{
 "required": true,
 "paths": [
  {
   "id": "grpo_dppo_custom_loss",
   "fixed_input": {
    "sha256": "aff0d4c84ff104219bf365a13dcb0de1c1cfc59b833ac72cb357c46e29a83f13",
    "description": "preflight_fixture.json: 16 ragged seqs (len 40-96), prompt-masked, stale current logprobs = sampler logprobs + N(0,0.6) noise (seed 0) so GRPO clip (35%) and DPPO mask (18.6%) both fire; one zero-advantage seq. Probe: code/preflight_local.py executes pinned source functions via ast extraction"
   },
   "baseline_revision": "modal-projects/spindle@e58ebef130b4cd02de6880e62d27ec358ff7132c:src/spindle/backends/megatron_runtime/common/forward_backward.py::_loss(loss_name='ppo', clip 0.8/1.2) \u2014 Spindle's stock Tinker 'ppo' loss (tinker-cookbook 0.5.9 has no client-side ppo; loss_fn='ppo' is server stock)",
   "reference_revision": "sail-sg/Stable-RL@c6af1d119c0fe30d16dd9a837d35fdf695acd226:verl/trainer/ppo/core_algos.py::compute_policy_loss_with_mask(loss_mode='dppo_binary_tv', clip_ratio=0.15, clip_ratio_c=10)",
   "implementation_revision": "code/losses.py sha256 36e744853b410170a874b80e523914ec5b0f04fa7781d96b9d04abd0e10271d3",
   "precision": "float32 CPU torch (Tinker logprobs dtype); errors computed in float64",
   "normalization": "sum over loss-mask tokens (Tinker forward_backward convention); reference agg_loss token-mean with batch_num_tokens=1, dp_size=1 == masked sum; ours per-datum unpacked list vs reference padded [B,T] layout (baseline: spindle per-sequence flattened rows)",
   "errors": {
    "baseline_loss": {
     "max_abs_error": 6.103515625e-05,
     "reference_max_abs": 429.4608154296875,
     "relative_l2_error": 1.4212043114790956e-07,
     "cosine_similarity": 1.0
    },
    "baseline_gradient": {
     "max_abs_error": 0.0,
     "reference_max_abs": 8.006952285766602,
     "relative_l2_error": 0.0,
     "cosine_similarity": 1.0
    },
    "reference_loss": {
     "max_abs_error": 3.0517578125e-05,
     "reference_max_abs": 420.54302978515625,
     "relative_l2_error": 7.25670762884611e-08,
     "cosine_similarity": 1.0
    },
    "reference_gradient": {
     "max_abs_error": 0.0,
     "reference_max_abs": 8.006952285766602,
     "relative_l2_error": 0.0,
     "cosine_similarity": 0.9999999999999999
    }
   }
  }
 ]
}