, no-shows, show-up) are guesses today and are not calibrated. Assumed · guess

Train — how the rules are learned

Not one tidy day but hundreds of imperfect ones per option. We randomise the inputs, not the rules: demand and supply from x0.3 to x3, heavy-tailed travel times with traffic and weather shocks, cancellations 2–25%, no-shows 3–35%. Assumed · board estimate Answers are ranges. If the ranking of options flips between runs, the answer is 'inconclusive'. Where a rule breaks is a finding.

Planned The randomized runs start with stage B, after the OIC link.

Transfer — how we check against reality

Planned protocol

Before launch we write down P10–P90 bands for six metrics: fill rate, median time to match, cancellation, no-show, runner earnings per hour, take rate. We then compare them with the first 50–200 live orders, outside the Sim. Under 50 orders: no verdict. With 50–200, at least 80% of metrics inside their bands and the ranking unchanged: we trust the Sim for comparisons only. Planned Otherwise the Sim is wrong on that point, and we refit and retest on fresh orders. Bay Area is city 1; Los Angeles is city 2, where we test the transfer.

Today none of this is validated: no live order has been compared with the Sim yet.

How to read the labels

Every number carries a label

One vocabulary, used everywhere on this page and in the simulation.

Measured
Taken from live FIXER data. Today there is none yet.
Sourced
Taken from a named public source, such as the Film SF permit counts or OSRM road routes.
FIXER spec (internal)
A rule taken from FIXER's own product spec, such as fees, payout timing and offer windows. Counted separately from sourced.
Assumed · subtype
Not measured and not sourced. The subtype says whose assumption it is: Assumed · guess Assumed · founder estimate Assumed · board estimate
Modelled
An output of the simulation, not an observation.

A Planned chip marks something not yet in the model or not yet run.

Limits

All people, photos and orders in the simulation are synthetic and labeled as such.

No real person, company or address appears here. Nothing on this page is a customer, a testimonial or revenue.

  • Real GMV today: $0

    Measured Live FIXER data: none yet.

  • The behaviour rates are guesses

    Assumed · guess 89 behaviour rates, all uncalibrated. Results are illustrative while the behaviour rates are guesses.

  • Not validated against reality

    Planned protocol No live order has been compared with the Sim yet.

Not modelled

  • Demand and conversion
  • Willingness to pay
  • Time to liquidity
  • Legal classification of runners
  • Deals (Go runs only)
  • Prices and fees: the Sim is not trusted for pricing until live data exists

Use it to compare options and to find breaking points. Do not use it to size the market or to set prices.

The live model

What the map at /sim/ shows

One 2D map of the Bay Area in FIXER black and yellow: 192 synthetic participants, synthetic projects and requests, and one shared model clock in Pacific time that the server owns. You can open any project or request and read who could serve it, who could not and why, and which project step waited or failed.

Modelled It is a bounded two-model-day experiment that started at model time 30 October 2026, 06:00 PT. Model time runs faster than the clock on your wall; when the two days are used up, the last state stays on screen. Earlier moments are kept as read-only history.

Sourced Public studios, stages and rental houses anchor the map as addresses only, with road geometry from OSRM. A place on the map is not a partner, a customer or a stock list. There is no live traffic feed: nothing moves that the model did not compute, so at a quiet moment you will see people resting and nobody travelling. That is the honest state, not a broken page.

Measured None of it. The counts on the map are not bookings, customers or revenue, and the model is not proof of continuous operation, not a validated forecasting engine and not evidence that FIXER improves production outcomes. If the model's server is unreachable, the map says so and shows no movement.

Open the live model Back to fixer-app.com

The live model is being set up.

One case from the model

A lens that nobody in the model could carry

Modelled Historical replay, synthetic people and projects — not the current state and not customer traction.

Model time Saturday 31 October 2026, 10:30 to 11:45 PT, read from world 99b39cea at revision 44 (model time Sunday 1 November 2026, 01:00 PT); the case reads the same at revision 30. A synthetic commercial shoot at a rental house in Cupertino needed a camera and a lens with a declared value of $50,000 brought over from a content studio in San Jose before its prelight could start, inside a three-hour request window. For the lens the model checked all 192 synthetic people and could assign none: 108 hold documents whose tier does not cover a declared value that high, 32 have only a commute vehicle (motorcycle or pickup) that the model does not accept for deliveries, 32 hold a role the city does not dispatch, 12 have expired documents and 8 have not completed identity, so the request closed as "search exhausted" with no one assigned. For the camera one eligible person was found (Person 005, a cargo van, a 21-minute detour, next commitment checked), but the model refused the request by rule because the fictional rental release had not been confirmed. With both inputs declined, the prelight step failed at 11:45 and the shoot, ingest, edit, review, revision and delivery steps were blocked behind it; the project ended as "required service unavailable", with no delay or saving attributed, because the model has no matched alternative to compare against. The case feeds the model's own question, whether an eligible person can meet a project's need inside its window without breaking another commitment, and here the binding constraint was documents and declared value, not travel time. One case is one trajectory under declared assumptions, not a finding; nothing in it happened to a real business or person.

Assumed · guess People, requests, prices, documents and release permissions in this case are declared assumptions of the model, not measured demand. Planned No live FIXER order has been compared with it.

Open the live model and its earlier moments

Contact

Attack an assumption, send a better number, or tell us what you would need to see next. This goes to the founder directly.

The simulator's source is private today. If it is released as open source, the date and license will be posted here.