{
 "required": true,
 "paths": [
  {
   "id": "gs1_advantage_rl_loss",
   "fixed_input": {
    "sha256": "a5f33a71f4a870d0758e9060568b92cd345b1b0702d573363982bb453bbb710a",
    "description": "baseline: step-0 pilot rollouts of grpo_gs8_seed0 (64 datums) sampled from shared init weights; one server forward (importance_sampling) on the never-updated init client for logprobs (sha256 5a3c4a874a1f1cc40ac52449a26c12c3d079d4b2e1e97e63593c0a3bebfb0fed, out/fixture_grpo_gs8_seed0.pkl); reference: step-0 pilot rollouts of gs1_seed0 (64 datums) sampled from shared init weights; one server forward (importance_sampling) on the never-updated init client for logprobs (sha256 b94b5a06547c019cb94c2c679970b1d01a7796aaf5767e2d9e12ca247d9f334a, out/fixture_gs1_seed0.pkl); optimizer state fresh from shared init"
   },
   "baseline_revision": "UNRESOLVED: pin tinker-cookbook commit SHA (rl/train.py, importance_sampling loss) before running",
   "reference_revision": "UNRESOLVED: pin trajectory.ai 'You Only Rollout Once' reference code commit SHA, or a reviewed independent float64 reference implementation at a pinned SHA",
   "measured_baseline_revision": "thinking-machines-lab/tinker-cookbook@6daabe0838736557f0e802020b710073eb340287 (compute_advantages + assemble_training_data, importance_sampling)",
   "measured_reference_revision": "independent float64 pure-Python reference code/gs1_reference.py sha256 0abdddae2cfbd20e575053c5d3727af414d3c07cba19b1cb8bc3356e93650f65",
   "implementation_revision": "content sha256 be66be41ae23d88b4a2cb73cfeea1756f8ee261950645877567c2b7ff1a181c2 (code/gs1.py, gs1_reference.py, preflight.py, common.py)",
   "precision": "float64 local loss/grad over server float32 logprobs; advantages sent as float32",
   "normalization": "importance_sampling sum over tokens per datum (Tinker loss:sum), no per-batch normalization",
   "errors": {
    "baseline_loss": {
     "max_abs_error": 3.0567533258363255e-05,
     "reference_max_abs": 492.42219406518177,
     "relative_l2_error": 6.937014880483375e-08,
     "cosine_similarity": 0.9999999999999986
    },
    "baseline_gradient": {
     "max_abs_error": 7.151933201221539e-08,
     "reference_max_abs": 1.135085226398739,
     "relative_l2_error": 6.609284206365096e-08,
     "cosine_similarity": 0.9999999999999982
    },
    "reference_loss": {
     "max_abs_error": 1.4858435463338537e-05,
     "reference_max_abs": 356.5939284235209,
     "relative_l2_error": 2.7288668728921185e-08,
     "cosine_similarity": 0.9999999999999998
    },
    "reference_gradient": {
     "max_abs_error": 3.2636295665966486e-08,
     "reference_max_abs": 0.8579570351141478,
     "relative_l2_error": 2.6648511609619557e-08,
     "cosine_similarity": 0.9999999999999999
    }
   },
   "lr_scaling": []
  }
 ]
}