{"schema":"cain.e89.frontier.preregistration.v1","registered_at":"2026-11-17T23:08:46Z",
 "authorization":"owner message 2026-11-17: 'yes use my anthropic key it has funds' then 'yes now proceed'",
 "model":"claude-opus-5-5 (same version every arm; Claude API first-party, $4 in / $20 out per 1M)",
 "hard_ceiling_usd":50, "per_arm_ceiling_usd":{"icl":30,"cain_structured":10,"cain_retrieval":10},
 "estimate_usd":{"per_arm_stateless_baseline":"~1.5","icl_rollout":"~14","cain_rollout":"~2 each","total":"~23"},
 "stopping":"each proxy refuses any request whose worst-case cost would cross its ceiling; a refused arm is reported as incomplete, never re-run past the ceiling",
 "benchmark":"CL-Bench blind_spectrum_monitoring, default schedule (90 scans), corpus mixed_grid_lifecycle, 1 run per arm, official scorer",
 "call_path":"every arm uses the harness's litellm chat route (provider_mode litellm_chat for icl), so all arms share one call path through the metering proxy",
 "arms":{"A_stateless":"harness reset-each-instance baseline inside each arm's run","B_icl":"upstream icl, context cap 150k","C_cain_retrieval":"cain_memory retrieval","D_cain_structured":"cain_memory structured"},
 "frontier_memory_arm_E_skill_F_causal":"NOT EVALUABLE on BSM (hidden reward)",
 "expected_information_gain":"whether CAIN memory beats the same frontier model's ICL and stateless baseline on an externally authored continual-learning task; the leaderboard has icl-claude-opus-4.7 (cumulative_gain 0.153) as a reference point, different model version",
 "hypotheses":{"H1":"D cumulative_gain > 0","H2":"D total_reward > B","H3":"C total_reward <= B"}}
