Multivariant balance study protocol — 2026-09-09

The owner requested broad testing, strategic variety and a separate review. All maps must be symmetric; terrain, deployment and nonoverlapping objective repairs are part of the new baseline. Original maps are archived for a separate paired diagnostic. Permanent offensive Enraged is the baseline. Candidate catalogs, maps and rule overrides are isolated; nothing is promoted automatically. Research uses official competitive-game sources. TASK-090 tracks execution.

Questions and criteria frozen before screening

Which observed swings come from physical board end, first activation, roster, policy or their interactions?
Can a candidate reduce the spread of per-family expected score against an equally weighted opponent panel by at least 5 percentage points without erasing counterplay? A complete symmetric matrix averages 50% by construction; its grand average is not evidence of balance.
Can reliability, opportunity to act or comeback improve without a harmful tradeoff? Secondary practical targets: at least 3 points better empty-Fire/comeback rates, or half an activation earlier engagement.
Command reserves must show actual use, gain, cap loss and opportunity costs. Compare fixed reserves, equal replenishment at scoring checkpoints, and casualty income. Automatic spend policies are experimental proxies for player choice, not a finished human mechanic.

Guardrails: rule out more than 2 points increased zero-body-damage proactive Fire probability, 5 points increased deaths before a first proactive attack, or 5 points added physical-side/first-activation advantage on any map. The latter two statistics are the maximum across maps of candidate minus baseline absolute deviation from 50%; signed and mean absolute map biases remain descriptive. A new family confidently outside a 35–65% uniform-opponent expected-score band is a warning. Existing failures remain visible. These are study design targets, not universal standards from other games. Inconclusive intervals do not pass a guardrail.

Experimental units and controls

Use six legal roster families: mixed, armored, mobile shooting, melee, gunline and control. Report actual costs/cap utilization and observed action mix. Roster labels express intent; bot behavior must substantiate them. Include individual hard limits and distinct pilots. Do not silently repair an illegal candidate roster. Preserve pilot, map and policy confounds in the interpretation; no claim of exhaustive kit-builder search.

For different rosters, each seed block includes roster X at physical A/B crossed with firstSide A/B: four executions. Same-roster mirrors need two firstSide executions. Policies travel with their roster. The firstSide override removes initiative-roll randomness for this factorial control. Separate ordinary random-initiative confirmation may be added and labeled.

Keep map-specific results. Reuse seed blocks within a phase for paired comparisons, but use disjoint seed ranges across phases. Resample entire seeds, retaining all linked assignments/maps/policies in a pooled statistic. Draws score half for strategy payoffs and remain separately reported. Movement-changing candidates need not share first-death histories. Baseline repeats and disabled optional mechanics must preserve exact events/RNG.

Fixed map baseline

The study uses the repaired 32×20 maps. Each rotates 180 degrees in the odd-r hex graph: cells, deployment, and complete objective flowers map one to one. Causeway and Foundry retain their outer objectives and replace the impossible singleton center with the non-overlapping pair [15,7] and [16,12]. Quarry now reflects its original left-side terrain into the opposite half and uses the same paired central objectives, plus [10,6] and its rotated mate [21,13]. Every scoring map therefore has four zones. This adds one board-wide scoring opportunity relative to the archived three-zone maps; it is fixed before the study and requires no scoring-rule change. Presence-based control and three deployed mechs still limit each team to three zones per checkpoint. Over seven normal rounds, the theoretical team ceiling remains 21 zone points, while the combined board ceiling rises from 21 to 28, ignoring travel and contestation. Pre-repair JSON is retained in docs/tuning/evidence/2026-09-09-multivariant/pre-symmetry/ for root-cause controls.

Stages

Mechanism and baseline audit: validate candidate effects, roster legality, exact replay, mirrors, physical side/initiative separation, objective symmetry and bot limits. Freeze source hashes before batches.
Individual screen: approximately 20–30 isolated candidates, eight diagnostic pairings covering all families, four maps, 16 seed blocks/cell. Screening is exploratory. Advance at most five based on declared targets and guardrails, not an arbitrary best win rate.
Wider/combinations screen: all 15 family pairings plus mirrors on 32 fresh seed blocks/cell for baseline and compatible finalists. Limit combinations to a few interpretable pairs/triples. Include each component-alone control and report interaction relative to additive changes. Lock one final combination/single before holdout.
Policy stress: a 3×3 aggression grid for baseline/finalists on declared contested matchups/maps; select a baseline-derived policy pair separately for each matchup/map and use that same locked pair in both baseline and finalist on fresh seeds. Independently adapted candidate policies may be a separately labeled secondary analysis. Aggression is one coupled heuristic knob, so this certifies only a finite bot-policy response set. Distinct command spending policies are part of this stress test.
Holdout: one finalist and baseline, all 15 different-family pairings and all four maps, exactly 256 fresh seed blocks × four roster/first-side assignments: 61,440 games per condition, 122,880 total. Run six same-family mirrors separately with both first sides. No selection of only favorable confirmation cells. Five primary claims are spread improvement and the four guardrails above. Use 5,000 paired whole-seed bootstrap replicates and 99% percentile intervals for Bonferroni-adjusted nominal 95% family coverage, subject to bootstrap approximation (especially for max/min and absolute-value statistics). The upper interval bound on candidate-minus-baseline spread must be ≤−0.05; each guardrail upper bound must be below its harm margin. Mean absolute family deviation is secondary. Do not extend sampling based on observed results; inconclusive precision stays inconclusive.

