Skip to content

Each bot in detail

The one-screen table is on Bots. This page says what each registered bot does, then how its fair-or-clairvoyant verdict was reached, with the code that decides it. The six below are bots/all's, in every binary that links it; the policies gorged adds follow. To select or connect a bot, see Using and plugging in bots.

The six registered bots

  • bot fair

    The default. Rules, a combat simulation, Forge's mulligan.

    gorge, dirac. Browser, HTTP, SpellBench agent. 29.7 µs a decision.

  • forge-heuristic fair

    bot plus rules re-expressed from Forge's AI.

    gorge, dirac. Browser, HTTP, SpellBench agent.

  • forge-heuristic-hold fair

    forge-heuristic that keeps instants back.

    gorge, dirac. Browser, HTTP, SpellBench agent.

  • lethal-pressure fair

    The older heuristic with an attack-into-lethal rule.

    gorge, dirac. Browser, HTTP, SpellBench agent.

  • sb-tactical fair

    SpellBench's scored tactical heuristic.

    gorge, dirac, bot-only. Reads decklists. Not for people.

  • az-redeal fair

    AlphaZero-style search, 100 simulations, no network.

    gorge only, bot-only. Reads decklists. About 106 ms a searched decision.

All six are fair about hidden cards. The two search bots read every seat's decklist, so they play only against other bots. The side-by-side table is on Bot comparison.

What each does

Six policies register themselves in the bot registry (bots/registry.go) from the packages bots/all links, so every binary that imports bots/all has them. Each entry's bots.Info carries its label, tier, formats, seat limit, description and cost. Bot speed measures them.

bot is the production heuristic: hand-written rules (bots/botpolicy, driven by bots/heuristic) that score each offered option. On a table with automatic mana payment, the default for hosted bots, it casts through the engine's payment plans and picks attacks and blocks with a two-ply combat simulation; on a manual-tap table it is the plain heuristic. It answers the London mulligan with a re-expression of Forge's keep and bottom rule (botpolicy.MulliganForge).

forge-heuristic is bot, built by the same constructor (bot.BaseWithSim), with every opt-in rule family re-expressed from Forge's AI switched on. Each family is a clean-room rule in gorge's own vocabulary, not translated Forge code:

Family Rules armed
Removal timing hold a one-shot destroy, damage or bounce spell until the best opposing creature is a real threat, the spell is lethal to a face, or it could never be cast later
Targeting prefer a land the opponent's colours depend on and the opponent's prominent colour; skip a target that is a strictly worse copy of another of their creatures
Forge AI hints a card's PlayMain1 hint earns a first-main-phase bonus; a RemoveDeck card is never chosen
Deck-shape posture an aggro deck releases its last held-back blocker, the combat simulation weighs life by deck shape, and the mana reserve scales by shape
Attachments aim an Aura by its AttachAILogic hint, rank its own creatures as carriers, hold a boon Aura while nothing can carry it
Timing counter only the top foreign spell; hold draw and token spells for Forge's window
Combat, inside the simulation gang blocks, a gang-aware predicted reply, Double Strike in blocked combat

The rules live in bots/botpolicy (TargetRules, RemovalRules, TimingRules, AIHints, AttachRules, PostureRules, CombatRules), and bots/forgeheuristic.New shows how to arm any subset on a policy of your own.

forge-heuristic-hold is forge-heuristic with its one option, hold-instant, switched on: it keeps instants back outside the respond, opponent post-combat and combat-trick windows.

lethal-pressure is the older configuration: the plain heuristic plus an attack rule (a creature whose unblocked power reaches the defender's life attacks even into a cheaper blocker). It has no combat simulation on any table, and it keeps the heuristic's original mulligan, a one-in-three coin whatever the hand.

sb-tactical is a scored tactical heuristic ported from the SpellBench project. It values creatures, timing, the damage race and removal over what its seat can see, reads printed card facts from the card registry, and at priority prices its plays through a payment planner built on an honest root: a copy of the game with every hidden card redealt from what the seat has seen. That redeal draws from each seat's real decklist, which is why the policy is marked Env and is never seated against a person.

az-redeal is an AlphaZero-style tree search (bots/azmcts) with no network: a uniform prior and a heuristic leaf value. At each priority, attackers, blockers and target decision it runs 100 simulations, each over a freshly redealt world built from its own observations, and bot answers every other decision. It is Env like sb-tactical. A hosted search seat has a 2-second budget per decision (hostedDecisionDeadlineMS); a search that runs out plays bot's answer.

The policies gorged adds

cmd/gorged, the server and the demo's binary, links further experimental policies beside bots/all (cmd/gorged/main.go). They stay out of bots/all so server/host's registry-wide gates keep the built-in set's cost; bots/hosttest runs their host-level gates instead.

