{
 "schema": "cain.e107.preregistration.v1",
 "written_before_sealed_runs": "2026-11-19",
 "title": "E107: the organism runs its own experiments - interventional attribution of a hidden precondition",
 "disclosed_development": {
  "dev_seeds": "practice grid 994000+ (600), edge 995000+ (120); use-check edge 996000+ (60), grid 997000+ (40); trial-1 sweep 992000+/993000+",
  "findings": [
   "trial 1 (default organism) solves edge 21/30 at both 400 and 1000 steps; door named correctly whenever solved",
   "grid: 164/600 worlds have no recorded opening (key collected before the door was bumped) - a selection that depends on key location; calibration measured on real rows instead of assumed",
   "dev attribution: 380/380 grid and 61/61 edge exact attributions equal the true key (oracle check, evaluation only)",
   "dev calibration: grid pooled gate OFF p=0.22 (1/19 batches of 20 fired); edge ON p=0.001, feature column 0",
   "prior_subgoal alone: edge 22 -> 23/60 (no gain); diagnosis: after the key the organism resumed frontier exploration and never retested the passage (KEY_NOT_CROSSED 8 -> 15)",
   "added retest_when_precondition_done; dev edge 22 -> 26/60 (4 vs 0, p=0.125), fewer steps 22 vs 0; grid unchanged 24/24"
  ]
 },
 "frozen_design": {
  "trial1": {"organism_config": {}, "max_steps": 400, "door": "opened cell with the longest believed-blocked window"},
  "experiment": {"max_steps": 1500, "visit": "one new suspect cell over visited cells, cheapest round trip to the door first", "test": "door after every new cell", "census": "near side mapped with the door closed"},
  "row_used_if": "exactly one candidate and census complete",
  "gate": {"hypothesis_class": "one cell, one column or one row", "statistic": "max z over features with expected count >= 1", "null": "cause uniform over each world's census support", "permutations": 999, "alpha": 0.01, "seed": 107, "min_worlds": 10},
  "sealed_arms": {"control": {}, "treatment": {"prior_subgoal": true, "retest_when_precondition_done": true}},
  "oracle_use": "true key/door/near side read after runs for accuracy reporting only; never given to the learner"
 },
 "sealed_sets": {
  "practice_edge": {"seed0": 13000000, "n": 150, "max_steps_trial1": 400},
  "practice_grid_kd": {"seed0": 13100000, "n": 300, "max_steps_trial1": 400},
  "test_edge_lock": {"seed0": 13200000, "n": 200, "max_steps": 100},
  "test_grid_key_door": {"families": [3, 8], "seed0": 13300000, "n_per_family": 50},
  "test_grid_other": {"families": [0, 1, 2, 4, 5, 6, 7, 9], "seed0": 13400000, "n_per_family": 10}
 },
 "hypotheses": {
  "H1_primary": "test_edge_lock: treatment solves more than control (McNemar p < 0.05, treatment-only > control-only)",
  "H1b": "test_edge_lock: treatment uses fewer steps (sign test p < 0.05)",
  "H2_gate_power": "practice_edge gate usable (p <= 0.01) and its feature is column 0",
  "H3_gate_guardrail": "practice_grid_kd gate NOT usable",
  "H4_attribution": ">= 95% of used practice rows name the true key (oracle check)",
  "H5_no_harm": "no significant loss (McNemar p < 0.05 the wrong way) on test_grid_key_door or test_grid_other"
 },
 "publication_rule": "all hypotheses reported pass or fail; a failed H1 is published as a negative result",
 "team": "Vincent Noriega (founder) and a three-person team"
}
