Fidelity and its limits

A simulator is only as useful as your knowledge of where it is wrong. This page is that knowledge, written down: what has been measured to be right, what is simplified and by roughly how much, where the simulator departs from reality on purpose, and which problems are still open. Nothing here is hidden elsewhere; if a limit matters to a flight, it should be on this page.

Two things frame all of it. First, there are two fidelities. The classic equatorial range keeps the original planar model, generalised to three dimensions, so that old programs fly as they always did; every real launch site flies the realistic model. Unless a line says otherwise, this page is about the realistic one. Second, the realistic model has been through two physics audits. The first compared the original model with the physics it claims to implement. The second, in September 2026, was four independent reviews in parallel — numerics, environment, vehicle, and control and landing. It found no critical fault and six major ones, all six since fixed, along with most of the minor findings; the rest are listed below with the reason they were left.

What is faithful

These are the parts measured against an analytic answer, a published table or an independent check, with the result.

Several things the second audit confirmed without needing a number: Coriolis and centrifugal effects come out of the inertial frame rather than being added; the pad's rotation is included in the vehicle's velocity from the start; airspeed, Mach and dynamic pressure are all built from the air-relative velocity; thrust uses the day's ambient pressure; and staging conserves momentum.

What is approximated

Right in kind, simplified in detail. Most of these do not change a flight's outcome; a few change a margin, and those say by how much.

Earth and its surroundings

• No geoid. Launch-site elevations are heights above mean sea level, but the simulator treats them as heights above the ellipsoid. The difference is −106 to +85 m depending on the site. Terrain, pads and the atmosphere all use the same reference, so it shifts the reference rather than the physics.
• Flat terrain. Land and sea come from Natural Earth's 1:10 m coastlines on a grid of 1/32° (about 3.5 km), but all land is at sea level except a plateau at the site's elevation, flat within 3 km of the pad and the landing zone and tapering to sea level by 8 km. There are no mountains.
• Earth's orientation among the stars. The inertial frame is tied to J2000 by mean sidereal time alone; the precession and nutation of Earth's axis since 2000 are left out. Every part of the simulator — the launch-window planner, the ephemerides, the flight — uses the same conversion, so it is consistent with itself. Against the real sky in 2026 it places Earth's orientation about a third of a degree off, which shifts a window's clock time by the order of a minute. It matters if you want to match a real launch to the second, not for flying the mission.
• Gravity beyond $J_2$. Earth's higher harmonics, from $J_3$ up, are about 0.2 % of $J_2$ or less, and are left out. The Moon and Mars are point masses plus $J_2$ only — no lunar mass concentrations, no Tharsis bulge. Real low lunar orbits are pulled out of shape by those concentrations within months, some within weeks; the simulator's are steadier than the real ones.
• The Sun, on an Earth flight, is fixed in inertial space for the whole flight: it really moves 0.07° in 100 minutes.

Air and weather

• The mean wind is known exactly. The predictors and the autopilot's feed-forward use the forecast mean wind, and in the simulator that forecast is the mean wind of the physics. Only the turbulence is unknown to them. A real flight also faces forecast error — the wind aloft at launch is not what the balloon measured two hours earlier — and there is no model of it yet.
• One model hour, frozen at launch. The live snapshot does not evolve during a flight, and a weather model's own error is not represented.
• The upper atmosphere keeps the continuum, Mach-dependent drag coefficients up to the 600 km drag ceiling. Above about 120 km real flow is free-molecular, and real thermospheric density rises and falls with solar activity by large factors; neither is modelled. A stage left for hours at 200 km decays at a plausible rate, not a forecast one.
• Gust tracking. The realistic autopilot does not reach the 0.1–0.5° RMS pointing in gusts that real launchers achieve at max-Q; expect a little worse.

The vehicle

• No slosh and no bending. Propellant is a rigid column that drains, and the airframe is rigid. The loads are judged against a bending limit, not computed through a structural model.
• Heat shields are a single lumped thermal mass with a 2 400 K limit. There is no ablation and no recession.
• Aerodynamic envelopes. Grid fins stall per axis, a square boundary rather than a circle; there are no aerodynamic roll moments except the fins'; the crossflow drag coefficient is the low-Reynolds 1.2, because the efficiency factor applied to it was correlated with that value and no flight data separate the two; and the reduction of base drag by the exhaust plume is on the conservative side, because the published data do not cover these cases.
• Tank walls do not thicken with the load they carry, so dry mass does not follow the design automatically. The real-scale presets match published dry masses through a structural scale factor instead.

The flight computer's predictions

• Point masses without lift. fc.impact, fc.predict() and fc.stopPoint() model the vehicle as a point with drag, in the mean wind. They cannot know the gusts, the per-engine thrust spread or the sensor errors they are fed, so a burn should be kept closed-loop.
• About the Moon and Mars the predictors — fc.impact, fc.predict(), fc.stopPoint() and fc.timeToAltitude() — use the body's own gravity, shape, rotation and air (until the third physics audit they used the Earth's, and said "crashed, in 0 s" on the Moon). A predicted lunar impact lands within 0.02 s of the flown one. They leave out the third bodies, since a landing spans minutes, and fc.timeToAltitude() ignores the body's $J_2$ there. fc.passes() is for Earth orbits only.

