AI Duel Simulation

Run AI-vs-AI game simulations to test decision quality, validate matchups, and catch regressions.

Quick Start

# Default: Red Aggro vs Green Midrange, 5 games, Medium difficulty
rtk cargo run --release --bin ai-duel -- client/public --batch 5

# Single verbose game (see every combat action and spell cast)
rtk cargo run --release --bin ai-duel -- client/public --seed 42 --difficulty VeryHard

# Batch with specific seed for reproducibility
rtk cargo run --release --bin ai-duel -- client/public --batch 20 --seed 1000 --difficulty Medium

# Full registered matchup suite in measurement mode
rtk cargo run --release --bin ai-duel -- client/public --suite --games 10 --seed 42 \
  --output target/duel-suite-results.json

# Compare two suite reports with paired-seed sign-test status
rtk cargo run --release --bin ai-duel -- compare crates/phase-ai/baselines/duel-suite.json \
  target/duel-suite-results.json

# Commander candidate-seat measurement
rtk cargo run --release --bin ai-duel -- client/public --commander-suite --games 8 --seed 42 \
  --difficulty Hard --baseline-difficulty Medium \
  --output target/commander-suite-results.json

CLI Options

Flag	Description	Default
--batch N	Run N games, print summary only	1 (verbose)
--seed S	RNG seed for reproducibility	time-based
--difficulty LEVEL	VeryEasy|Easy|Medium|Hard|VeryHard	Medium
--matchup NAME	Deck matchup preset	red-vs-green
--list-matchups	Show available matchups	-
--verbose	Print every action (full trace)	off
--suite	Run every registered MatchupSpec in measurement mode	off
--games N	Games per matchup/suite cell	10 for suite, 4 for commander suite
--output PATH	JSON report path	target/duel-suite-results.json
--suite-filter STR	Run only suite matchups whose id contains STR	all
--show-attribution	Capture phase_ai::decision_trace policy attribution	off
--commander-suite	Run 4-player Commander candidate-seat rotations	off
--baseline-difficulty LEVEL	Baseline seats for --commander-suite	Medium
--feed PATH	Commander feed under data root	feeds/mtggoldfish-commander.json

Measurement And Gates

ai-duel uses AiConfig::into_measurement(seed) for single, suite, and
Commander regression runs. Measurement mode disables wall-clock search budgets
and bounds search by node/depth budgets, so outcomes are functions of the
binary, config, matchup, and seed.

Use rtk cargo ai-gate for the normal regression gate. It runs the pinned suite
against crates/phase-ai/baselines/, compares paired seeds, and reports
FAIL/WARN/PASS with a binomial sign-test p-value. WARN means movement was
observed but not significant enough to fail the gate.

When intentionally changing AI behavior, refresh the baseline in the same PR
after review:

rtk ./scripts/refresh-ai-baseline.sh
rtk cargo ai-gate

Exception — deadline / wall-clock-budget changes are inert under the gate; do
NOT refresh the baseline for them. Because measurement mode nulls the
wall-clock budget (PlannerServices::with_deadline forces Deadline::none()
whenever execution_mode.is_measurement(), and Deadline::none().expired() is
always false), any code guarded by self.deadline.expired() is a dead branch
during ai-gate, so the run is byte-identical and zero baseline movement is
the expected result — not evidence the change did nothing. If ai-gate
diverges after a deadline-only change, that is a bug (a guard fired on a live
path), not a baseline event: fix the guard, don't refresh baselines to silence
it. Unit-test such guards by injecting
with_deadline(..., Some(Deadline::after(0))) on a non-measurement config —
a .into_measurement() config nulls the injected deadline, so the guard never
fires and the test passes vacuously — and assert services.deadline.expired()
as a witness before the negative assertion. Example: PR #6255 bounded the
quiesce / sample_backfilled_continuations search apply-loops this way with
no baseline change.

Performance Guide

All times are release mode (--release). Debug mode is 5-10x slower.

Difficulty	Time/Game	Search	Use Case
VeryEasy	~1s	None (random)	Stress testing
Easy	~3s	None (heuristic)	Baseline sanity
Medium	~24s	Depth 2, 24 nodes	Primary testing
Hard	~60s	Depth 3, 48 nodes	Quality validation
VeryHard	~126s	Depth 3, 64 nodes	Final verification

Deck Configuration

The ai-duel binary resolves --matchup and --suite entries from
crates/phase-ai/src/duel_suite/spec.rs. Inline starter/metagame deck builders
live in crates/phase-ai/src/duel_suite/inline_decks.rs; snapshot decks live
under crates/phase-ai/duel_decks/.

Available Matchups

Use --matchup NAME to select a preset. Use --list-matchups to see all options.

