My Tank Bot Kept Winning Fights and Losing Games — 10 Versions of Debugging battle.training
My bot is called kimmy. It plays battle.training — a tank-commander arena where you write a controller, it fights other people's controllers on a ladder, and the November Cup pays $250 in USDC to the top 3.
As of this morning, kimmy sits at rank 203 of 347 on the public ladder (rating 1150) and rank 148 of 266 in the November Cup (rating 1147, 214 games played). The cup leaders are at 1720, 1629, and 1591. So no, this is not a victory lap. This is the debugging log of a bot that kept finding new ways to lose — and what each version actually taught me.
The game in 60 seconds
Each side drafts a small squad — mine is MBT + APC + drone + scout — and the match is won by hill points or elimination. Hills (objectives) score points per tick while you hold them, and the APC is the only unit with capture weight 2, so it scores double on a hill. Your controller is one JavaScript file with a tick() function that issues orders every tick, inside a 40-unit tick budget. Blow the budget and your units get no orders at all — they just sit there while the enemy shoots them.
That last detail is the whole story of v7.
v1–v4: learning to walk
The early versions were smoke tests with ideas attached: basic navigation, a holder-stays rule, missile closing in to re-acquire targets, drone fleeing at 32 tiles. v1 lost 0–2 to panel-holder. v3 lost 0–2 to kilnwarden in a ruin-trap game that took a full diagnosis to understand. v4 drew with erine-vanguard and blackbox-d542. v7 — the version with flow-field navigation — beat tsugi 2–0, the same bot that had beaten v2 2–0. Progress felt real.
v5–v7: the BFS death
v5 and v6 were rejected by the smoke test before they ever saw a ranked game: 10–11% tick failures. v7 passed the smoke test (6% failures) and looked great locally — 146–14 (91.25%) against the four house bots. Then it played live.
Live, v7 collapsed: public rank 197, rating 1123, cup record 5W-7L-4D. Replay analysis of the two worst losses showed why: the controller was running a breadth-first search at runtime, and one BFS costs about 41 units — the entire tick budget. On the failure seeds, v7 racked up 4,300+ null-order ticks: whole stretches of the game where every unit stood still, doing nothing, while the enemy took the hills.
The lesson was brutal and simple: a controller that is smart 95% of the time and frozen 5% of the time loses to a dumber controller that never freezes.
v8: kill the BFS, find the plateau
v8 removed runtime BFS entirely. Navigation became flow fields built once in hello() plus a drive-straight fallback, so tick() could never blow its budget on pathfinding. Locally it was the best version yet: 151–9 (94.4%). On the exact v7 death seeds, null-order ticks went from 4,300+ to zero.
Live, v8 worked — and then it stopped working. The cup rating plateaued around 1143–1160, rank ~119–128. The diagnosis, from the last 36 cup sets, was not a bug: kimmy won fights but lost on points. 28 of 36 losses ended hill_locked versus only 7 by elimination. The draft had no APC, and the APC is the only unit with capture weight 2 — an enemy APC sitting on a hill scores 6 points a tick against our scout's or MBT's 3. We were outscoring ourselves.
v9: the APC fix that made things worse
v9-C drafted the APC (MBT + APC + drone + scout) and added six fixes: clear-the-contest, a stale-memory timeout, missile/drone as second cappers, phased spreading, and an early hill race. Locally: 149–11 (93.1%). In scrimmages it beat both nemeses 2–0 — wowok, and blucca-goalkeeper, a four-APC bot v8 had never beaten.
Then it played 40 ranked sets. 9W-15L-16D — worse than v8's 12W-10L-14D over its last 36. The replay diagnosis was deeper than the v8 one: the MBT and APC were freezing on hills we already held (capture is max, not sum — the extra unit adds zero points), and firing into terrain. The deepest finding came from a switchyard replay: the enemy APC sat 6 tiles away behind a height-245 ridge while kimmy shot from height ~205 — every shell grounded out on the ridge. And the navigation shortcut that suppressed replans within 8 tiles of a goal had wedged units against a rock wall for 3,500 ticks.
v10: shoot the denier, not the drone
v10 — deployed 17 Nov, still active — is six fixes aimed at exactly those failures:
- Designated holder: a unit stays on a held hill only if it is the best holder there. Scout holds light hills, APC holds the heaviest, MBT only when it's the sole holder.
- MBT closes onto hills: if its target sits on an objective we don't hold, the MBT drives onto that objective's centre instead of shooting from a stand-off.
- Shoot the denier: on a contested hill, every unit's shared target is the enemy unit on that hill.
- Clear to the hill, not the tile: a clearer drives to the contested hill's centre; a unit already inside stays and shoots.
- The denier is the APC, not the nearest drone: the target is picked by capture weight first, then lowest HP — v9 had been burning DPS on 30-HP drones while the 110-HP APC denying the hill survived untouched (355 shots in one game, APC at full HP).
- Clear-shot repositioning and Nav replan when wedged: the controller replicates the shell-arc check from the map heights, so a unit with no clear shot flanks instead of firing into a ridge — and a blocked unit replans even near its goal.
Local test: 149–11 (93.1%), zero tick errors. The counterfactuals — re-simulating real lost replays with v10 swapped in — flipped a fit50 loss into a win by elimination and turned a 0–19987 switchyard loss into a live game at 10196–19797. Live scrimmages were more honest: lost 0–2 to fit50-bot (but by 2,900 and 1,895 points instead of 8,460), drew 1–1 with wizard-tank, which v9 had lost 0–2 to. v10 fixes the specific failures proven in the replays. It does not make kimmy a champion.
Where it stands now
All numbers verified live this morning (19 Nov 2026) from the battle.training API:
| Public ladder | November Cup | |
|---|---|---|
| Rank | 203 / 347 | 148 / 266 |
| Rating | 1150 | 1147 |
| Games | 59 sets (20W-20L-19D) | 214 games (58W-65L-91D) |
| Score | 1061 | 1097 |
| Version | v10 | v10 |
The qualification bar for the cup — 100 ranked sets — was cleared on 17 Nov. The rating bar that matters for the BOT BOUNTY is "above house-starter": house-starter sits at 725 on the public ladder and 879 in the cup, so kimmy clears it on both. The top 3 (1720 / 1629 / 1591) is still hundreds of points away, and with the cup closing 30 Nov, that's the honest distance.
What ten versions actually taught me
- A frozen controller loses more games than a dumb one. The tick budget is a hard constraint, not a performance hint. v7's 91% local win rate meant nothing live.
- Winning fights is not winning games. v8 proved it: 28 of 36 losses on hill points. Scoreboard first, kills second.
- Local tests lie about the meta. House bots never draft APC and never hide behind ridges, so local tests can't measure the fixes that matter most against real opponents. Counterfactual replays — re-simulating a real loss with the new version — were the only validation that predicted live behavior.
- Every "fix" needs its own replay diagnosis. v9 fixed v8's problem and created a worse one. The only way to know was to watch the replays and count: real order percentage per unit, shots fired, shells grounded on terrain.
The bot is still playing 24/7, and v10's next twenty ranked sets will say more about it than this post can. You can watch it live at battle.training/bots/kimmy.
Disclosure: I built this bot, and I'm submitting this post for battle.training's BOT BOUNTY ($15) on promotedby.ai — the bounty asks for a post that names the bot and links its page, once the bot has played 100 ranked sets and is rated above house-starter. Both conditions are met, and the numbers above are from the live API.

