Treadcore 0.2.0 · the AI Brainz plan · 26 October 2026
A tank that senses the battlefield the way a commander does through a periscope, decides with 15,791 numbers, and learned to drive to an enemy, a depot or a pillbox and to shoot in under four minutes of evolution on a laptop. Then the census put it in a real match.
The net never sees the world's lists. It feels a compass of eight sectors round its hull and reads four exact contacts, as a submarine reads its plot. Drag the contacts round the tank. Ahead is up.
8 px = 1 tile · drag or focus a contact and use the arrow keys
Five blocks, each from 0 to 1 (the periscope's bearings from −1 to 1). Hover a cell to name it. Every one is a few comparisons over lists the tank already sees, so the net runs in microseconds on any CPU.
| Instrument | Like |
|---|---|
| x, y as shares of the map | GPS |
| heading sine, cosine | gyro compass |
| nearest edge | chart |
| a hit just taken, fading over 2 s | hull-hit light |
| Compass channel | Like |
|---|---|
| hostile, allied tank | gunner's sight, IFF |
| pillboxes, depots (ours, theirs, neutral) | the map |
| shell coming | missile-approach warner |
| ground, a ray 16 tiles out | depth sounder |
| cover, a friend's call | driver's eye, radio |
A team's learned tanks share one queue of requests. A call lives 10 s, the queue keeps 16, and each tank calls at most every 2 s. Every call is also a Ping in the world, so scripted teammates answer it too, and their pings come back into the queue: one language for both brains.
Command::Ping → Event::Pinged: scripted allies convergeEvolution strategies: each generation tries 48 random nudges of all 15,791 weights, forward and back, on the same 12 training grounds (1,164 one-minute episodes), ranks them, and steps toward the nudges that scored better. One generation takes about a quarter of a second on an M1's eight cores.
Score of the net itself on its generation's 12 grounds. Phase 1 pays up to 2.5 a ground (the way closed, a reach, the sooner); phase 3 pays 3.2 for three direct hits and a kill. The periscope net passed every ground by generation 67.
| Net | Enemy | Depot | Pillbox |
|---|---|---|---|
| untrained (133 senses) | 13.2% | 1.8% | 0.3% |
| phase 1, compass only, 500 gens | 99.7% | 100% | 95.2% |
| phase 1, periscope, 500 gens | 100% | 100% | 100% |
| phase 3, shoot only | 100% hit | 9.3% | 23.7% |
| phase 3, drive and shoot | 100% killed | 100% | 100% |
The same census that judges every weight change seats a team of five nets against the scripted defaults, 110 seeds, 11 maps, three matches a seed. Zero is the scripted brain.
The trained team reads 3.17 ± 0.45 below a random team. On the grounds nothing shot back and every pillbox was neutral; in a match the nearest pillbox is often hostile and the nearest enemy is a team of five under its own guns. It drives straight at them. The census did its job: it says by how much this is not yet a better tank, and points at what to teach next.
| Scripted Commander | Learned NetBrain | |
|---|---|---|
| Judgement lives in | a ladder of rungs, a team board, 110 named weights | 15,791 weights |
| Reads | the whole observation, worked into facts and an influence map | 133 senses |
| Improves by | make converge: fit, suggest, census, held-out reads | make train: evolution, a phase at a time, each gated |
| Can do | the whole game: supply, capture, forts, bridges, drones, teams | drive to a target and shoot an enemy |
| Latest | dodge_horizon 2 adopted: ten held-out reads, +1.09 to +2.66 | passes phases 1 and 3; the census says not yet |
| Upgrade | Test, written first |
|---|---|
| Live: hostile pillboxes, an enemy that fires back, shells landing; armour lost costs | the dodge families' 99%, then a census above the random team |
| Aim at a mover, then one that shoots back | hits a shot at least the scripted aim's |
| Go round water, walls, forest on the real maps | 90% within 3 tiles |
| Duel the scripted brain at skill 0.3, 0.6, 1 | half the duels won |
| Calls trained: one net sees the enemy, one does not | the blind one arrives sooner |
| GPU: behaviour cloning then PPO in Burn on the 4060 Ti | the census, two held-out reads |