  • cast-profile is bot playing from a cast profile. The embedded default profile is intent-identical to bot, so the entry exists to serve a fitted profile beside it.
  • sb-first, sb-uniform, sb-heuristic are SpellBench's three builtin policies behind one hosted adapter. Each answers from its seat's view; at priority it uses the honest root, when the host built one, as its payment planner.
  • sb-search-lite-atk is sb-search (bots/spellbench/sbsearch): at priority it prices sb-tactical's pick and its next best plays over four worlds redealt from its own observations, rolled two turns forward with sb-tactical on both sides, and plays an alternative only when it beats the pick by the margin. It searches an attack declaration the same way and answers everything else as plain sb-tactical.
  • search is the search teacher (bots/searchseat): it replans the decisions it covers over sampled worlds rebuilt from its own observations and plays bot everywhere else.

Fair or clairvoyant

A fair bot uses only what its seat may see. A clairvoyant bot reads hidden information: the opponent's hand, the library order, or a clone of the real engine. No registered bot is clairvoyant. The two search bots are fair about hidden cards but read every seat's decklist, so they never seat against a person. Verdicts, with where the code decides them:

  • The four heuristics are fair. Decide takes the seat's redacted view.View, and DecideBoard takes a Board of public facts (bots/heuristic/bot.go:616, :627, BoardFromView at :642). They register without Env (bots/bot/bot.go:12).
  • sb-tactical is fair on hidden cards and reads decklists. Env: true, Search: false (bots/sbtactical/sbtactical.go:50). Its planner root is redealt from the seat's own feed (bots/searchseat/honestroot.go:30-53), but the redeal draws from every seat's real decklist (bots/registry.go:205), so -bot-policies refuses it (server/gorged/botpolicies.go:53).
  • az-redeal is fair on hidden cards, determinised, with no belief model, and reads decklists. Env: true, Search: true (bots/azredeal/azredeal.go:52); Hosted sets WorldRedeal (azredeal.go:66-70, bots/azmcts/redeal.go:19-21).
  • cast-profile is fair. It is a heuristic.Bot registered without Env (bots/castprofile/castprofile.go:15).
  • The three SpellBench builtins are fair on hidden cards and read decklists. Env: true (bots/sbbuiltin/sbbuiltin.go:40); they answer from env.View and take the honest root only as a payment planner (sbbuiltin.go:75-81).
  • sb-search-lite-atk and search are fair on hidden cards, determinised, and read decklists. Env: true, Search: true (bots/sbsearch/sbsearch.go:60, bots/search/search.go:44). sb-search deals its worlds from the seat's own feed (bots/spellbench/sbsearch/sbsearch.go:709); the teacher's hosted config is not clairvoyant (searchseat.Defaults, pinned by TestHostedSearchConfigIsHonest).

Two further points.

  • The clairvoyant world exists, for benches only. azmcts.WorldClairvoyant clones the real engine (bots/azmcts/redeal.go:14-18), and searchseat.Options.Clairvoyant searches a clone of the actual engine (bots/searchseat/searchseat.go:149-153). They are ceilings for measuring what hidden information is worth. The hosted az-redeal leaves SeatConfig.Source nil, which NewSeat refuses for a clairvoyant world (bots/azredeal/azredeal.go:16-18, :64-65), and a seat whose honest root is refused plays its fallback, never a clairvoyant one (bots/searchseat/honestroot.go:49-52).
  • A belief model exists. azmcts.WorldRedealClosed closes the opponents' lists: their unknown cards are dealt from a list drawn from a belief corpus (bots/azmcts/redeal.go:22-28; bots/azredeal/azredeal.go:100-105, ClosedFactory). Such a seat declares ClosedLists and may sit against a person; the host checks the declaration at match start (server/host/closedlist.go:23-38). gorge ships no trained network and no belief corpus. A binary links a policy that needs a network with gorged.RegisterNetBot (server/gorged/netbots.go:76). Net runtime details are on Search and nets.

Cost and engines

Costs are on Bot speed, measured 2026-10-31. The registry's own Cost field is a declared number, not a measurement, for every bot but az-redeal (bots/bot/bot.go:14, bots/sbtactical/sbtactical.go:45-50, bots/azredeal/azredeal.go:47-52). On dirac, bot, lethal-pressure, forge-heuristic, forge-heuristic-hold and, at a bot-only table, sb-tactical take a seat. Searching policies are refused because dirac deals no search roots (server/host/table.go:340, engine/game/gorgegame/gorgegame.go:130; dirac server). dirac's hbot is a port of bot and sbt runs gorge's sb-tactical through dirac/oracle (dirac/cmd/dplay/game.go:20-21).

Not registered

bots/searchseat (a determinized rollout seat, which search hosts), bots/searchprobe (the sampler, redealer and belief) and the learned-policy runtime in bots/policynet are building blocks, not registered policies. bots/explain is the per-decision note a match record carries, and bots/sanity is the guard -bot-common-sense wraps around a policy.