{
 "run_id": "run_21b1618aa7",
 "experiment_version": "exp_e6b0965af3",
 "kind": "bounded_activation_diagnostic",
 "authorized_by": "requester instructions in launch-instructions.md (sha256 b42946cb...ad51)",
 "original_spec_unchanged": true,
 "original_gate_report_preserved": "run_19badd842c pilot-dppo-steps.jsonl (sha256 ee68c6b8951cb73d3949c28b2986dda1429624e8cd3c6ec28eb2b07871be590b); original second-minibatch unweighted-mean gate failed (DPPO mask 0, PPO clip 0.00045687)",
 "model": "Qwen/Qwen3.5-2B", "lora_rank": 32,
 "loss": {"trained": "dppo_binary_tv", "delta": 0.15, "clip_ratio_c": 10.0, "variant_note": "Stable-RL@c6af1d1 released-code C=10 variant, not paper C=5", "counterfactual_ppo_eps": 0.2, "counterfactual_ppo_trained": false},
 "optimizer": {"type": "adamw", "learning_rate": 2e-05, "beta1": 0.9, "beta2": 0.95, "eps": 1e-08, "weight_decay": 0.0, "grad_clip_norm": 0.0},
 "renderer": "qwen3_5_disable_thinking", "convo_prefix": "MathEnv.standard_fewshot_prefix", "temperature": 1.0, "max_tokens": 512, "advantages": "group-mean centered (cookbook compute_advantages)",
 "initial_state": "restore immutable shared init tinker://21e32976f18f971463410adff4e364e4/weights/rx19badd842c_init (untrained, saved before any update in run_19badd842c); never trained pilot weights",
 "data": "ONE batch: Gsm8kDataset(train, seed=0) batch index 0, 32 groups x 8 responses, sampled once from the restored init sampler; no further sampling",
 "schedule": {"minibatches": 8, "groups_per_minibatch": 4, "complete_groups": true, "epochs": 2, "max_optimizer_updates": 16, "old_logprob_refresh": false, "frozen_hash_assert_before_every_reuse": true},
 "sentinels": "first datum of each of the 8 minibatches (same training batch); trainer logprobs via forward at init and after updates 1,2,4,8,16",
 "analysis": {"population": "ALL 8 second-epoch loss calls (updates 9-16), pooled by loss-token counts",
   "checks": {"dppo_masked_token_fraction": [0.001, 0.3], "ppo_counterfactual_clip_fraction_min": 0.0005, "policy_drift_abs_p_q90_pooled_min": 0.005, "nonzero_removed_gradient_per_present_adv_sign": true},
   "revisions_vs_original_gate": ["aggregation: pooled-by-token over all 2nd-epoch calls instead of unweighted mean of 2nd-minibatch fractions", "schedule: 8 minibatches x 2 epochs on one frozen batch instead of 2 disjoint minibatches per fresh batch"],
   "no_step_selection": true, "no_threshold_change": true},
 "budget": {"max_wall_minutes": 12, "gpu_hour_cap": 2.0, "stop_at": "16 updates or 12 minutes regardless of activation"},
 "excluded": ["full training", "GRPO trained arm", "GSM8K test evaluation", "LR/delta sweeps", "held-out selection", "automatic follow-on launch"]
}