Independent final review: multivariant balance study
The completed study is approved as a reviewed diagnosis. No candidate is approved for live adoption or advancement to the planned holdout. The game is not certified balanced.
This verdict follows the completed cautious-policy grid and the final selection record. The combination remains a promising experimental candidate for a narrow, explicitly defined roster comparison. Withholding advancement is a response to unresolved policy/first-activation effects; it is not a claim that every measured outcome worsened or that a full-panel confirmation already failed. No final 600000-range outcome was used.
Evidence and implementation checks
The program retains 280,320 full-game executions, including 6,144 exact control/parity repeats, and a separate 153,600 isolated melee exchanges. Shared seeds link comparisons; these counts are not independent experimental units. Original conditions, source snapshots and adaptive selection records remain available.
Independent review covered clone-only candidates, hard build limits, optional CP lifecycle/dice order, directional movement costs, complete scoring footprints, unreachable-route handling and saved-session capabilities. New optional controller/combat experiments do not change the selected live defaults. Exact baseline/prior-capability history and future-RNG checks protect the controls.
For the completed cautious grid, independently verified all four 16,000-row files against the exact five-pair/four-map/25-policy/eight-seed/four-assignment design and their retained hashes. Recomputed all 80 control/candidate policy matrices and pure-security selections, plus all 640 leave-one-seed selections and their held-out payoffs. The 60-cell profile contains exactly 20 trained cells and 40 explicitly untrained .75/.75 fallbacks. Candidate-adapted policies remain secondary.
The nine generalized analyzer tests pass independently. Earlier focused reviews also verified the policy-diagnostics tests, event-only melee observer, isolated loadout harness, raw-seed aggregates and histograms, and exact ranged-accuracy negative controls. The map-family review independently reproduces all 144 metric changes and 288 bootstrap endpoints from the retained mirrors.
The completed policy appendix also passes final review. Its action counts, exchange-role outcomes, policy stability and matched payoff interaction agree with retained evidence. Independently reproduced the 5,000 whole-seed bootstrap's bias intervals and the equal-map Fire supplement using the same retained resampling design. The appendix correctly separates pooled-attack and equal-map ratios, and labels these fixed-profile 95% intervals as exploratory: they omit policy-selection uncertainty and do not replace the planned 99% held-out guardrails.
What the policy results support
Uniformly averaging policy settings is not an equilibrium. On the same eight fresh seeds, baseline melee payoff is 85.6% at .75/.75, 81.3% over the original three-level policy mixture, and 53.0% over all five levels. These are different policy distributions, not competing estimates of one universal balance value.
At the common control-selected policies, the four conditions instead produce:
| Condition added to safeapproach | Melee payoff against five sampled opponents | Empty proactive Fire | Death before first proactive attack | Wipes |
|---|---|---|---|---|
| None | 75.31% | 13.37% | 15.39% | 67.97% |
| Melee pool −1 | 66.41% | 13.27% | 15.52% | 58.13% |
| Ranged accuracy +2 | 69.06% | 9.06% | 18.05% | 72.81% |
| Both | 53.52% | 8.69% | 17.71% | 66.56% |
Each row uses 640 training observations and eight shared seed blocks. Empty Fire follows the protocol's equal-map average of within-map shot ratios. These are descriptive selected-training results, not adjusted confirmation intervals or a full six-family payoff matrix.
The combination therefore warrants retention as a narrow candidate. Its uniformly pooled 36.1% melee payoff cannot justify calling it intrinsically excessive. Conversely, 53.5% at the selected profile does not establish healthy individual matchups: its melee payoff is 68.75% into mixed and 67.97% into gunline, versus 38.28% into armored and 36.33% into mobile. Interactions must be calculated on matched payoff cells, not by adding changes in nonlinear family spread.
Why no finalist advances now
Causeway exposes a concrete unresolved first-activation effect under the common selected profile:
| Condition | First-side payoff | Added absolute deviation from the control's 50% |
|---|---|---|
| Control | 50.00% | — |
| Melee pool −1 | 40.63% | 9.38 percentage points |
| Ranged accuracy +2 | 39.38% | 10.63 percentage points |
| Both | 36.88% | 13.13 percentage points |
All three point changes exceed the declared five-point review margin in the observed stress-test panel. Eight training seeds and selection uncertainty prevent treating them as confirmed population effects. Joint leave-one-seed policy agreement ranges from 37.5% to 100%, and the pure-security pair need not be an equilibrium.
These observations do not prove failure of the unrun full-15-pair holdout. Its ten additional opponent pairings and explicit fallback policies change the estimand. They do provide a defensible reason to resolve the mechanism before advancing a finalist. Neither more decimal places nor a large simulation count removes that limitation.
The catalog issue is separate. The isolated loadout audit shows that a legal dual-Spike substitute is lighter and retains more opportunities than the dual-Blade roster used in full games. Its sampled one-exchange kill advantage grows under pool reduction. Those controlled exchanges do not prove global dominance, and they do not invalidate a precisely scoped six-roster comparison. They do prevent claiming that such a comparison establishes catalog-wide variety or human strategic balance.
Bounded prerequisites for future confirmation
- Trace matched Causeway first/second-side outcomes under the common policies. Separate exposure, scoring boundaries, activation availability and targeting before proposing a turn-order rule change.
- Address the documented ranged-policy limits: current-position threat omits enemy Move-plus-Melee reach, and the activation-value heuristic does not model the interactive exchange. Preserve movement legality and control parity while testing any correction separately.
- Include matched legal Blade/Spike, shield and two-handed substitutions, relevant pilot synergies, and ranged/support alternatives in a declared roster panel. Evaluate mobility, survival, opportunity use and utility alongside damage.
- Only after those controls support a credible candidate, preregister a bounded new stage and keep fresh full-panel confirmation, mirrors, paired seed analysis and guardrails intact. Do not reopen the completed study as an unlimited tuning search.
The study has successfully identified meaningful controller, map, policy and equipment mechanisms. The accepted deliverable is that evidence and its limits. Permanent offensive Enraged remains selected; experimental CP, combat tuning and routing candidates receive no automatic live promotion, deployment or repository publication from this review.