Corrected-controller balance selection — 2026-09-09

This stage is selected after the four-condition reachability diagnostic and before any new balance runs. It uses safeapproach in every condition. The old playable controller remains unchanged. Its apparent 20.0-point family spread concealed Foundry melee failures. Avoiding unreachable roofs raises melee's uniform-opponent payoff to 83.7% and the full family spread to 51.8 points. This is a bot-controller finding; it does not measure human balance.

Targeted single-mechanism screen

The panel is all 15 different-family pairings, all four repaired maps, both physical ends crossed with both first sides, aggression .75 on both sides, and 16 fresh seed blocks 420000–420015. Each of 13 conditions has 3,840 games, for 49,920 total. The exact manifest is docs/tuning/evidence/2026-09-09-multivariant/corrected-screen-manifest.json. Sources will be frozen after independent implementation/design review.

The corrected control is safeapproach. Each other condition adds one of:

meleepoolminus2 is deliberately excluded: Blade starts at two base dice, so the legal minimum of one makes it identical to minus one on every fielded study weapon. Its catalog implementation can support future Spike comparisons. Shields and two-handed weapon semantics remain intact. Damage reductions can saturate on other profiles; exact effects are covered by candidate tests and retained catalog snapshots.

Follow-up selection rules

This is a fresh exploratory screen, not untouched confirmation. Preserve all results. Advance at most five interpretable combinations only if their singles show a plausible mechanism. Include their component controls on the same fresh seeds. Prefer reductions in the complete family spread with acceptable reliability, early-death and map-bias behavior. Report family payoffs and counter-matchups; a large average improvement from 83.7% melee does not by itself establish strategic diversity.

Use a finite aggression grid to test whether a candidate advantage is merely a fixed-policy choice. Select control-derived policies and apply the same profile to both arms of any primary confirmation. Pilot-neutralization candidates remain diagnostic. Do not erase pilot identity to manufacture acceptance. Do not promote optional bot routing or combat numbers automatically.

The original protocol's full holdout and guardrails still apply to a finalist. If no credible candidate remains, or the controller lacks the choices needed to support the claim, record that before touching holdout data and deliver a reviewed diagnosis instead of a false balance verdict.

Launch review

Independent reviewer approved the fixed panel and clone-only implementation as exploratory. All 13 focused candidate/map tests and TypeScript checks pass before freeze. Approval does not certify ranged counterplay or authorize live melee reductions. A read-only range-management probe is being retained separately; the original .5/.75/1 policy grid excludes cautious negative-closing behavior and needs a prospective amendment before policy experiments. No policy or holdout data have been used.

Prospective cautious-policy and combination amendment

Before any policy outcomes, the independent reviewer approved a bounded replacement for a separate corrected 32-block combinations stage. Use all five melee-involving roster pairings on all four maps, eight fresh seeds 500000–500007, all 25 pairs of aggression [0,.25,.5,.75,1], and all four roster/first-side assignments. This is 16,000 games per condition. Compare corrected control, at most two distinct eligible singles selected after the complete 13-condition screen, and their compatible combination: at most four conditions and 64,000 executions. Declare the actual IDs before running. If the singles lack credible complementary effects, omit the combination prospectively. Do not stack redundant intensity doses or incompatible CP economies.

The wider aggression range is necessary because the read-only Foreman/Reaver probe shows that .5/.75/1 omits available cautious range management. This is focused cautious-counterplay sensitivity, not global strategy training. Eight independent blocks support more stable policy selection than spreading four blocks over all 15 pairs, though selection remains uncertain and finite.

Learn pure maximin/minimax policies from safeapproach control only. Retain the full 5×5 payoff matrices and response-gap diagnostics. Apply those same locked policies to both primary holdout conditions. Explicitly populate each of the other 40 map/non-melee-pair cells with .75/.75 and identify these as untrained fallbacks; do not silently fill missing training cells. The final all-15 comparison would estimate this declared mixed profile, not an equilibrium. Candidate-adapted policy results are secondary only.

With both singles present, describe interaction on matched payoff cells as combination minus single A minus single B plus control. Do not call the nonlinear difference of spreads an additive mechanism interaction. This design replaces, rather than automatically adds to, a separate corrected combination screen. Any finalist still needs the original full-15/four-map/256-block holdout and separate mirrors. A failed or methodologically inadequate candidate does not proceed to a ceremonial holdout.

Selected components after the complete corrected screen

All 13 conditions finished (49,920 executions); final source hashes match. All results and output hashes are retained in corrected-screen-analysis.{json,md} and its comparison manifest. No cautious-policy or holdout outcomes have been used.

Select meleepoolminus1 and accuracy2, plus their combination, always with safeapproach. The exact four raw IDs are safeapproach, safeapproach+meleepoolminus1, safeapproach+accuracy2, and safeapproach+meleepoolminus1+accuracy2.

Pool minus one reduces the full family spread from 54.8 to 37.5 points (paired change −17.3, exploratory 95% interval [−22.2,−12.8]). It lowers melee's payoff from 84.8% to 70.5%, but gunline remains 33.0%. Ranged accuracy plus two reduces spread to 38.3, raises gunline to 39.9% and reduces empty proactive Fires from 15.5% to 9.7%. Its early-death rate rises from 6.5% to 7.9% and wipes from 84.5% to 87.0%. The two components address different resource and reliability problems; their combined effect remains unmeasured.

Accuracy plus two has unresolved first-activation uncertainty: worst-map added absolute first-side bias is +2.4 points, but its 95% interval reaches +8.2. This is not a passed guardrail. Pool minus one also remains a play-feel tradeoff because it removes both strike and parry resources. Neither single passes the intended broad family band; neither is accepted for live use.

Damage minus one/two barely alter spread (−0.5/−0.2 points). Ability-only removal and free-Lock deletion do not improve it. Neutral pilots are diagnostic and erase identity. Hull ×1.20 and defensive CP reduce early deaths but worsen the family spread to 64.1 and 60.9. Offensive CP reduces empty Fires to 11.2% and gives descriptive first-loss wins of 38.8%, but its spread of 47.6 is less promising than the chosen reliability component. These results remain in the report; no unfavorable condition is discarded.

Before the next source freeze, add observational melee-exchange counters so the pool tradeoff can be evaluated directly. This adds no engine rule or RNG call. Historical rows remain unchanged and missing exchange counters are not treated as zero. The predeclared primary empty-Fire guardrail stays unchanged; exchange damage and ineffective exchanges are additional descriptive checks.

Cautious-policy launch review

Independent review approved the selected four conditions and pure observational melee counters. All 18 focused tests pass independently and TypeScript checks pass. The new helper is included in the runner source snapshot. The exact 64,000-execution manifest is approved for a new freeze; no live rule or final confirmation is approved by this step.