The town that fell silent

July 18, 2026 · land · balance

The land probe kept logging the same crime scene: by minute four, all twelve hulls frozen, the closest enemy pair parked at 90 m, and the score untouched for six straight minutes. Ten minutes of “war” produced three kills from 67 shots. Nobody was stuck on a treeline. Nobody was out of ammo. The whole front had simply… stopped.

Watching the freeze form

The probe grades the battle; it doesn't explain it. So the stall got its own magnifying glass — a diagnostic that steps the same AI-only match and, minute by minute, dumps every hull's decision state: where it is, how far it actually displaced, what its bow is grinding on, whether its own eyes (not the team's) cross its target, and — the tell — which gunnery gate is refusing the trigger. The freeze formed on camera. Both platoons converged on the market town, parked 30–260 m apart with targets spotted and sightlines clear… and every stalled gun showed the same reading: dPitch ≈ 0.05–0.21, pinned, for hundreds of seconds.

Lock one: the gun that couldn't look down

Land mounts ride high on purpose — the world's terrain steps in quantized 3 m blocks, so a low bore stares into the first uphill lip in front of the hull. The sight head sits 7.2 m over the dirt. But the mount still used the flak battery's depression stop, tuned for guns at head height. Do the trigonometry and it lands like a verdict: from 7.2 m up, a stop of −0.1 radians means everything inside ~56 m sits below the barrel's reach. A whole town block of dead zone. The AI aims exactly — solves the ballistics, slews the mount, checks alignment — and the alignment check waited politely, forever, for a lay the trunnions made impossible. Players had it no better: park an enemy hull under your nose and your own barrel refused to touch it.

The fix is three small honest pieces: land bores get a real depression stop (−0.35 — the dead zone shrinks to ~15 m, two hull lengths, closer than armor ever fights), the client sight mirrors it, and the AI now chases the clamped solution — it fires the best lay the mount can give instead of holding fire on a perfect one it can't. A regression test pins the scene: a Paladin and a Warden on level ground at 35 m must open fire and connect. Without the fix, that test sits silent to the end of time.

A Paladin main battle tank at a treeline with a burning enemy hull directly in front of its bow, reload timer counting down, HULL BLOCKED banner overhead — Point-blank works now: an enemy hull burning a bow's length away — a shot that was geometrically impossible last week.

Lock two: standing a flag against ghosts

With the guns talking, the diagnostic caught a second, quieter deadlock. Two Jackals — one Crimson, one Cobalt — drove to the south village, parked 257 m apart on opposite sides of the same rise, both inside the zone, each spotted on the other's map, neither able to see the other with its own eyes. The garrison doctrine said stand the flag. So they stood. Eight minutes, zero movement, the zone contested and scoreless the entire time. Two parked trucks staring at the same hill was the whole war.

Now a garrison holding its zone against a spotted intruder it cannot see sweeps toward the contact — a committed straight-line advance, no wheeling (armor advances, retreats, or holds; it does not dance) — and the moment the sight crosses the target, the existing halt-and-fight logic takes the gun. Zones flip again. Rounds end again.

The after-numbers

Same probe, same ten minutes: shots 67 → 165, kills 3 → 7, and the score marches every minute instead of flatlining — one run put a cap win on the board at four minutes, redeployed, and had the second round at 581 points before the probe ran out. The hit ledger changed shape too: the baseline was front-plate shatters (parked hulls plinking armor they couldn't beat); now it's penetrations through sides and turrets — movement makes flanks, flanks make kills. And from the player's seat, driving into the brawl got an ending the old build couldn't deliver: our playtest Paladin died gloriously at 300 m, mid-maneuver, to bots that a week ago would have watched us park.

One confession from the process: the first version of the regression test passed without the fix. Its terrain scan had picked a duel ground where the target sat a couple of metres uphill — inside the old clamp's reach after all. A test that can't fail isn't evidence; it took a level-ground constraint (and a run against the unfixed code to watch it go red) before the pin was real. The diagnostic that found both locks stays in the tree (LAND_DIAG=1), because the next frozen war will want a camera on it.