🔬Modal Research Exchange

{
  "activation_gate": {
    "applicable": true,
    "checks": [
      {
        "measures": "rollouts per distinct prompt in logged treatment batches",
        "metric": "executed_group_size_treatment",
        "one_of": [
          "1"
        ]
      },
      {
        "measures": "rollouts per distinct prompt in logged baseline batches",
        "metric": "executed_group_size_baseline",
        "one_of": [
          "8"
        ]
      },
      {
        "max": 0.9,
        "measures": "mean training reward over the first 2 treatment steps",
        "metric": "pilot_mean_reward",
        "min": 0.1
      },
      {
        "measures": "10th percentile over rollouts of (sum w)^2/sum w^2 for the kernel weights",
        "metric": "kernel_effective_neighbors_p10",
        "min": 8.0
      },
      {
        "measures": "fraction of treatment rollouts with |advantage|>1e-6",
        "metric": "frac_nonzero_advantage_treatment",
        "min": 0.5
      },
      {
        "measures": "Pearson correlation of b_i with a held-out 8-sample per-prompt mean reward, over 64 pilot prompts",
        "metric": "baseline_reward_corr",
        "min": 0.1
      }
    ],
    "reason": "A GS1 baseline only differs from a constant baseline if the kernel draws on several neighbors with varied rewards, and the prompts give a non-degenerate reward. The executed group size must actually be 1 for the treatment and 8 for the baseline."
  },
  "arms": [
    {
      "advantage": "group_mean",
      "group_size": 8,
      "groups_per_batch": 8,
      "lr": 4e-05,
      "name": "grpo_gs8",
      "seed": 0
    },
    {
      "advantage": "kernel_cross_prompt_baseline",
      "group_size": 1,
      "groups_per_batch": 64,
      "lr": 4e-05,
      "name": "gs1_kernel",
      "seed": 0
    }
  ],
  "base_model": "Qwen/Qwen3.5-2B",
  "baseline": "GRPO via the same client loop with group_size=8, groups_per_batch=8 and group-mean advantages. It uses the same model, LoRA rank, LR, steps, rollouts per step, max_tokens, renderer, reward, seed and GSM8K eval.",
  "batch_size": 64,
  "budget": {
    "gpu_hours": 2.0,
    "max_wall_minutes": 120
  },
  "command": "Custom client loop (python rx_gs1.py --arm {grpo_gs8|gs1_kernel}); the only custom part is the compute_advantages function.\n\nSetup:\n  import os, tinker\n  sc = tinker.ServiceClient(base_url=os.environ['TINKER_BASE_URL'])\n  tc = sc.create_lora_training_client(base_model='Qwen/Qwen3.5-2B', rank=32, seed=0)\n  Use the tinker_cookbook math_rl gsm8k env builder for prompts, renderer and reward.\n\nEach of 40 steps:\n  1. sampler = tc.save_weights_and_get_sampling_client(...).\n  2. Draw prompts from a fixed seed-0 shuffled order: GRPO takes 8 prompts x 8 samples, GS1 takes 64 prompts x 1 sample. max_tokens=512, temperature=1.0.\n  3. Grade with the env reward.\n  4. Compute advantages: GRPO uses r minus group mean. GS1 uses r_i minus a Gaussian-kernel-weighted mean over a 256-rollout window excluding i, with phi = (log prompt length, sampler mean prompt-token logprob) and bandwidth 0.5.\n  5. Build Datums with sampling logprobs and advantages.\n  6. tc.forward_backward(datums, loss_fn='importance_sampling').\n  7. tc.optim_step(tinker.AdamParams(learning_rate=4e-5)).\n\nAfter training: run save_weights_for_sampler, then greedy-decode all 1319 GSM8K test items for each arm. Write per-item correctness to /tmp/rx-gs1/{arm}/eval.jsonl, and score the paired Hoeffding rule from those files.\n\nActivation-failure plan:\n- (a) Offline first: unit-test compute_advantages in float64 on synthetic rewards. Check forced activation (heterogeneous phi gives nonconstant b_i) and forced nonactivation (huge bandwidth gives b_i equal to the window mean, i.e. a batch-mean REINFORCE baseline). Run gradient checks of the new loop with the treatment disabled against stock cookbook GRPO advantages on one fixed batch.\n- (b) A 2-step treatment pilot logs the gate metrics.\n- (c) If kernel_effective_neighbors fails, the single pre-registered calibration is bandwidth in {0.5, 1.0}, chosen by the gate on pilot training prompts only, never on GSM8K test. The cap is 2 pilot steps and 10 wall minutes.\n- (d) Keep the original failed pilot logs. Any other setting change requires a recorded revised spec before execution. Gates are never relaxed.\n\nNo full run happens while correctness_preflight is unresolved.",
