{
  "objective": "Predict whether the sealed target separation reduction (before - after) is strictly greater than its threshold. Each apparatus has one stable hidden A/B coefficient in the stated public dynamics family. Apparatus IDs are neutral matching identifiers; different apparatuses need not have the same coefficient. Target geometry can change the event outcome. The target cannot be executed or inspected before your prediction. Choose your investigations and arguments yourself. You may search and retrieve memory, execute a legal calibration, estimate from authentic endpoints, integrate explicit public dynamics, or finish. Evidence acquired in this task is accessible. Previous observations are accessible only through permitted memory search and retrieval. Controller assistance: calibration endpoints are automatically stored with apparatus, experiment and task provenance; current-task observations and requested tool facts are delivered in later decisions. No automatic estimation or target integration. Retain the full evidence-derived admissible parameter interval. Select your own final category: OCCURS, DOES_NOT_OCCUR, or INSUFFICIENT_EVIDENCE if available evidence cannot determine the event. Calibration is available when more evidence is needed. Use at most eight decisions, with the eighth reserved for finish only. One action per decision; earlier finish is allowed. Invalid responses consume a decision; no repairs, extra calls or feedback about target correctness. Keep rationale concise.",
  "contracts": {
    "search_memory": "Supply query:string, limit:1..6 and optionally apparatus_id. Searches permitted authentic calibration observations by ordinary apparatus metadata and content relevance then recency. Returns catalog IDs and metadata, not endpoint bodies; no hidden parameters or target correctness used.",
    "retrieve_memory": "Supply source_ids:list of 1..6 IDs returned by search_memory in this task. Returns the permitted authentic endpoint records. Unavailable or unsearched IDs are rejected.",
    "run_calibration": "Supply experiment_id from the two legal calibrations for this task. Executes only that calibration and returns authentic before/after sensors. Costs six simulation steps and one experiment; it never executes the target. Automatically persists provenance.",
    "estimate_parameters": "Supply source_ids from retrieved or current-task authentic endpoints and dynamics_assumptions=chronomata-symmetric-ab-v1. Supports the two offered six-step calibration geometries. Returns full inclusive admissible k interval, sensor uncertainty and provenance. IDs must belong to the current apparatus. [] returns missing_evidence; no hidden substitution.",
    "integrate_public_rules": "Calculate from explicit parameters={k:number or inclusive [lower,upper]}, the complete supplied symmetric p1/A,p2/B initial_state, horizon=1..6, threshold=number, dynamics_assumptions=chronomata-symmetric-ab-v1. Separation change=after-before; reduction=before-after in unit-box distance. Event comparison is reduction > threshold, strictly. Returns full-range event status and numerical basis from directed interval propagation; split ambiguous bounds deterministically or establish both outcomes with certified witnesses. Unresolved cases report a limitation. Arguments are your selected inputs, not evidence of a hidden coefficient; retain the entire evidence-derived interval.",
    "finish": "Supply summary exactly OCCURS, DOES_NOT_OCCUR or INSUFFICIENT_EVIDENCE; flow=stop. Select the categorical prediction yourself. The eighth decision permits only finish."
  },
  "scientific_runtime_sha256": "e668e42492ade662b3b3f2f2f5754cca54be45c01b94ef1fcff6fb603d87ad64",
  "model_digest": "25b843619e944cd0ae6069f94ff4e5e26a16e109ccbc0a66a0f05979ed70098e",
  "rules": "Only authentic observations persist. Full coefficient intervals are recomputed per task. Rule controller receives the same public context as the LLM."
}