Balance study stage selection — 2026-09-09
This record is written after the complete original 61,440-game screen and before the wide-stage games. The eight-pair screening panel is uneven across families, so its reported spread is only an exploratory prioritization signal. Full opponent-panel spread will be measured on all 15 pairings. No candidate is selected for live rules by this record.
Wide-stage conditions
Advance five mechanisms, retaining baseline and four combinations (ten conditions total):
| Single | Reason to retain | Known screen drawback |
|---|---|---|
accuracy1 |
Empty proactive Fires 15.1%→12.0%; exploratory panel spread 35.5%→25.8% | Wipes 70.4%→75.7%; deaths before first attack 8.8%→10.1% |
cp-round-offense |
Empty Fires 10.5%; explicit shared, limited resource with equal replenishment | Wipes 77.3%, deaths before attack 10.7%; no observed comeback improvement |
hull120 |
Survival control: deaths before attack 6.4%, wipes 59.9%; test whether it offsets added offense | Panel spread worsens to 50.6%; not a standalone balance recommendation |
nofreelock |
Controlled test of unpriced free support; panel mean absolute deviation 10.1%→7.8% | Empty Fires 16.8%; worst added first-side bias 5.1 points; not yet acceptable |
goalpath |
Independently confirmed direction error in the AI cost field; test routing separately from combat stats | Supplementary matched screen and baseline repeat required before wider execution; no inferred benefit |
The combinations are accuracy1+nofreelock, accuracy1+hull120, goalpath+accuracy1+nofreelock, and cp-round-offense+hull120. They test whether broad reliability can replace concentrated free support, whether survival absorbs offense's lethality cost, and whether corrected routing changes the combined result. All component singles are controls; the goal-path triple also has its two-stat combination control. Interaction is candidate change minus the sum of component changes, treated descriptively.
The fixed-reserve offensive CP option was competitive in the screen, but the round-income option had a smaller observed first-side harm and exercises replenishment. This is a bounded selection, not proof that its economy is better. Accuracy +2 improved reliability further but had worse exploratory spread and early-death exposure than +1. Low-variance damage and larger Hull/Core increases worsened the panel gap. Heat reduction increased lethality and showed 6.1 points worst added first-side bias. Scoring-only changes barely altered outcomes while elimination remained decisive. Melee damage/accuracy did little to resolve the major map interactions. Range reduction improved early survival but did not materially reduce panel spread. These unadvanced alternatives remain in the complete retained screen.
Fixed execution design
- Original screen: seed 300000–300015, eight pairings, four maps, four assignments, 30 conditions; complete and source-verified.
- Goal supplement: the same screening cells/seeds for baseline and
goalpath, 4,096 games. Compare every repeated baseline observation with the original baseline before interpreting the correction. - Wide stage: seed 400000–400031, all 15 different-family pairings, four maps, both roster ends crossed with both first sides, aggression .75 for both; 7,680 games per condition, 76,800 total. Use separate source snapshots after the pathfinding addition.
- Subsequent policy training and holdout will have a separate selection record. The final comparison remains one locked finalist versus baseline over all 15 pairings and four maps, 256 fresh blocks. Mirrors are a separate diagnostic, not extra opponents in family payoff.
Independent review of the widened design and implementation precedes execution. If a mechanism fails implementation or replay checks, repair it and repeat the relevant validation before freezing sources. Do not discard unfavorable completed runs or change the wide sample size in response to outcomes.
Pre-wide review and parity outcome
The independent reviewer approved the optional goal-field implementation, old-save capability handling, corrected analyzer, locked-policy method and wide-stage selection. The goal supplement completed 4,096 games with matching final source hashes; all 2,048 repeated baseline observation rows exactly equal the original screen. Full TypeScript suite: 498 tests across 74 files pass; typecheck passes. Goal-path effects remain exploratory and are retained as a diagnostic even if adverse. Map diagnostics use separate seed 800000–800063 to keep policy-training seeds unused.
The goal supplement did not solve the Foundry melee failures: all three tested melee matchups remained winless there. Its partial-panel spread increased from 35.5 to 44.5 points, deaths before attack from 8.8% to 11.0%, and worst added first-side bias was 5.5 points. It therefore advances only as a routing diagnostic/control, not as an acceptable balance candidate. The retained supplement analysis was available before wide-stage execution.