phase: 4S
name: Preregistered CS-US timing repair and open-loop credit validation
experiment_root: /data/digital-drosophila/experiments/phase4s-timing-credit
model_root: /data/digital-drosophila/models/male-cns-lif/timing-credit-v1
runtime_root: /data/digital-drosophila/models/male-cns-lif/deterministic-runtime-v2
phase4p_root: /data/digital-drosophila/experiments/phase4p-deterministic-runtime
phase4q_root: /data/digital-drosophila/experiments/phase4q-credit-v3
phase4q_model: /data/digital-drosophila/models/male-cns-lif/credit-v3
phase4r_root: /data/digital-drosophila/experiments/phase4r-timing-audit
phase4r_model: /data/digital-drosophila/models/male-cns-lif/timing-audit-v1
duration_ms: 120
candidate_US_ms: [10, 15, 20, 25, 30, 40, 60, 90]
windows_ms: [5, 10, 20, 50, 100, 250]
tau: window/3
calibration_seeds: [54101, 54102, 54103, 54104, 54105]
heldout_seeds: [55101, 55102, 55103, 55104, 55105]
odor_seed_offset: 100000
target_compartment: "y5B'2a"
matched_DAN: PAM15
timing_gate:
  mean_matched_random_ratio_each_odor: 3.0
  per_seed_ratio_at_least: 3.0
  qualifying_seeds_each_odor: 3
  selection: shortest qualifying window, then earliest qualifying US; formal odors and heldout seeds forbidden for selection
  primary_random: exact uniform expectation on all 1200 timesteps, including activity peak
  historical_random_30_110: diagnostic only, never used for selection
formal:
  train_seed_bases: [61000, 62000, 63000, 64000, 65000]
  evaluation_seeds: [71001, 71002, 71003, 71004, 71005]
  checkpoints: [0, 1, 2, 5, 10, 20, 30, 40, 50, 60, 75, 100]
  isolated_checkpoints: [10, 20, 30, 40, 50, 75, 100]
  primary_endpoint: 100
  random_seed_offset: 975000
  fixed_groups: {A-specific: 105, shared: 369, B-specific: 221, inactive: 13537}
  source_events: unchanged base brain, one fresh independent seed per pairing; all eligibility arms share exactly the same events
  shadow_update: E=max(0.75,E*(1-0.01*eligibility)); never fed back during event collection
  arms: [v2_same_US, v2_90ms, v3_matched, v3_random, D0]
  minimum_A_credit_matched_random_ratio: 3.0
  minimum_B_reduction_vs_v2_same_US: 0.25
  minimum_seed_majority: 3
  isolated_gate: at predeclared 100 pairings, at least 3/5 seeds show positive A spike-count L1 greater than paired B; matched effect must exceed random A effect in at least 3/5, and random must not meet specificity majority
  no_early_stopping: true
  whole_brain: only if both formal credit and isolated specificity gates pass; evaluate fixed 100-pairing endpoint
prohibitions:
  modify_frozen_brain_or_eligibility_v2: true
  change_learning_magnitude_or_floor: true
  change_odor_rates: true
  expand_or_reclassify_plastic_mask: true
  online_conditioning_or_memory_claim: true
  choose_parameters_on_formal_or_heldout_results: true
  game_RL_decoder: true