Missions beyond Earth

• The Mars atmosphere is a mean model — no dust storms, no daily cycle, no variation with latitude. Below 30 km it is a hydrostatic column anchored on Viking Lander 1's mean surface pressure with Seiff's temperatures; against the two measured entry profiles it is 10–37 % denser than Pathfinder's night-time one and 5–18 % thinner than Phoenix's afternoon one. Between 75 and 105 km, where two spacecraft datasets disagree, it is good to a factor of about two. Heights are measured from the IAU ellipsoid, not the areoid — within 0.3 km at low latitudes, 1.5–2 km near the poles. The real atmosphere scatters about these means by a factor of up to 2.7.
• Ephemerides are truncated series, good to about 270 km for the planets and 7 km for the Moon, on time scales taken from the leap-second table. Checked against published equinox, eclipse and lunation times they agree to within the references' own rounding — the Moon to about 15 km. They could not be checked against JPL's DE440.
• Third bodies — the Sun, the Earth, the Moon and Mars — are held for 5 s at a time in the flight model, and carried on the rails of a destination mission too. Jupiter and the other planets are left out, about $10^4$ km over a Mars transfer. Solar radiation pressure is implemented and not switched on.
• The Earth's orientation is turned by the Earth Rotation Angle, so a window opens when it would against the real sky, but it still spins about the pole of J2000, 0.15° from the pole of date.
• The window planner assumes a circular parking orbit, quotes each window at a ±2° plane tolerance, and does not model the parking orbit's $J_2$ drift during the coast. Its lunar search stops at five days of flight, so lunar solutions sit on that edge.

Where it departs on purpose

These are decisions, not oversights.

• Structural limits are generous. Break-up is at 250 kPa·° of bending load and 15 g. Real launchers are rated nearer 100–200 kPa·° and 5–8 g. The limits are kept everywhere so that a first program has room to be wrong; fc.limits reports them.
• The classic presets are not the real vehicles. Their payloads are 75–91 % below those of the launchers they resemble, one is scaled up, and the hopper is short with a thrust-to-weight of 2.9. They are the original simulator's vehicles, kept as they were. Aster 5 and Colossus II are the real-scale presets, modelled on Falcon 9 Block 5 and Super Heavy with Starship.
• The classic range is the original model. Turbulence there is a sum of sine waves in time and height rather than a field; a landing is decided by fixed rules — more than 6 m/s vertical collapses the legs, more than 6 m/s sideways or 35° of tilt tips the vehicle — rather than by the leg physics; landing predictions use the original engine timing, which stops a planned burn 6–11 m high; and powered flight is first order in the step. It agrees with the original simulator to about $3\times 10^{-4}$ relative — not bit for bit, as was once claimed, but well inside its 1 % acceptance test.
• The weather board advises. It never refuses a launch. And several of its rules — lightning, precipitation, icing, cloud, visibility — have no physical counterpart in the simulation at all; they are there because a real range would scrub for them.
• Some audit findings were deliberately not applied: the separation impulses, which conserve momentum and are plausible; the pad clamp; the attitude overshoot of a reaction-control-only vehicle, which is realistic; the gravity-gradient torque, about 10 % of an already small term; and an early max-Q, which belongs to the reference program's trajectory rather than the vehicle.

What is still open

The state of things in September 2026, from the reference programs flown across sites and seeds.

Leaving Earth

• The Mars lander does not land: 0 of 20. Since the third physics audit it reaches the entry corridor with its full 1.96 km/s, and the aeroshell is what falls short: at a ballistic coefficient of 507 kg/m² it is still at 2.6 km/s at 30 km, and once the engines' plume collapses the drag, no ignition point stops it. A 9 m aeroshell came close in a test flight. The flyby, the lunar flyby and the lunar landing now close from all four sites (the tables are in missions and destinations); their failures were bugs, now fixed.
• The lunar landing does not aim for a site. It comes down under its capture orbit, about 5,000 km from any chosen point and not always in daylight.
• Cold gas runs short on long lunar flights. A large slew in vacuum uses far more reaction-control gas than it should; from Kourou the gas can run out mid-cruise, and a later correction then costs 400–500 m/s.

Coming home

• Vandenberg cannot return a booster to the launch site with the reference payload: 0 of 40. A polar launch gets no help from Earth's rotation, and the upper stage ends up short — it now burns its landing reserve to reach orbit, deploys the payload and is expended. The ceiling is about 1.0 t; a lighter payload or a drone ship would do it.
• Aster 5's booster is recovered on 67 of 80 flights at its 13 t payload, 79 of 80 at 11 t.
• Colossus II's upper stage lands with 100–160 t of propellant aboard on legs sized for about 30 t, and crushes or collapses them. It needs a larger landing reserve on that stage, or a way to vent.
• The classic Aster's upper stage lands with 3–7 t on legs built for 2 t, and 8 of 160 landings are graded Hard. The third audit's more accurate orbit insertion saves it another 37–91 kg, so a few more land Hard at bullseye speeds. Fixing it would change a classic design, so it has been left.
• The drone-ship booster lands with its single engine at minimum throttle giving a thrust-to-weight of 0.98–1.05. On about one flight in eight it could hover, which a real booster cannot.
• Sideways touchdowns on the ship: 9 of 80 ship landings are graded Hard for 2.5–4 m/s of lateral speed. The final lateral correction needs its own pass.

What could not be validated

• The ephemerides against JPL's DE440, which was unreachable from the machine they were built on.
• Live Open-Meteo data beyond the recorded snapshots used as fixtures.
• Structural bending and slosh, which are not modelled at all.
• Real centre-of-pressure positions for Falcon 9 and Super Heavy, and a Merlin's nozzle-flow separation at 40 % throttle.
• Landings of the three-engine Colossus and of Sparrow.
• Browsers other than Chromium, and a GPU other than an integrated one.

How to read the numbers

What transfers to reality is the shape of the problems: why a gravity turn works, why max-Q is where the wind hurts, why a landing burn has to be timed rather than flown gently, why a window opens when it does. The margins are another matter. A vehicle that lands with 200 kg to spare here might not in reality, and a load that passes the 250 kPa·° limit here would fail a real review. Treat the simulator as a laboratory whose instruments are good and whose tolerances are stated — not as a flight-qualification tool.