{
  "schema": "cadence-attention-content-control/1",
  "scope": "Supplied unselected content through existing recall source ports; no learned selection, semantic representation or continuing-life claim.",
  "development_seeds": [0, 1],
  "confirmation_seeds": [300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 319],
  "confirmation_status": "reserved and unopened; this runner implements development only",
  "rounds": 512,
  "neutral_events": 1,
  "training_seconds_per_arm": 90,
  "total_seconds": 600,
  "arms": ["recall_ports", "sensory_exposed"],
  "brain": {"modules": [32], "episodic": false, "teaching": "normalized", "learning_qualified": true, "learning_damping": 0, "learning_free_steps": 1024, "learning_nudged_steps": 1024, "recall": true, "recall_next": false, "recall_address": 0.0, "recall_amplitude": 0.0, "recall_capacity": 4},
  "dt": 0.1,
  "content_amplitude": 3.0,
  "acquisition_floor": 0.95,
  "erasure_loss_minimum": 0.20,
  "donor_follow_minimum": 0.80,
  "relevant_follow_minimum": 0.80,
  "irrelevant_change_maximum": 0.05,
  "sham_change_maximum": 0.05,
  "data": "The same sixteen selective_recall cases and actual two-fact/neutral prefixes. Both witnessed fact events are summed, never selected by the answer. The query observation names A or B and carries neither fact. All cases occur in training; no novel binding claim.",
  "boundary": "After the actual prefix, clear working trace and neural warm state explicitly while retaining witnessed slots. Recall-port control supplies both unselected fact events together on the existing one-way recall source neurons. The matched sensory control supplies the same four fact coordinates on the original exposed sensory coordinates. Automatic recall amplitude is zero; no successor region, no selected slot oracle, no new module or learned answer head.",
  "teaching": "Capture the complete supplied-content drive before the query act, qualify from rest, act greedily, and teach only wrong rows from those saved drive rows using the existing local learner. No-wrong rounds take no lesson. All accepted free/positive/negative phases must qualify. Trial resets preserve the acquired parameters.",
  "evaluation": "Final acquired state only, private forks, all sixteen cases. Keep query, zero trace and rest warm boundary fixed while erasing content, flipping relevant/irrelevant content, complementing both facts from a donor, or making an identical sham copy. Record every refused/unrun answer as -1. Initial frozen readings and independently seeded uniform random are controls. Source/checkpoint fingerprints verify isolation.",
  "gate": "Every declared development arm/founder must complete training and every evaluation condition without answer refusals. Intact overall, unequal-fact and each-query accuracy must reach 0.95. Erasure loses at least 0.20, relevant flips and donor content are followed at least 0.80, irrelevant/sham changes at most 0.05. This admits decoding supplied content only, not an attention candidate.",
  "budget": "One fixed recipe, no parameter search. One local run at a time; development shares 600 host seconds, at most 90 per training arm. Deadlines checked between bounded trials; bounded initial/final evaluation follows and is recorded. Confirmation remains unopened. Host time and solver counters are not physical energy. Preserve failures."
}
