Rules

Rules — cogame-vizdoom-deathmatch

Eight cogs, four RED and four BLUE, in one walled arena with cover, glass windows, med kits and nothing to capture. Kill an enemy, take a frag. Die, lose one. After 108 seconds the two teams' net frags are compared and the margin is the score. Nobody is capturing anything; the only thing that happens in this game is that somebody shoots first.

This is not a ViZDoom port — see Divergences. What it reproduces is the ViZDoom multiplayer benchmark's shape: eight agents, one map, an egocentric observation, frags − deaths.

Seats, cogs, teams, aliases

• num_agents = 8, always — in both shipped variants, in the certification fixture and as SMOKE_SEATS in tools/ci/docker_smoke.sh. One cog per seat (cogsPerTeam: 1), so eight seats drive eight bodies.
• Teams are assigned by the server, never chosen. Every variant's slots[] is red, blue, red, blue, …, so slot parity is the team: seats 0, 2, 4, 6 are RED and 1, 3, 5, 7 are BLUE. closedRoster is true, so a joining player takes the slot its token names or is refused.
• Two name spaces. In game a cog is RED-alpha…RED-delta / BLUE-alpha…BLUE-delta, and those aliases are the only names in an observation, a prompt, an order, a shout, a radio line or a sprite label. A seat's real policy name exists spectator-side only: the replay config's players[].name, results.names, and the viewer's scorebug plates and endcard. showPlayerLabels is false in every variant.

The arena

The board is 1235 × 659 px in every shipped variant. Two variants, two ways of choosing the geometry, the same dimensions either way:

Zones. The board is divided once into a 5 × 3 lettered grid: columns A…E left to right (247 px each), rows 1…3 top to bottom. A1 is the top-left corner, E3 the bottom-right. Zone centres are published to every seat in its first observation and are the only place names in the game. RED spawns in column A, BLUE in column E. Each zone carries a one-word terrain label derived at load from its wall fraction: open (< 8 % wall), cover, or corridor (> 22 %).

Weapons, health, respawns

A wipe is unreachable, which is why wipe is not a legal end rule. A cog cannot die more often than once per respawnTicks + 1 ticks, so its ceiling in a 2592-tick game is 52 deaths, and lives is 60.

The clock

• Tick = 1/24 s. Turn = one order round every 108 ticks (4.5 s).
• One game = 2592 ticks = 24 turns = 108 s of sim, maxGames: 1.
• Between turns the loop runs uncapped (fastMode), so the episode's wall clock is the 24 LLM turns.
• Playback is 1 tick per frame at 24 fps, so a full episode plays for 108 s.

Scoring

Per cog c, from counters the sim already maintains:

frags[c]      = kills of ENEMY cogs by c
teamFrags[c]  = kills of OWN-TEAM cogs by c
deaths[c]     = every death of c, whatever the cause
net[c]        = frags[c] - teamFrags[c] - deaths[c]

net[c] is literally frags − deaths, with a team kill charged to the killer as a lost frag — the Quake/ViZDoom convention, and the only thing that stops friendly fire from being a free way to deny an enemy a frag.

teamNet[T]       = sum of net[c] over the four cogs of team T
margin(T)        = teamNet[T] - teamNet[other(T)]        (exactly antisymmetric)
DecisiveMargin   = 12
scorePermille[s] = 500 + clamp(margin * 500 div 12, -500, +500)
scores[s]        = scorePermille[s] / 1000.0             in [0.0, 1.0]
win[s]           = scorePermille[s] > 500

Higher is better. 1.000 is a 12-frag rout, 0.500 a dead-even game, 0.000 the reverse. A red seat's score and a blue seat's score sum to exactly 1.000 for every legal margin.

There is no tiebreak, deliberately. A margin of exactly 0 is a draw: all eight seats score 0.500 and every win is false. A manufactured tiebreak would be a second scoring rule, it would break the exact zero sum, and Elo handles draws natively.

Everything else is measured and shown, never scored: per-seat frags, teamFrags, deaths, net, damageDealt, damageTaken, shotsFired, shotsHit, medkits, longestStreak. damageDealt / damageTaken are recorded so an episode's damage graph is reconstructable from results alone.

End conditions

The episode ends at the first of full time, the wall-clock stop, or a fault. results.reason is exactly one of:

• complete — 2592 ticks elapsed. results.endRule = "full_time". The healthy value.
• deadline — the engine's own wall-clock stop at wallClockBudgetSeconds (660 s). The engine settles from the real frag counters so far, so a deadline episode is still rankable and still exactly zero-sum. results.endRule = "wall_clock".
• fault — an unexpected exception, caught. Both sides score 0.500, artifacts are still written, and tools/ci/docker_smoke.sh fails the build if the smoke episode reports it.

results.endRule is therefore also closed: full_time | wall_clock | sim_fault | host_error. mercy and wipe are gone with the mechanics that produced them.

Budget guard. At the start of each turn, if two more full turns would not fit inside the wall-clock stop, the LLM is switched off for every remaining turn (all seats fall to rusher, microseconds per turn), the remaining ticks run at full speed, and the episode still ends complete / full_time. The comparison uses the LARGER of turnBudgetMs and the derived rate floor.

Orders

Each seat gives its cog ONE order per turn and a deterministic driver carries it out until the seat changes it.

The trigger rule. A is pressed iff all of: the cog is alive; the gun is off cooldown with no windup in flight; a live enemy is known with ticks_ago == 0 (seen this tick, never a memory); it is inside gunRange; the aim error is inside 24 brads; the line of sight is clear; and no teammate's body box intersects the bullet corridor. There is no suppressive fire and no firing at memories. Friendly fire is on in the sim; the driver simply refuses to cause it, so a team kill in a replay is always an accident of movement, never a bot bug.

The two scripted baselines

Both emit the same order object an LLM does, through the same validator. Neither ever emits radio or notes.

rusher (PLAYER_SCRIPTED=rusher) — also the per-turn fallback and the default for a seat that registers with neither field. First matching rule wins: dead → hold at the spawn anchor; a live enemy known within 520 px → hunt it; hp ≤ 1 with a live med kit within 360 px → move_to it; otherwise move_to the contested zone (the zone with the most distinct enemy contacts in the last 72 ticks, ties toward the centre, C2 with no contacts).

sentry (PLAYER_SCRIPTED=sentry) — deliberately weaker and different in shape: it never crosses the centre line unless it is already hunting, and its hunt radius is 260 px. Otherwise it holds an assigned post on its own half, facing the map centre.

rusher — not sentry — is the fallback because a degraded episode must still be a legible deathmatch: eight sentries hold eight posts and produce a 0-0 draw with an empty kill feed.

Divergences

1. This is not a ViZDoom port. No ZDoom, no vizdoom bindings, no WAD, no raw pixel buffer, no depth/labels image.
2. The observation is text, not pixels. The 16-ray depth strip and the labelled contact list are ViZDoom's depth and labels buffers written as JSON. See OBSERVATION.md.
3. Teammates are not fogged (the starter fogs everyone). Four independent policies on one team, each driving one body at a 4.5 s cadence, cannot play team deathmatch blind. Enemies are fogged exactly as the starter fogs them.
4. visionConeDeg 60 → 45. A 90° total cone is ViZDoom's default FOV, and a 120° cone at 1575 px on a 1235 px board is very close to no fog at all.
5. One cog per seat. num_agents is the cog count.
6. The driver refuses to cause friendly fire, though the sim allows it.
7. mercy and wipe are gone. Full time, wall clock or fault — nothing else can end an episode.