Cautious-policy diagnostics — 9 September 2026

The combination remains promising for the six fixed roster families, but there is no eligible finalist for holdout now. Under policies selected from the corrected control, the melee family's expected score falls from 75.3% to 53.5%, and empty proactive Fires decline. However, all three candidate conditions show a Causeway first-activation bias increase above the predeclared five-point margin. Eight shared seeds leave substantial uncertainty, and policy selection itself is unstable in several cells. This is an unresolved stress-test concern, not a failed full-15-pair confirmation or proof that the combination cannot work.

No holdout games, additional condition selection, live tuning or controller changes were performed for this analysis. The fresh seed-600000 holdout remains untouched.

What was measured and locked

All four approved conditions completed 16,000 games each: 64,000 total. Each includes four maps × five melee-involving roster pairings × all 25 aggression combinations [0,.25,.5,.75,1] × eight seeds 500000–500007 × four physical-end/first-side assignments. The actual conditions are safeapproach, safeapproach+meleepoolminus1, safeapproach+accuracy2, and their combination safeapproach+meleepoolminus1+accuracy2.

The reviewed selector learns pure maximin/minimax choices from control only, with its fixed tie preference nearest .75, then lower aggression. Its common selected profile evaluates 640 games per condition: 20 map/pair cells × eight seeds × four assignments. Every candidate uses the same control-selected policies in the primary descriptive comparisons below. This reuses training outcomes; it is not independent confirmation.

The exported locked profile explicitly includes all 60 map/pair cells. Twenty are trained here; the other 40 are labeled untrained .75/.75 fallbacks. Those 40 have no outcomes in this policy stage. A future full comparison would measure that declared mixed profile, not an equilibrium or comprehensive strategy training. A complete six-family spread cannot be estimated from the five-pair panel alone.

Cautious choices are active and consequential

Control selects non-melee aggression 0 in 11 cells, .25 in seven, and .5 in two. Melee aggression is 1 in 14 cells and .75 in six. The wider policy range therefore changes the observed comparison materially: on these same seeds, the control melee payoff is 85.6% at .75/.75, versus 75.3% at the selected profile. The combination is 66.1% at .75/.75 and 53.5% under the common selected profile. Neither uniform pooling over all 25 policies nor the old .75-only result substitutes for the selected-policy estimate.

Holding the melee team's aggression at 1, the following descriptive rows each contain 640 games per condition. Action counts combine both teams; they do not identify which team moved or fired.

No game in these slices had both zero proactive Fires and zero melee exchanges. This supports observable cautious counterplay within the available bot action set. It does not establish optimized ranged play: the documented missing enemy Move+Melee threat horizon and legacy action valuation remain.

Selection stability and security gaps

Whole-seed leave-one-out fits remove all four assignments of one seed from every policy cell. Each omitted fit trains on seven seeds/28 games per aggression pair. The selected joint policy remains unchanged in 81.9% of fits averaged across cells; eight of 20 cells retain it in all eight omissions. Two retain it in only three of eight: Quarry mobile versus melee and Foundry mixed versus melee. These are descriptive stability counts, not confidence levels.

Six selected cells have a positive observed pure-response gap. The largest is 25 points for Mobile on Quarry: a different row policy has a 25-point higher observed payoff against the selected opposing policy. The independently secured choices need not be a Nash equilibrium. All 5×5 payoff matrices, both response gaps and every omitted-seed choice are retained.

Retraining the selector separately on each candidate changes 12 of 20 selected policies in each case. Its secondary melee payoffs are 62.2% for pool−1, 65.9% for accuracy2 and 51.8% for the combination. These candidate-adapted results change policy alongside rules and are not the common-policy treatment contrasts used below.

Common-policy outcomes and the holdout blocker

¹ The primary-compatible Fire statistic equally weights four within-map empty-Fire/Fire ratios. It uses attacks as denominators within maps, unlike game-payoff rates. The separately retained pooled-attack values are 13.32%, 13.20%, 9.01%, 8.66%; those are descriptive and differently weighted. ² Deaths before any recorded proactive attack, divided by all six starting units per game. First attack includes proactive Fire, melee initiation and qualifying grapple actions; this is not death before any activation.

