THE FIELD NOTES · CRITTERFRONT
Good evidence has boundaries.
What 90,118 scheduled battles can tell us—and what they cannot.
Real combat, a limited question.
The experiment ran 90,118 scheduled battles in the actual deterministic Quantum combat simulation. It compared already-owned legal rosters across gold allocations and round rules. Two living players fought with both attacker and defender forms simultaneously; two remaining seats were empty and eliminated. Classic abilities were enabled and match modifiers were disabled.
Seeded roster generation, evolutionary counter-search and unseen-opponent validation searched for unusual strength. This was not reinforcement-learning training, a player survey, or a complete match economy simulation.
Points are not win rates.
A win earns one point, a draw half a point, and a loss zero. Outcomes compare applied opposing Nexus damage, capped at six per round; damage statistics do not break ties. A roster winning half its battles and drawing the rest would score 75% points—not a 75% win rate.
Each opposing composition has equal weight. The 95% intervals resample opposing compositions, not individual repeated seeds or seat swaps. They describe the generated panel, are not corrected for all candidate comparisons, and do not estimate a live ranked population.
Gold means net roster investment.
Stars I / II / III / IV consume 1 / 2 / 4 / 8 base copies. Each binary merge refunds one gold, so net investment is purchase total − (base copies − 1) per recruit. For example, a two-gold star-III recruit needs eight gold in purchases and produces three gold in merge refunds: five net gold. Two such recruits therefore cost ten net gold but require sixteen gold of purchase transactions.
Round capacity, tier availability and the twelve-copy shared family pool were enforced for both armies together. But shop rolls, reroll/search costs, interest, purchase liquidity and the other players’ pool holdings were not modeled. A legal round-seven roster does not imply it can reliably be acquired by round seven.
Not every guide has the same evidence.
- Six-gold pairs: the complete panel contained 50 legal round-one compositions and 14,700 battles. Guide scores use only the nine other fully invested opponents: 108 battles each across two seeds, three formation policies and paired seats.
- Main validation: 36 finalists were frozen after screening and counter-search. Each faced 48 physically unseen opposing compositions with at least 85% of budget spent, three seeds, three formations and both seats: 864 battles per finalist, 31,104 total.
- Counter notes: small, deliberately selected direct matchups answer a narrower question. Six wins against one army do not establish six independent counter-strategies or universal dominance.
Formations, opponents and caps matter.
Three policies tested planned defense with front-row offense, randomized legal defense and Nexus placement, and planned defense with randomized full-depth offense. None is a complete search over all possible placements. Planner links in these guides are illustrative, not extracted winning placements.
Shared-pool legality can give finalists different opponents. At 60 gold, only 22 opponents were common to all four finalists; the leading three were not clearly separated by the descriptive paired intervals. Do not compare scores from different budgets or opponent panels as if they were a universal tier list.
Actual combatant caps and denied allocations remain in the results. Recursive Bunnie armies can hit the 384-body ceiling; a cap is neither automatically a loss nor proof of excessive strength. Four-player contention and rendered frame rate were not measured.
What changed after the experiment?
Hermit King’s defender reflection changed from 10%/25%/40%/75% to 10%/20%/30%/50%, and both forms lost 5% health. Its separate attacker fifth-shot reflection is unchanged. The website’s creature reference and planner use the new values; all experiment scores here retain the old snapshot. Even a roster without Hermit can face different opposition after that change.
No fresh post-nerf balance panel has been run for these guides. The playable browser alpha has its own build date. Treat these pages as documented team ideas and historical evidence, not a current-patch ranking.
Inspectable evidence
Download the compact evidence snapshot. Each guide includes its run, stage, roster ID, sample size, formation results and source hashes. The development repository retains the full plans, raw battles and analysis; the website build uses this small frozen extract without shipping the full experiment.