Accuracy

Accuracy

Every figure on this page came off a flown or benched drop. None of them is an estimate, and each one is quoted with what it is actually worth.

Measured

Radial error at touchdown.

Drop	Error
Unguided C-130J, live wind, 1485 m at ~383 km/h	19.3 m
Unguided C-17, same conditions	21.1 m
Guided cargo (JPADS), zero wind	0.76 m
Guided cargo, 10 m/s wind	1.15 m
Jump run exit cue	7 m and 12 m on two flown runs

What each one is worth

The unguided figures

Flown, one drop each, in live wind. They are what the recorder reported, not an average over a campaign. Expect that order of accuracy on a well-flown pass in the calibrated airframes; a single number from a single drop is not a distribution.

The guided figures

Bulk-validated over many runs, which makes them the best-supported numbers here — with one real caveat:

They were measured with the steering running on a client. On a dedicated server the canopy is usually owned by the server and flown at the server's frame rate. That case is unmeasured. Treat 0.76 m and 1.15 m as very good rather than exact until someone re-measures them on a dedicated server.

Guided accuracy also degrades above about 12 m/s of wind, for a simple reason: the canopy cannot fly upwind faster than its glide speed, so beyond that it is being carried. Raising Canopy Glide Speed pushes the limit out; it does not remove it.

The jump figures

Two flown runs, cue-flown by hand, inside a 200 m box. Two data points.

Calibration state, by airframe

Airframe	State
C-17	Fully calibrated. Zero-wind baseline and wind deltas both measured.
C-130J	Measured zero-wind baseline (48 runs). Borrows the C-17's wind deltas, counter-shifted so its wind predictions match what the borrowed tables gave.
Everything else	Uses the C-17's numbers.

Other aircraft

Any aircraft that is not the C-17 or C-130J flies a drop using the C-17's model. Two things follow, and they are easy to get backwards:

• It solves normally. There is no warning tier and no degraded state — a solution is GOOD or it does not exist. Auto drop is not gated, and nothing amber appears on the HUD. (Older documentation said otherwise; that tier was removed in v0.14.0.)
• The accuracy figures above do not transfer. They were measured on the calibrated airframes. The maths runs; the claim does not.

If a particular airframe misses consistently in one direction, that is worth measuring into its own profile rather than guessing at.

Where the model has no hand-placed release point for an aircraft, it derives one from the bounding box — six metres aft of the rear extent, 1.5 m below the hull. That derivation reproduces the C-17's hand-placed geometry to within 2 cm along and 14 cm vertically, which is good enough to carry and drop from, but it is not a calibration.

Where the accuracy actually goes

In order of how much they cost:

1. Canopy scatter in strong wind. The dominant term for unguided drops in wind, and it is not a calibration error — it is the canopy. Guided cargo is the answer to it.
2. Flying the pass. A load released off the line misses by about as much as you were off, which is why the release gate refuses rather than dropping anyway.
3. Lateral drift at release. Measured: about +1.3 m/s of sideways aircraft velocity produced +30 m, and −2.9 m/s produced −68 m, over roughly 24 seconds of fall. This is what the drift limit exists for.
4. Release delay, at 0.5607 s — already in the model, and a time rather than a distance, so it does not scale with speed.

What is not yet measured

Stated plainly, because a gap quoted as a number is worse than no number:

• Multiplayer is source-verified, not flown. Everything from v0.10.0 onwards — the release split, cargo loading, stick slots, the jump broadcast and the force autopilot — still needs a two-player dedicated-server pass.
• Guided accuracy on a dedicated server, as above.
• Signing has verified against the signature checker but has not yet been through a live verifySignatures = 2 connection.
• Broader C-17 envelope. The calibration covers the conditions it was flown in.

How the numbers were got

See How It Works → How the numbers were got. The short version: a bench that flies many independent drops per descent, feeding a model in which every value records its sample size and method — and a standing rule that the bench cannot validate anything involving the cockpit, so after any calibration change, somebody flies one.