{
  "study_id": "OTHER-ROAD-2026-11-06-001",
  "version": "0.1.0",
  "status": "EXECUTED_SYNTHETIC_LEARNING_STUDY",
  "question": "Can expected checkpoint cost induce an authorized detour or undisclosed undercompletion, without a hostile objective?",
  "evidence_boundary": "Synthetic contextual-bandit agents only. No language model participants, real guardrails, external transactions, or inference about Astra's cause.",
  "design_frozen_before_first_local_run": true,
  "external_preregistration": false,
  "checkpoint_probabilities": [0.0, 0.25, 0.5],
  "checkpoint_costs": [0.0, 12.0],
  "alternate_available": [false, true],
  "audit_probabilities": [0.0, 1.0],
  "seeds": 64,
  "training_rounds": 3000,
  "exploration_probability": 0.15,
  "checkpoint_interval": 250,
  "direct_cost": 1.0,
  "alternate_cost": 3.0,
  "partial_cost": 0.5,
  "full_reward": 10.0,
  "honest_partial_reward": 5.0,
  "silent_partial_unchecked_reward": 9.0,
  "silent_partial_audited_reward": 1.0,
  "required_items": 4,
  "partial_items": 2,
  "action_order": ["direct", "alternate", "honest_partial", "silent_partial", "stop"],
  "seed_pairing": "Four Random(seed) draws per round in every cell: explore, action, checkpoint, audit. Shared exogenous draws; changing actions changes exposure.",
  "learning": "Force each available action once, then epsilon-greedy sample-average value learning. Final action is greedy; ties use action_order.",
  "primary_contrasts": ["Costly versus costless checkpoint, audited, alternate present", "Audit on versus off, costly checkpoint, alternate present", "Alternate present versus absent, costly checkpoint, audit on"],
  "primary_probability": 0.5,
  "null_controls": ["Probability changes at zero checkpoint cost", "Interchangeable narrative labels excluded from reward and state; this is structural invariance, not a model framing test"],
  "exact_grid": {"p_step": 0.01, "audit_step": 0.05, "costs": [0.0, 4.0, 12.0], "alternate_costs": [2.0, 3.0, 6.0]},
  "reporting": "All seeds/cells, full event tapes, sample-average values, greedy actions, task completeness, nondisclosure, expected regret, and nonoptimal choices retained. No confirmatory p-values or population-prevalence claims.",
  "tape_encoding": "One unsigned byte per training round, ordered by conditions.csv then seed then round; bits 0-2 action index, bit 3 direct checkpoint event, bit 4 silent-partial audit event; remaining bits zero. Gzip mtime=0."
}