Stay alive. Stay inside the lines.
Defense is the half of the fight nobody films: not dying, not walking into the same kill zone twice, and never doing what it was told not to do. Here is how the brain does it, and how much it helps, game by game.
We put the AI models in the tank. Here is what we learned.
A real 3D tank game, each tank with its own driver: our rule pilot, Claude, DeepSeek, a local Llama model, and the game’s own built-in AI — years of tuning. Every match recorded. Totals across every match each one played:
| driver | matches | kills – deaths | K/D | decisions in one 10-min match |
|---|---|---|---|---|
| The game’s built-in AI | 45 | 804 – 437 | 1.84 | — |
| Our rule pilot | 43 | 394 – 649 | 0.61 | 60,140 |
| DeepSeek | 7 | 30 – 55 | 0.55 | 415 (1.0 s each) |
| Claude | 7 | 3 – 47 | 0.06 | 58 (7.6 s each) |
| Llama 3.2 3B, local | 7 | 2 – 33 | 0.06 | 213 (2.6 s each) |
In the same ten minutes, our rule pilot made 60,140 decisions. Claude made 58. A tank that thinks for seven and a half seconds before every move is a tank that is already dead. On a written exam of 8 tank situations with no clock, the rules scored 8/8, Claude 6/8, DeepSeek 5/8 — the models know the answers; they cannot act on them in time.
Five things that keep it alive
- The floor. Before it picks a move it works out what its orders allow. Orders can narrow that list; nothing it learns can widen it. In simulated driving, a controller built to crash sent 64,952 bad commands over 40 km: every one overridden, 0 collisions. Pull the floor and the same commands crashed within seconds.
- Failure memory. One bad break is not a lesson. When a spot keeps getting it hurt, it becomes a rule. A rule reads: died 13 of 19 times here, 3.6× the base rate.
- The map of kill zones. It draws its own map from what its eye has seen, marks where it died, and routes around it.
- The unstick rule. Pinned against a wall, it pivots to the open side, then backs out the other way. In the 3D tank arena, time spent pinned went from 15% to 0%.
- An eye that won’t guess. A sensor claim is checked against the pixels, or the eye says can’t confirm. Gate passed on 7,863 frames: recall 98.4%, false confirmations 0.48%.
Cross-compare: does it survive better?
| game | with the brain | what went against us |
|---|---|---|
| Fallout (1997) — learning off vs on, same save | 15 deaths per hour vs 80 | stalled on one quest for hours; the XP climb also rode 112 commits of ours |
| Drone course — starts it had never seen | 10 of 10 clean laps, 0 contacts; a fresh brain: 0 of 10, 18 contacts | at chase speeds of 2–3 m/s it still crashes |
| Simulated driving — hostile controller | 64,952 of 64,952 bad commands overridden, 0 collisions | this is a boundary; nothing here learns |
| 3D tank arena — pinned against walls | 15% of the time → 0% | still loses to the game’s own AI |
| Quake III engine — hazard floor v1 | fixed in v2 | v1 of the floor caused lava deaths: 38 vs 19 |
| Freeway (Atari) — same mechanism as Space Invaders | Space Invaders +32% | Freeway −12%: caution costs points when danger and the goal are the same road |
Every row comes from a recorded run. The red column is not hidden: a defense that only shows wins is a sales sheet.
When it doesn’t know, it says so
A confident wrong answer gets people hurt. The eye abstains instead. That has a cost and we show it: in the tank arena it held 12 shots at point blank because it will not confirm a target closer than 12 m. A fix is a candidate, not done.
Why a floor, not a bigger neural network
Same situation in, same decision out, every time. That means you can replay any decision at the after-action review and get the same answer. It runs on the machine with no network, so there is no link to jam and no server to take down.