{
  "schema": 1,
  "id": "feint-interactions-v1",
  "status": "Specified before measurement; local prospective diagnostic, not public preregistration. No mechanism candidate is selected using the reserved synthetic outcomes.",
  "question": "Which authored tactic interactions receive ambiguous individual credit, and which learned role-preference changes account for the original-map feint decline under the exact deterministic outcome table?",
  "seed": 42,
  "styles": ["orbit", "reverse", "feint"],
  "worlds": ["original", "south-entry", "wide-crossings", "central-obstacles"],
  "profiles": "All eight joint authored tactic combinations; one actual combat per deterministic world/style/profile. Repeating a fixed profile across policy seeds would not add independent observations.",
  "expected_encounters": 96,
  "reference": "docs/experiments/adaptation-v1.results.json",
  "late_rounds": [17, 24],
  "analysis": "For each role, compare its two actions with other roles fixed. Retain team damage and individual reward changes. On original-map feints, compute expected damage from the product of each role's actual pre-choice probabilities. Attribute the online-minus-frozen expected difference by averaging marginal replacements over all six role orders (Shapley decomposition). Average rounds within seed, then the 32 historical seeds.",
  "limits": [
    "Outcome-table maxima are offline diagnostic upper bounds, not fair deployable baselines.",
    "The decomposition accounts for changed probabilities within this outcome table. It does not identify the causal learning mechanism that changed them.",
    "The independent-role probability product models sampling expectations, not the realized RNG sequence in each encounter.",
    "No player labels or counterfactual outcomes enter live learning. Core, rewards, stats, observation rules and shipped gameplay remain unchanged."
  ]
}