MEASURED, NOT GUESSED ✦THE CENTRE STAYED EMPTY FOR MOST OF A MINUTE ✦0.0 MEDIAN HEIGHT ERROR ✦AN ENSURE, NOT A CRASH ✦2° OFF HEADING ✦7 / 7 ✦MEASURED, NOT GUESSED ✦THE CENTRE STAYED EMPTY FOR MOST OF A MINUTE ✦0.0 MEDIAN HEIGHT ERROR ✦AN ENSURE, NOT A CRASH ✦2° OFF HEADING ✦7 / 7 ✦
To film the war we first had to find it. A probe ran the simulation for ninety seconds and wrote down every soldier once a second: position, faction, whether they were swinging. The plan was to fly the camera down onto the mesa in the middle, where a war ought to be.
The numbers said otherwise. Adjacent factions enter through gates that sit about a hundred and thirty units apart at each corner of the battlefield, so the first enemy anyone meets is around the corner from their own door. Within three seconds of the gates opening, a hundred and forty soldiers were fighting, nearly all of them in the four corners. The centre stayed empty for most of a minute.
The camera path was rebuilt from that data: the landing target is whichever corner has the most sustained fighting in the window when the camera arrives, weighted toward corners with elites in them. That is the southwest corner in the clip on the Lab page. Nothing in the flythrough was staged; the war just got measured before it got filmed.
UNDIVIDED · GREYBOX CAPTUREthe centre stayed empty for most of a minute05
Five render passes, four of them inside somebody's shoulder.
A corner melee here is about four hundred units across and packed right up against the wall. Every eye-level camera we tried ended up inside the crowd: the first frame of the landing was a blue chest, the second a green elbow. Moving the camera back did not help, because the crowd moves too.
What works is altitude and distance together. The landed shot hovers over the wall corner, roughly six hundred units up and looking down at fifty degrees, where no body can reach the lens. The close-up then comes from the lens rather than the camera: the focal length pushes from 30 to 45 millimetres over the hold. Elites at 1.6 times troop scale read as elites at that distance; at eye level they were just bigger blue.
One more thing for the notebook. A cine camera set to manual exposure in this scene renders black, because the sun is only a few lux in this greybox. Auto exposure with fast adaptation, and the war comes back.
Walkable terrain went in. The steering quietly pulled everyone back down.
Elevated ground is an analytic height field: a plateau has a flat top and a linear skirt, and the ground height at any point is the tallest contribution. Entities snap to it, the puppets read their height from the same field, and a tagged mesh becomes climbable without a navmesh.
First test: the soldiers reached about a fifth of the mesa height and stopped. The steering was the culprit. A move target on the ground below has a lower Z than a climber, so the steering vector pointed partly downhill and fought the slope every frame. Sixty percent on the second try, still wrong.
The fix is to make movement planar: targets carry the mover's own height, vertical velocity is zeroed, and the height field owns Z outright. After that the error between soldiers and terrain was zero at the median across nearly two thousand samples on the mesa top. Rule kept for later: anything that steers should not be allowed an opinion about altitude.
UNDIVIDED · GREYBOX CAPTUREsteering had opinions about altitude03
The crash reporter was telling the truth about a different problem.
Play a one-shot attack animation on a frame that runs long and the clip can overshoot its own end. The single-node instance clamps it back, the animation warping code sees time go backwards, and an ensure fires. On a machine with long frames that happens a lot, and every time it does the crash reporter pops up and looks like a crash.
The war was never actually falling over. Still, a swing that stutters at the end is a bad swing, so the fix is in the puppet: attack clips loop and a timer cuts them at their real length, a swing already in flight is never restarted, and death clips stop a little before their last frame and hold the pose. Zero ensures since across every verification run.
The lesson we keep relearning is that a scary dialog and a bug are two different objects. Measure which one you have before fixing either.
UNDIVIDED · GREYBOX CAPTUREa scary dialog is not a bug02
Four hundred soldiers facing ninety degrees off their direction of travel, and one root component to blame.
The Manny skeletal mesh wants a minus ninety degree yaw to face forward. When that mesh was also the actor's root, the crowd system wrote the entity's heading straight over it every frame, and the offset vanished. Everyone jogged with perfect conviction, sideways.
The fix is boring and permanent: a plain scene component is the root, the mesh hangs off it with its own yaw, and the heading is applied to the root only. Measured afterwards, the fast movers face within about two degrees of their velocity.
The same pass replaced our hand-picked jog loop with the mannequin's own idle-walk-run blend space, driven by direction and speed. It looked better in one afternoon than anything we had tuned by hand, which is the argument for stock animation in one sentence.
Before any tuning is allowed to stand, it has to survive a functional test. Undivided now carries a seven-test suite on its own test map: the rules that decide who parries what, the dominance arithmetic that turns kills and populations into pressure, the wound and reinforcement systems, and the gates that let a faction climb the class ladder. The suite is order-independent and runs from the editor's automation window.
With all seven green the baseline was frozen. Every number that shaped the war lives in one data asset, and that asset is the thing under version control, not a memory of what felt right on Tuesday.
The practical effect is that the visual work that followed, weapons, terrain, the flythrough, could be reckless, because nothing it touched could change the maths without a test going red.
UNDIVIDED · GREYBOX CAPTUREseven green, then frozen
every note above was true on the day. some will be wrong by next month. that is the point of dating them.
