Validated adaptation opportunities

On prospectively qualified tasks where the best legal response changes, how much does learning reduce regret and how quickly does it recover?

18432 admission-table encounters and 27648 learning/control encounters; 46080 new reserved executions. All candidates remain in the audit, including rejects.

Six comparisons: each solo role and the joint squad online-minus-exact-switch-frozen mean damage over all 32 B encounters; each terrain-context host minus pooled context on 32 frozen alternating test encounters. Damage difference equals reduction of mean realized oracle regret. Require >=12 qualified seeds, >=3 HP mean gain and adjusted lower interval bound >0. Context also requires >=3 HP over the shuffled-label capacity control with a positive secondary 95% lower bound. No production promotion.

Solo and squad: train A for 32, branch B for 32 into online, frozen, uniform, and cell-reset online; return A for 32 from the B-online checkpoint into online, frozen and cell-reset; sham A for 32 from the original A checkpoint into online/frozen. Squad uses a real joint eight-action policy with shared reward; an additional independent-role host receives the identical reward, with its own 32 A training and 32 B online/frozen branches. Context: pooled/terrain/shuffled hosts each train for 32 alternating A/B encounters and test frozen for 32 alternating B/A encounters. Only map geometry changes between context A and B.

Recovery and regret

First ending encounter of three consecutive PRE-CHOICE distributions whose exact expected regret is <=35% of that state’s best-minus-worst action range. Report initially displaced cases separately; initially acceptable policies are not evidence of recovery. Unreached cases are right-censored after 32; restricted-check summaries encode them as 33 and never discard them. No mean over successes alone.

Offline best legal action damage minus actual damage; cumulative regret sums over encounters. Expected regret integrates the actual joint-choice probabilities against full conditional outcome tables (products for independent roles). Tables never enter choose/feedback. This oracle is an optimistic reference, not a deployable agent.

The summary JSON contains each branch’s cumulative regret, initial displacement, censored recovery records, probability-integrated regret, clear/death rates and four-encounter curves. These secondary measures are exploratory.

Admission and scope

Enumerate every candidate and every legal action in both states before running any learner. Keep all outcomes. Select the first qualifying candidate per group/seed by index. Never backfill a seed with no qualifying candidate. No learner performance enters selection.

This is a deliberately action-sensitive challenge population, not a random sample of games or players. Report both candidate-level and seed-level qualification. The reserve contains new parameter/world seeds for familiar controller families, not unseen behavior families.

• Development enumerated 4,608 fights and informed the generator and reward eligibility criterion; development outcomes are not reserved evidence.
• Authored controllers use delayed restricted public observations. Slower firing schedules intentionally allow differences between tactics to emerge; player weapon capabilities and all enemy stats stay unchanged.
• Individual means isolated enemy, not player-identity recognition. One tactic is chosen per encounter, not within-fight replanning.
• The joint/independent sampling distributions differ despite matched scalar exploration and temperature. This comparison does not isolate factorization from exploration.
• Seed bootstrap coverage is approximate at n<=16. Cases and repeated rounds are not independent population draws. Secondary controls, return results, curves, recovery and expected regret are exploratory.
• A/B tasks repeat exactly conditional on actions; new learning experiences differ through stochastic choices and feedback, not an independently resampled physical world each round.
• Damage opportunity does not guarantee practical learnability within 32 encounters; reward saturation, noisy sampling and memory interference may still matter.
• Rejects and missing qualification are part of the result. No thresholds, sample sizes, labels or candidates change after freezing.
• Public terrain is the only contextual cue. No strategy labels, future actions, oracle choices or rewards are supplied to the learner.
• Previous measured sources, results and the production game remain unchanged.