Documentation

sim-v2.5 is the official match engine

You do not move units directly. You submit a WarBrain policy and the server automatically matches it and runs the battle to completion. No model is called during a match, no tokens are spent, and prompts cannot be added mid-battle.

The official engine is sim-v2.5; the policy JSON contract remains sim-v2.2. These are different version lines: one identifies battle rules, the other the compatible strategy format.

What happens after publishing

1. Sign in and choose the country you represent this season. Country is identity only and grants no stat bonus.
2. Build a policy by hand, or copy the server instruction from Connect to your AI so it can read rules and records and save a draft.
3. After you confirm publish, the server immediately completes five placement battles.
4. Every 5 minutes the server starts a small matchmaking batch, prioritizing the strategies that have waited longest. Each rated strategy plays about once every 60 minutes with a 55-minute cooldown; when selected, it faces a nearby-rated opponent from another account and country. No page or battle button is required.
5. Read the stored replay and prediction ledger, revise the policy, and publish a new version.

One continuous 150-second battle

Operation score is the primary win line. The server draws one operation and the higher accumulated score wins at 150 seconds; an exact tie is a draw. Operations define which actions turn into points, not a fixed doctrine. The seed also creates one mirrored, fair seven-node supply graph shared by both sides.

Destroying a core ends the match immediately, but full damage requires the attacker to own a supplied relay beside that core. A raid without that bridgehead deals only 15% core damage. Losing ownership or supply removes full damage immediately.

Plan windows and command points

• Both sides seal a major plan simultaneously every 30 seconds; neither can wait to see the other's new order.
• Every 5 seconds the policy reads allowed observations. Rules are checked in array order and the first match fires.
• A major window refills command points to three. Tactical changes between windows cost command points, replan allowance, and retool time.
• Intel investment can buy more tactical reactions, but it never reveals hidden enemy orders directly.

Observe, hypothesize, deploy, test

predict is not a global buff. It is a battlefield hypothesis that expires every plan window: you predict the enemy’s dominant fast, safe, or flank route. The route must appear in observed in-match plan history; without evidence, the command is rejected.

• A prediction costs one command point.
• A correct call with no matching counter posture at the actual engagement has zero effect.
• A correct, deployed counter gives a smaller informed force enough local leverage to beat a larger blind one.
• A mismatch causes no hidden damage and no direct operation-score penalty. Its costs are the command point, mission commitment, and the opportunity lost by deploying to the wrong place.
• The replay records evidence, prediction, deployment contact, match rate, and confirmed / mismatch / untested. Reinforcements show dispatch, transit, and arrival instead of appearing from nowhere.

The WarBrain policy contract

A policy is bounded JSON data, never code. It contains opening investment, production, three squad missions, and a limited rule list. The cap is 12 rules, 8 memory values, and complexity 96; loops, network requests, and arbitrary code are impossible.

{
  "schema": "sim-v2.2",
  "id": "example-counter",
  "name": "Example Counter",
  "authoredWith": "your model or Human Only",
  "memory": {},
  "opening": {
    "allocation": { "force": 0.5, "logistics": 0.25, "intel": 0.25 },
    "production": {
      "units": { "scout": 0.25, "striker": 0.5, "engineer": 0.25 },
      "reinforce": [2, 4, 3]
    },
    "squads": [
      { "target": "weakest_visible", "posture": "evasive", "route": "flank" },
      { "target": "highest_value", "posture": "guarded", "route": "safe" },
      { "target": "nearest_neutral", "posture": "aggressive", "route": "fast" }
    ]
  },
  "rules": []
}

Automatic matching, player-owned solutions

The platform provides fair rules, pairing, and settlement, but never solves opponents for a player. The ladder reads every placed strategy from the database; all 196 country seeds are full system competitors under the same pool, engine, and rating rules. A country uses its strongest current strategy rating, so volume cannot inflate its rank.

A replay is evidence, not re-simulation

• At settlement the server stores the v2.5 map, observable snapshots, event timeline, and prediction ledger.
• Produced units are stored as transport records from core dispatch to squad arrival; supply cuts can delay them and the replay shows the updated journey.
• A replay contains neither policy source, hidden memory, nor server entropy keys.
• A future engine update cannot change a finished result; playback shows the facts stored at that time.
• Animation stops on the actual terminal frame and restarts only when the viewer presses Replay.
• Historical v1 records remain readable and are explicitly marked as old-engine records, never disguised as v2.5.

A server-only API workflow

1. The AI reads public /api/agent/context for current v2.5 rules, observations, commands, and budgets.
2. With an Agent Key it reads /api/agent/workspace for server-persisted policies, records, and briefs.
3. It states evidence first, then develops a native sim-v2.2 revision.
4. It fixes every /api/agent/validate issue and saves through /api/agent/drafts.
5. The AI can save only a pending draft; only the signed-in player can publish it to the ladder.

FAQ

Does country affect combat?

No. Country is season identity and a leaderboard grouping. It never enters the engine and no real-world territory is occupied.

Can offline token spending guarantee first place?

The platform cannot know how many tokens someone spent offline, so it makes no false claim to limit that. What it can enforce is equal in-match budgets, a hard policy-complexity cap, observable and counterable fixed scripts, and no uncounterable universal solution from more search.

Will the platform beat past champions for me?

No. Test programs use champion pools to audit the mechanism, but online policy design, opponent research, and iteration belong to the player or their authorised AI.