AI Break Duel Balance Regression

AI Break Duel Balance Regression

Note: This skill covers the biased-cost stress-deck regression check only. For the full balance-tuning workflow (league verification, excitement metrics, adoption criteria, and docs/balance-history.md recording), see .agents/skills/ai-break-duel-balance-tuning/SKILL.md, which uses this skill as one of its verification steps.

Workflow

1. Confirm the working directory is an ai-break-duel checkout (the main repository or one of its git worktrees; CPU/balance work may live on a non-main branch — follow the branch noted in docs/balance-history.md or the active plan document).
2. Inspect current uncommitted changes with git status --short; do not overwrite unrelated changes.
3. Run the stress-deck regression with the TypeScript CLI (same engine as the browser game):

npm run balance:cost -- --games-per-order 1000 --seed 3000000 --out tmp/cost-balance-3000000.json

The TS CLI supports --rule-set current only (the experimental Python rule sets were retired with the Python simulator on 2026-07-08). It always uses challenger versus challenger. Use --candidate <id> to restrict to a single stress candidate.

4. For an existing six-deck round-robin league (previously --include-preset-league), run:

npm run sim -- league --games-per-pair 1000 --seed 3000000 --decks break control fire water wind earth --out tmp/preset-league-3000000

Use lower --games-per-order values such as 100 only for quick smoke checks. Use 1000 or more for user-facing balance reports. Candidates use the deck-builder template (updated 2026-07-04 for the 25-card deck rules): 19 summon slots from the target band, plus 4 generic commands and 2 generic relics, 25 cards total. Candidate generation also applies the current construction rules: at most 2 copies per card ID, and power 3+ summons are capped at 5 total under --rule-set current, with remaining summon slots filled by legal low-power summons when a target band cannot supply 19 legal summons.

5. Report candidate_win_rate against all six existing decks combined, but do not use that number alone as the final risk call when the loss is mostly concentrated in the mono-attribute decks. The mono-attribute decks (fire, water, wind, earth) are allowed to be weaker because they are theme / onboarding samples. The competitive baseline is the multicolor pair break and control; treat a stress candidate as a stronger balance risk when it also beats the break/control subset combined above 50%, or when match-shape metrics such as one-sided rate are unacceptable. Always report six-deck combined, break/control combined, and per-existing-deck rates.
6. When --include-preset-league is used, report existing-deck standings, first-player win rate, average turns, one-sided game rate, and resource-exhaustion rate before candidate stress-deck results.
7. If the user asks to preserve the guard in tests, keep src/game/costBalance.guard.test.ts aligned with src/sim/costBalance.ts candidate definitions and thresholds.
8. After implementation changes, run npm run check. If npm is not on PATH (Codex runtime), use:

PATH="/Users/user/.cache/codex-runtimes/codex-primary-runtime/dependencies/node/bin:$PATH" npm run check

Candidates

The regression set is:

• p1: power 1 stress deck
• p2: power 2 stress deck
• p3: power 3 cap stress deck; low-power filler may be added
• p4: power 4 cap stress deck; low-power filler may be added
• p1_2: power 1-2 stress deck
• p2_3: power 2-3 stress deck; high-power cap may add filler
• p3_4: power 3-4 cap stress deck; low-power filler may be added

Do not call p3, p4, or p3_4 "power 3 only", "power 4 only", or "power 3-4 only". Under the high-power cap, those candidates are cap-stress decks with legal filler, not pure high-power decks.

The baseline opponents are the existing six deck archetypes: break, control, fire, water, wind, earth.

Evaluation Criteria

• break and control are the current competitive multicolor baselines. Multicolor decks are expected to have better strategic depth and higher deck efficiency than mono-attribute samples.
• fire, water, wind, and earth are mono-attribute theme / onboarding decks. It is acceptable if they underperform optimized or biased stress decks, as long as their play patterns remain clear and they are not the only evidence for a balance conclusion.
• Keep reporting the six-existing-deck combined win rate because it shows the full preset environment, but distinguish it from the competitive break/control subset. If a candidate is high only because it farms mono decks, say that explicitly instead of calling the whole rule set broken.
• One-sided game rate remains a primary quality signal. A candidate can still be a problem even with acceptable break/control rates if it creates too many one-sided games.
• When updating preset decks for this rule family, preserve mono-attribute identity unless the user explicitly asks to make mono decks competitive.

Reporting

Include:

• games per ordered matchup and seed
• rule set name when using --rule-set
• CPU profiles; default is challenger/challenger
• existing-deck league standings when included
• first-player win rate, one-sided game rate, resource-exhaustion rate, and average turns
• six-deck combined win rate for each stress candidate
• break/control combined win rate for each stress candidate
• per-opponent win rates when a candidate is near or above 50%
• whether any candidate violates the design goal
• the exact validation command run

Historical note: the retired Python script supported experimental rule sets (high_cap_*, p3_cap_*, proposed_action_cost*, etc.) used for one-off rule audits. Their conclusions are recorded in docs/balance-history.md (see the 2026-07-06 entry for the power 3+ cap curve). The TS CLI intentionally supports only the adopted current rule set.

Do not call a candidate "strong" just because it beats one or two archetypes. For the main conclusion, separate the full six-deck environment from the competitive multicolor baseline. A candidate that only farms mono decks is a mono-deck weakness finding, not automatically a rule-set rejection.