Exact candidate IDs, diagnostic pairings, seeds and selected confirmation cells must be written before each relevant run. Preserve selection records before touching holdout data. Simulations may run in independent processes; do not mutate source while a frozen batch is in progress.

Acceptance and delivery

A candidate must improve a practical target and rule out the harmful guardrail margins on holdout. Report tradeoffs, counter-strategies and maps where it fails. Existing map/bot bias prevents a universal balance verdict. A useful result may be a root-cause diagnosis with no accepted replacement.

Deliver official-source research, baseline and candidate artifacts, reproducible commands, plots, an explanation of major swing causes, candidate recommendations and a separate reviewer verdict. Keep current rules documents aligned with the selected live rule; label study mechanics as experiments. Human play remains necessary to evaluate agency, bluffing, resource choices and strategy diversity outside this bot family.

Pathfinding diagnostic amendment (before supplementary candidate runs)

The independent reviewer confirmed that the AI reads a distance-from-goals field as cost-to-goals although ascent, descent and terrain-entry costs are directional. See 2026-09-09-goal-pathing-diagnostic.md. This is a demonstrated heuristic error, not yet a measured balance explanation. Complete the original frozen screen unchanged, then test an optional reversed-edge goal field with an unchanged-baseline replay check under a new source freeze. Include the correction alone and with selected balance candidates. Do not compensate a bot routing defect with live melee stat buffs.

Reachability amendment (before geometry-control runs)

Further read-only traces demonstrated two remaining planning defects: unreachable routes fall back to geometric distance, and objective-center goals can be blocked even when their scoring footprints are accessible. See 2026-09-09-foundry-roof-trap.md and independent 2026-09-09-safe-approach-review.md. All six baseline/goalpath × aggression .5/.75/1 traces at the specified Foundry seed trap the three healthy melee units on roofs and score no melee-side zone points. The original wide stage finishes under its existing freeze; do not use those trapped matchups to certify melee's human viability.

Four mechanism controls are baseline, goalpath, zonegoals (directed costs plus complete objective footprints), and safeapproach (the same plus a finite penalty preserving unreachable routes). Freeze all 15 pairs, all four maps and all 32 wide seed blocks 400000–400031 for the latter two: 15,360 new exploratory games. Reuse the complete immutable baseline/goalpath wide controls only through an explicit comparison manifest retaining original phases and hashes. Before that reuse, prove structural default/prior-goal no-ops and repeat all 2,048 baseline plus 2,048 goalpath observations from the supplemental screen under the new executable. Report this as subset parity; do not imply all wide rows were rerun. This is adaptive exploratory evidence, not independent confirmation.

If reachability improves, evaluate a bounded stat/resource shortlist using the same corrected controller in control and candidates on fresh seeds before policy selection. The playable default and historical baseline remain separately labeled. No data from the exploratory/reused stages count as untouched holdout. The final confirmation design and guardrails above remain fixed unless a methodological blocker makes a balance verdict inappropriate; document such a decision before accessing any final data.

Cautious-counterplay amendment (before policy outcomes)

The corrected controller exposes melee dominance under .75 aggression, but a read-only open-board probe shows that the original .5/.75/1 stress grid excludes the bot's available cautious range-management behavior. The independent reviewer approved expanding the grid to [0,.25,.5,.75,1] on all five melee-involving pairings/all four maps/eight fresh seeds, with all four assignments. Corrected control, two distinct selected singles and their compatible combination make at most four conditions/64,000 games. This replaces the separate corrected combination screen; it does not change the original historical wide stage. Exact IDs will be selected after the complete 13-condition corrected screen and before any policy results. See 2026-09-09-corrected-controller-selection.md and its separate review.

Control-derived pure-security policies cover 20 map/pair cells; the other 40 cells will explicitly retain .75/.75. The resulting full holdout measures that declared policy mixture, not equilibrium or all strategic counterplay. The original full confirmation sample, five primary claims, guardrails and independent final review remain unchanged. Untouched holdout remains conditional on a credible candidate and valid scope.