Final balance-study selection — 2026-09-09
No finalist advances to holdout in this study. This decision follows all four completed cautious-policy conditions and independent review of the control-selected profile. No seed from the final 600000 range has been accessed. The study delivers a root-cause diagnosis and experimental candidates, not an accepted balance replacement. Permanent offensive Enraged remains the playable combat default. The requested map-symmetry repair is implemented separately.
Evidence behind the decision
The completed program contains 280,320 full-game executions, including 6,144 exact control/parity repeats, plus 153,600 isolated adjacent melee exchanges. Conditions sharing seeds are dependent comparisons, not independent replications. Original source snapshots, unsuccessful conditions and adaptive selection records remain available.
The final exploratory grid comprises four conditions, five melee-involving pairings, four maps, eight shared seeds, 25 joint aggression settings and four roster-end/first-side assignments: 64,000 executions. It does not supply the other ten family pairings under the new policies. Control-derived pure maximin/minimax policies select cautious non-melee behavior in 18 of 20 trained cells. The full retained profile explicitly marks the other 40 map/pair entries as untrained 0.75/0.75 fallbacks.
At the same control-selected policies, melee payoff across the five opponents is 75.3% for control, 66.4% for pool minus one, 69.1% for ranged accuracy plus two and 53.5% for both. The combination remains promising for this narrow sampled-roster question. Its 36.1% payoff when pooling every policy equally does not establish a general overshoot. Neither distribution is a Nash equilibrium, and selection from eight training seeds is uncertain.
The unresolved harm is first-activation sensitivity. On Causeway, control-selected first-side payoff is 50.0% for control, 40.6% for pool minus one, 39.4% for accuracy plus two and 36.9% for both. Added absolute bias is therefore 9.4, 10.6 and 13.1 percentage points, above the predeclared five-point review margin for all three candidates. The combination's early-death increase is only 2.3 points, below its five-point point-estimate margin, and it improves empty proactive Fires. This is a specific unresolved stress-test problem, not a blanket failure of every outcome.
These selected-training observations do not prove failure of the unrun full-15-pair holdout: its opponent weighting differs, and adding ten untrained pairings changes the estimand. They do justify withholding advancement while the turn-order mechanism and policy stability remain unresolved. Joint leave-one-seed policy agreement ranges from 37.5% to 100%; some cells have nonzero pure response gaps. A larger confirmation of the present mixture would not resolve those methodological questions.
A separate scope limit is the catalog. Full games field dual Blade melee rosters; the contact-only audit shows lighter dual Spike builds with more opportunities and higher sampled exchange kill probability. This does not disqualify every narrow six-roster experiment, but it prevents catalog-wide or human strategic-diversity claims.
Next work justified by these results
- Trace Causeway first/second-side outcomes under the shared selected policies, separating exposure, scoring boundaries, activation availability and targeting. Do not change the alternating-turn rule from an aggregate correlation.
- Make bot threat evaluation account for an enemy's Move followed by the actual interactive melee exchange. Validate reachable objectives and safe roof exits before using the controller as a balance instrument.
- Expand the legal roster panel with matched Blade/Spike, shield and two-handed swaps, plus ranged/support alternatives and relevant pilot synergies. Evaluate opportunity, survival, mobility and utility costs separately rather than collapsing them into one uncalibrated scalar.
- Preregister any new candidate screen after these mechanisms are addressed. If one credible candidate survives, retain the existing full-panel fresh-seed confirmation and independent mirror design, with guardrails and policy limitations visible.
This is a future work sequence, not another tuning ladder within the completed study. No automatic combat, CP, AI-default, deployment or GitHub change follows from these results.