Starter decks (mono-colored, simple cards for baseline testing):

Matchup	P0	P1
red-vs-green (default)	Red Aggro	Green Midrange
blue-vs-green	Blue Control	Green Midrange
red-vs-blue	Red Aggro	Blue Control
black-vs-green	Black Midrange	Green Midrange
white-vs-red	White Weenie	Red Aggro
black-vs-blue	Black Midrange	Blue Control
red-mirror	Red Aggro	Red Aggro
green-mirror	Green Midrange	Green Midrange
blue-mirror	Blue Control	Blue Control

Metagame decks (real competitive lists from MTGGoldfish feeds, 100% engine coverage):

Matchup	P0	P1	Tests
azorius-vs-prowess	Pioneer Azorius Control	Mono-Red Prowess	Aggro vs control
azorius-vs-gruul	Pioneer Azorius Control	Gruul Prowess	Control vs aggro variant
delver-vs-prowess	Legacy Izzet Delver	Mono-Red Prowess	Tempo vs aggro
azorius-vs-green	Pioneer Azorius Control	Green Midrange	Control vs midrange
delver-vs-green	Legacy Izzet Delver	Green Midrange	Tempo vs midrange
prowess-vs-green	Mono-Red Prowess	Green Midrange	Aggro vs midrange
prowess-mirror	Mono-Red Prowess	Mono-Red Prowess	Mirror match

Changing Decks

To add new matchups, add a MatchupSpec in duel_suite/spec.rs. Use an inline
builder only for small stable decks; prefer snapshot decks for real metagame
coverage. Card names must match entries in client/public/card-data.json. Use
jq 'keys[]' client/public/card-data.json | rg -i "card name" to find exact names.

To find high-coverage metagame decks for testing, check the feed data:

# List all feeds
ls client/public/feeds/

# Check a deck's card coverage against the engine
python3 -c "
import json
with open('client/public/card-data.json') as f:
    db = {k.lower(): v for k, v in json.load(f).items()}
with open('client/public/feeds/mtggoldfish-pioneer.json') as f:
    feed = json.load(f)
for deck in feed['decks']:
    sup = sum(1 for e in deck['main'] if e['name'].lower() in db)
    print(f'{sup}/{len(deck[\"main\"])} {deck[\"name\"]}')
"

Matchup Triangle (Expected Results)

The classic archetype triangle should hold:

• Aggro > Control — kills before control stabilizes
• Control > Midrange — removal + card draw outgrinds
• Midrange > Aggro — bigger creatures brick aggro attacks

Control decks improve more at higher difficulty levels (they need search to time removal correctly).

Mirror Match Testing

For testing AI quality independent of deck matchup advantage, use mirror matches (prowess-mirror, red-mirror, etc.). Win rates should be close to 50/50.

Commander Suite

--commander-suite loads four resolvable decks from the Commander feed, runs
one candidate seat against three baseline seats, and rotates the candidate seat
across P0-P3. Metrics are candidate win rate, survival turns, and elimination
order. Use this as the first strength signal for multiplayer/Commander changes;
it is slower than the two-player suite and should run nightly or on targeted AI
changes rather than on every parser-only change.

Interpreting Results

Healthy signs:

• 0 draws/aborted games
• Games complete in 10-20 turns
• Win rates match expected archetype matchups
• Higher difficulty = longer games (smarter defensive play)

Warning signs:

• Any draws/aborted games → AI might be stuck in a loop
• Games > 30 turns → AI might not be attacking efficiently
• Same player always wins regardless of seed → deck balance issue
• Higher difficulty = worse results → search/evaluation regression

Verbose Output Patterns to Watch

When running single verbose games, look for:

• Self-targeting: "X deals N damage to X" — anti-self-harm policy failure
• Wasteful spells: Combat tricks cast outside combat, counterspells with empty stack
• Suicidal blocking: Blocking at low life when the block damage kills you
• Not attacking with lethal: Having lethal on board but not swinging
• Tapping out into lethal: Casting sorcery-speed when opponent has lethal on board

Community Scenario Fixtures

When a Discord #ai-suggestions thread includes a saved game-state zip, turn it
into a deterministic AI scenario before landing the fix:

1. Download the zip and commit it under crates/phase-ai/fixtures/scenarios/.
2. Add an entry to crates/phase-ai/fixtures/scenarios/community-scenarios.json
   with the Discord thread id, archive name, and expected action assertion.
3. Run the fixture through crates/phase-ai/tests/community_scenarios.rs; it
   loads the same { "gameState": ... } export used by ai-bench-state, applies
   saved-state compatibility migrations, and evaluates the action in measurement
   mode with a fixed seed.

Related Files

File	Purpose
crates/phase-ai/src/bin/ai_duel.rs	Duel simulation binary
crates/phase-ai/src/bin/ai_tune.rs	CMA-ES weight optimization
crates/phase-ai/src/bin/ai_gate.rs	Pinned regression gate wrapper
crates/phase-ai/src/auto_play.rs	AI action driver
crates/phase-ai/src/combat_ai.rs	Combat decisions
crates/phase-ai/src/duel_suite/	Matchup registry, runner, compare, attribution
crates/phase-ai/src/search.rs	Action selection + search
crates/phase-ai/tests/ai_quality.rs	Regression test suite
crates/phase-ai/tests/scenarios.rs	Scenario integration tests
crates/phase-ai/tests/community_scenarios.rs	Discord saved-state scenario tests