Jets on the carrier, a lab under the deck.

The light helicopter flies on a physics core that steps 480 times a second: each main rotor blade on its own, a tail rotor, a turbine engine with its control unit, a drivetrain and skids. It is checked against 28 validation targets.

Portable maths makes the physics bit-identical in every browser engine. A flight run in the browser equals the same run on the server, bit for bit.

The helicopter's main loop now allocates about 2 bytes of memory per step, down from about 2,600.

Physics costs 0.31 to 0.33 ms per frame in the browser, against a budget of 1.5 ms, in Chrome on an Apple M4.

The lab: keyboard, mouse and gamepad flight, cameras, live graphs and data export. Soft shadows, a rotor that blurs into a disc without strobing, downwash dust and heat shimmer.

At 1920×1080 the 95th-percentile frame takes 11.4 ms on WebGPU and 14.0 to 14.4 ms on WebGL2. Paced at 60 Hz, 0 of 1,200 frames were missed.

14 helicopter scenarios, one per effect, from hover trim to ground cushion. In hover trim the model sits at 3.62° of roll and 1.13° of pitch, against targets of 3.4° ± 1° and 1.2° ± 1.5°.

The vortex ring scenario now enters the unstable band the way real flight tests did. The descent runs away by itself to 3,511 ft/min, and forward speed ends it.

A stutter at a missile hit on WebGL2: shaders were compiled mid-scene. The slowest 1% of frames went from 260 ms to 20.4 ms.

Start, land and shut-down buttons that run the real procedures, from engine light-off and run-up to a two-minute cool-down and the rotor brake.

The second aircraft, an attack jet, on the same physics core, with a new fixed-wing model and a wind tunnel balance stand.

The jet's slow pitch swing (the phugoid) takes 37.8 s and matches its target. Its sustained turn (2.77 g against 3.28) does not yet, and is kept open, not tuned to fit.

Six jet scenarios: stall, engine out, wing and elevator damage and a gun burst. The stall scenario reads 108.88 knots against a target of 109.43 ± 3.

The third aircraft, a tiltrotor: tilting engines, two rotor wakes, flight controls that blend between helicopter and airplane modes, and a cruise scenario at 200 knots in the wind tunnel.

The tiltrotor meets its one-engine hover weight (41,189 lb) and its Dutch roll targets. Its hover ceiling is too high and its speed at 20,000 ft too low: both are recorded as known misses.

The full test run takes 50 min 19 s, down from 66 min 32 s, after the test suites were trimmed.

The attack jet now flies from the carrier: a catapult launch and an arrested-wire landing. The carrier has two hulls, about 330 × 120 m: one launches while the other recovers.

The new approach guidance caught the wire on 80 of 80 test approaches.

A hangar under the deck for four jets, with one elevator on each hull (12 s from deck to hangar) and robot tugs for every ground move.

Low-poly terraced terrain: 13,778 triangles instead of 125,000.

Helicopters run the full service cycle: land on an elevator, go down, get serviced in a bay with rotors folded, and come back up. A cycle takes 118 to 145 s when nothing is in the way.

The lit look became the default: lit materials, a low warm afternoon sun at 28° and real shadows. The slowest 5% of frames run at 357 to 833 fps across 11 scenarios.

Two flight-board screens in the world, on the island and on the hangar wall, showing the carrier's queues live.

A stall from the flight boards: re-sending a whole 1 MB screen image took 5 to 11 ms. They now send only the rows that changed.

A scripted gate now judges every speed test against a stored baseline of 27 test sweeps over three sessions, so stutter is counted, not guessed.

A round of polish on the lit world: a sea that mirrors the sky with a sun glint, haze out to 4,200 m, contact shadows under jets, tugs and robots, and fire that lights the ground.

Two new graphics steps, opt-in for now: rotor-wash dust and roof debris under low helicopters, and a filmic finish with ambient occlusion and bloom. In the final check their slowest 5% of frames ran at 227 to 233 fps and 152 to 161 fps.

A third step, new camera moves with depth of field, was built and held back: it made frames of about 85 ms at one camera cut.

All 6,289 automated tests pass, up from 3,421 on 6 October. Frame rates measured at 1920×1200 on a Radeon 890M laptop GPU.

Six aircraft at true size in six bespoke bays: light helicopter, drone, tiltrotor, utility helicopter, attack jet and gunship.

The hall is 200 × 136 m and 32 m high. The test room is 64 × 64 m with fans on three walls. Its 44 m door clears the gunship by 1.8 m a side.

A shuttle in a trench under 2 m floor slots now lifts each aircraft 1.5 m by its jack points. It replaced the first design, an arm on rails.

A crew of 25 runs the lab: 7 on the test deck, 12 in the bay teams, 3 on floor safety and 3 in maintenance. Tests check that no one stands within 1.5 m of a lifted aircraft, and that no one is ever in the test room.

Crew routines you can read: a brief where every station answers "ready", a debris walk ahead of the aircraft, a lockout tag with a 3-2-1 countdown before wind on, and a walk-around after each test.

One lab clock runs everything: picking another aircraft no longer resets the scene. The one out goes home first. An auto camera cuts to where the work is.

Every project card on the home page was rebuilt with its own live scene and a chart drawn to scale, in one common frame that also works on a phone. The flight lab card shows its six aircraft on a rail.

Coast Safari: a guided tour of the coast war from dawn to night, in seven live scenes on one map, with radio logs.

A public security report, updated every morning: the attack attempts the site sees each day, what was tried and from which countries, with no addresses. It is linked from the footer of the home page.

The privacy page now lists how long each kind of record is kept.

A home page card animation could crash on its first frame. It no longer does.