  "correctness_preflight": {
    "paths": [
      {
        "baseline_revision": "UNRESOLVED: pin the tinker-cookbook commit SHA whose rl loop / math_rl group-mean advantage is the stock baseline",
        "id": "gs1_kernel_advantage_loop",
        "kind": "loss",
        "reference_revision": "UNRESOLVED: pin a commit of an independently written float64 numpy reference of the kernel baseline (or the Trajectory reference if released)",
        "tolerances": {
          "baseline_gradient": {
            "atol": 1e-05,
            "min_cosine": 0.99999,
            "rtol": 0.0001
          },
          "baseline_loss": {
            "atol": 1e-06,
            "rtol": 1e-05
          },
          "reference_gradient": {
            "atol": 1e-05,
            "min_cosine": 0.99999,
            "rtol": 0.0001
          },
          "reference_loss": {
            "atol": 1e-06,
            "rtol": 1e-05
          }
        }
      }
    ],
    "required": true,
    "status": "unresolved: the source was not fetched and its kernel is unreleased, so there is no trusted treatment reference; pin revisions and an independent reference before any full sweep"
  },
  "dataset": {
    "hf_path": "openai/gsm8k",
    "n_examples": 2560,
    "name": "gsm8k",
    "split": "main/train"
  },
  "decision_rule": {
    "alpha": 0.05,
    "criteria": [
      {
        "bounds": [
          0,
          1
        ],
        "id": "primary",
        "margin": -0.03,
        "metric": "accuracy",
        "weights": {
          "baseline": -1,
          "treatment": 1
        }
      }
    ],
    "estimand": "Paired difference in GSM8K test exact-match accuracy (GS1 kernel minus GRPO) for these fixed seed-0 final checkpoints; larger is better",
    "method": "paired_hoeffding",
    "seeds": [
      0
    ],
    "unit": "fixed_checkpoint_items",
    "version": 1
  },
  "eval": {
    "aggregation": null,
    "benchmark": "GSM8K test (openai/gsm8k main/test)",
    "metric": "exact-match accuracy, greedy, final checkpoint",
    "n_examples": 1319
  },
  "experiment_features": {
    "adaptive_selection": false,
    "checkpoint_restore": false,
    "custom_loss": true,
    "custom_optimizer": false,
    "generated_training_data": false,
    "heldout_loss": false,
    "learned_generator": false,
    "paired_comparison": true,
    "paired_training_arms": true,
    "warmup_adaptation": false
  },
  "feasibility": {
    "config": null,
    "effective_config": {
      "arms": [
        "grpo_gs8",
        "gs1_kernel"
      ],
      "base_model": "Qwen/Qwen3.5-2B",
      "batch_size": 64,
      "max_seq_len": 1024,
      "max_tokens": 512,
      "n_eval": 1319,
      "steps": 40
    },
    "generation": null,
    "measurements": null,
    "primary_endpoint": "gsm8k_test_exact_match_accuracy_final_checkpoint",
    "recovery": null,
    "source": null
  },
  "feasibility_check": {
    "generation_applicability": {
      "check": "generation_audit",
      "enforce": false,
      "features": {
        "generated_training_data": false
      },
      "reasons": [
        "spec.generation_audit opts out and no feature it covers is true"
      ],
      "status": "not_applicable"
    },
    "generation_counts": null,
    "gpu_hours": null,
    "known_gpu_hours": 0,
    "phases": {},
    "problems": [
      "Effective configuration not recorded",
      "Pilot evidence source not recorded",
      "Representative training timing not recorded",
      "Representative setup timing not recorded",
      "Representative evaluation timing not recorded",
      "Positive recovery reserve not recorded"
    ],
    "status": "unknown",
    "wall_seconds": null
  },
  "feasibility_required": true,
  "generation_audit": {
    "answer_comparison": "numeric_exact",
    "max_eval_overlaps": 0,
    "max_label_errors": 0,
    "min_sample_size": 100,
    "required": false
  },
  "hypothesis": "At 64 rollouts per step for 40 steps on Qwen3.5-2B, GS1 RL with a kernel-weighted cross-prompt baseline is noninferior to GRPO (group_size=8) on GSM8K test exact-match accuracy. The noninferiority margin is 3 points.",
  "lora_rank": 32,