The combination lowers melee payoff by 21.80 points versus control. Pool−1 alone lowers it by 8.91 and accuracy2 by 6.25. On matched cells, the melee-oriented combination interaction is −6.64 points, calculated as combination − pool−1 − accuracy2 + control, with the payoff direction consistently oriented toward the melee family. It is a conditional payoff interaction, not a nonlinear spread comparison or an established synergy. Across eight shared seed clusters its value ranges from −22.5 to +10.0 points.

The central eligibility concern is on Causeway, where first-side expected score changes from .500 to .40625, .39375 and .36875. The corresponding added absolute deviations from 50% are 9.375, 10.625 and 13.125 points; each exceeds the predeclared five-point harm margin. The direction favors the team activating second in this panel. All three conditions show this observed concern, so switching to a single does not supply a clearly acceptable fallback.

A fixed exploratory paired whole-seed bootstrap produced the following descriptive 95% intervals. It used 5,000 replicates with seed509901, sharing each seed weight across all maps, pairs, assignments and conditions. The profile stays fixed during resampling; selection uncertainty is not incorporated. With only eight independent blocks and no multiplicity adjustment, these are not the predeclared 99% held-out acceptance intervals.

The worst-map statistic is recalculated in every resample; it is not simply the interval for Causeway alone. Its uncertainty does not establish population-level harm, but it also does not rule out the stated margin. Physical-side uncertainty remains unresolved despite smaller point changes. Empty-Fire improvements therefore do not settle overall acceptability.

The decision is no finalist now, while retaining the combination as a promising narrow candidate. No full-six-family spread claim or confirmation failure is inferred from these five pairings. A future holdout would require a credible resolution of this stress concern under a prospectively documented decision, rather than treating the existing holdout as a ceremonial step toward promotion.

Whole melee exchanges expose the play-feel cost

The new observer measures matching completed melee_start→melee_end spans using actual accepted-strike damage events. It does not count each strike as another exchange or use final game hull to reconstruct damage. Both roles below combine whichever family actually initiated or defended; they are not synonymous with melee and ranged rosters.

Every exchange here completed and its empty/non-empty status was observed. Damage totals use actual durability removed, including critical effects, rather than rolled damage or overkill. Counter-strikes, parries, failed upfront opportunities and credited kills remain separately retained. Defender “empty” can include useful parrying; it is not automatically a failed defensive experience.

Pool−1 reduces attacks and parry resources. More exchanges per game and lower damage per initiation accompany a higher chance of an initiation doing no body damage. The combination improves the Fire-only reliability statistic while raising initiator-empty melee exchanges from 21.4% to 33.8%. These attack-weighted mixtures include changed contact timing, wounds and eligible Enraged states; they are not fixed-weapon probabilities or causal isolation of a single die.

The separate fixed-position loadout audit adds a further scope limit: the game study's dual-Blade family does not cover cheaper dual Spikes, which retain two extra opportunities under pool−1. That risk does not automatically disqualify a finite-roster result. It prevents extending such a result to catalog-wide balance, and warrants explicit kit-budget coverage in any later broader claim. The present holdout decision rests on the observed first-side stress concern and uncertain policy selection.

Retained evidence and reproduction

• Full selector diagnostics: exact input hashes/IDs, matrices, selection, leave-one-out fits, conditional contrasts, full policy-cell interactions and candidate-adapted profiles.
• Locked 60-cell control profile.
• Observed action/exchange summaries and reproduction script.
• Conditional uncertainty and eight-seed sufficient statistics and script. Its zeroFire field is the explicitly pooled-attack statistic.
• Primary-compatible equal-map Fire supplement and script. It reuses the exact same 5,000 bootstrap weights, not additional replicates or simulations; the earlier pooled artifact remains unchanged.
• Approved 64,000-game manifest and execution log.

Exact analysis commands are retained. The scripts read only this completed policy stage and refuse output overwrites. Reproduction requires Python and NumPy. Copy the evidence directory and omit only its derived outputs to regenerate summaries; retain all four raw condition files, the manifest and source records. No holdout file is an input.