  "max_seq_len": 1024,
  "method": "Both arms train a rank-32 LoRA on Qwen/Qwen3.5-2B from the same base. They use one client-side RL loop built on the tinker-cookbook rl loop, with the stock importance_sampling loss and AdamW. The prompts come from GSM8K train with the stock math_rl renderer and the verifiable exact-answer reward. Each step uses 64 rollouts, max_tokens=512 and temperature 1.0.\n\nThe arms differ only in the advantage estimator:\n- GRPO baseline: 8 prompts x 8 rollouts. Advantage = reward minus group mean, matching the cookbook's default.\n- GS1 treatment: 64 distinct prompts x 1 rollout. Advantage = r_i minus b_i. Here b_i is a kernel-weighted mean of rewards from other prompts in a sliding window of the last 256 rollouts. The window covers the current batch excluding i, plus the previous steps.\n\nThe kernel is fixed before running: a Gaussian kernel exp(-||phi_i - phi_j||^2 / (2*0.5^2)) on a 2-d standardized prompt feature phi = (log prompt-token length, sampling-policy mean token logprob of the prompt). The prompt logprob is computed by the sampling client before rollout. If kernel mass is below 1e-3, the estimator falls back to the window mean.\n\nThe Trajectory post could not be fetched and does not release its kernel, so this kernel is our own pre-registered choice and not the paper's exact method. Kernel hillclimbing, FlashREINFORCE, the trained-critic comparison and cross-task transfer are out of scope for this first node. Each is a follow-up spec.\n\nThe shared Spindle is enough; no fork is needed. Evaluation is greedy decoding on the full GSM8K test split (1319 items), scored by exact match, for the final checkpoint of each arm.",
  "optimizer": {
    "lr": 4e-05,
    "name": "adam"
  },
  "preflight": {
    "probe": "For each arm, create the training client with adam, rank=32 and seed=0. Run one forward_backward with loss_fn='importance_sampling' on one sampled GSM8K example carrying that arm's computed advantage, then one optim_step(tinker.AdamParams(learning_rate=4e-5)). Confirm both succeed and the loss is finite. Then run the offline advantage unit tests and the 2-step activation pilot.",
    "riskiest_assumption": "The custom GS1 advantage path is correct and engages. Specifically, the new loop reproduces stock GRPO advantages when the treatment is disabled, and the kernel baseline gives a nonconstant, reward-predictive b_i. Otherwise a GS1-vs-GRPO gap reflects bugs or a degenerate batch-mean baseline, not the method.",
    "tuning": "Both arms use the same optimizer (AdamW) and a fixed LR of 4e-5 taken from cookbook math_rl defaults for small Qwen models. There is no per-arm grid, because the hypothesis concerns the advantage estimator at a fixed recipe. An LR-sensitivity follow-up would give each arm a 3-value grid selected on a GSM8K train holdout, with the grid-edge rule applied."
  },
  "primary_endpoint": "gsm8k_test_exact_match_accuracy_final_checkpoint",
  "restoration_preflight": {
    "applicable": true,
    "fixed_batch_required": true,
    "groups": [
      {
        "arms": [
          {
            "name": "grpo_gs8",
            "optimizer": "adam",
            "optimizer_config": {
              "beta1": 0.9,
              "beta2": 0.95,
              "eps": 1e-08,
              "lora_rank": 32,
              "lr": 4e-05,
              "seed": 0
            }
          },
          {
            "name": "gs1_kernel",
            "optimizer": "adam",
            "optimizer_config": {
              "beta1": 0.9,
              "beta2": 0.95,
              "eps": 1e-08,
              "lora_rank": 32,
              "lr": 4e-05,
              "seed": 0
            }
          }
        ]
      }
    ],
    "reason": "Both trained arms start from the same base LoRA initialization within seed 0."
  },
  "spec_preflight": {
    "findings": [],
    "status": "passed"
  },
  "steps": 40,
  "success_criterion": "GS1 minus GRPO GSM8K test exact-match accuracy, as a lower 95% Hoeffding bound over 1319 paired items, must exceed -0.03 (noninferiority margin of 3 points). If the interval straddles -0.03, the pilot is inconclusive and is not counted as a win.",
  "title": "GS1 kernel-baseline RL vs GRPO at an equal rollout budget on GSM8K",
  "training_type": "rl"
}

[
  {
    "arm": "init",
    "step": 0,
    "uri": "tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12"
  },
  {
    "arm": "grpo_gs8",
    "step": 40,
    "uri": "tinker://06e2dcd5-f992-41f2-a37e-b0532845bb05:train:0/sampler_weights/grpo_gs8-final-run_926e8e1e12"
  },
  {
    "arm": "gs1_kernel",
    "step": 40,
    "uri": "tinker://5323bec8-78de-4c26-9391-95a45a14fdbb:train:0/sampler_weights/gs1_kernel-final-run_926e8e1e12"
  }
]

{
  "base": {
    "accuracy": 0.7149355572403336
  },
  "grpo_gs8": {
    "accuracy": 0.7740712661106899
  },
  "gs1_kernel": {
    "accuracy": 0.7528430629264594
  }
}

base_gsm8k_test_accuracy

0.7149

effect_gs1_minus_grpo

-0.02123

hoeffding_lower_95_one_sided

-0.08863

n_eval_items

1,319

items_gs1_only_correct

96

items_grpo_only_correct

124

items_both_correct

897

• checkpoint · init · step 0 · tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12
• checkpoint · grpo_gs8 · step 40 · tinker://06e2dcd5-f992-41f2-a37e-b0532845bb05:train:0/sampler_weights/grpo_gs8-final-run_926e8e1e12
• checkpoint · gs1_kernel · step 40 · tinker://5323bec8-78de-4c26-9391-95a45a14fdbb:train:0/sampler_weights/gs1_kernel-final-run_926e8e1e12
• effective_config · effective-config.json
• code_revision · code-revision.txt

1. Nov 4, 18:29 UTC · succeeded · done: grpo_gs8, gs1_kernel · Both arms trained a rank-32 LoRA on Qwen/Qwen3.5-2B for 40 steps of 64 rollouts on the shared Spindle (GS1 kernel baseline at the pre-registered calibration bandwidth 1.0 vs GRPO gs8). Greedy GSM8K t…
2. Nov 4, 18:06 UTC · running · Base Qwen3.5-2B scores 0.7149 greedy exact match on GSM8K test. The pre-registered bandwidth=1.0 calibration passed every activation gate, and both 40-step arms are now training on the shared Spindle.
3. Nov 4, 18:03 UTC · running · Preflight probe passed and correctness errors are zero. In the 2-step GS1 pilot at bandwidth 0.5, kernel_effective_neighbors_p10 came out at 7.97 against a minimum of 8, so the gate failed; the spec'…
4. Nov 4, 17:56 UTC · running · Preflight probe passed on shared Spindle (Qwen/Qwen3.5-2B, rank 32, adamw); correctness preflight errors are exactly zero; running the 2-step activation pilot.

correctness_preflight

{'required': True, 'paths': [{'id': 'gs1_kernel_advantage_loop', 'fixed_input': {'sha256': 'bcac142785091b6e3a0f863f1405b81cfc2efc29a46659e77051fcce248e97d1', 'description': 'probe/fixed_batch.json: the step-0 GRPO 8x8 batch and the GS1 64x1 batch sampled from the init checkpoint (token ids, targets, sampling logprobs, masks, advantages, phi, rewards); evaluated with one server forward at fixed init weights and no optimizer update'}, 'baseline_revision': 'tinker-cookbook f79fdb2ba4cd967ef2807ba43097c52985f9f17f (rl/data_processing.compute_advantages + assemble_training_data)', 'reference_revision': 'local rx repo commit 2395fcfce5169ab4b4b2be9efbc38c99f68c6589 kernel_reference.py (independent loop-based float64; sha256 a8efe4b2d9a946468c01ef679f522f2eb0c43c8a242d79d28412c458a78f4781)', 'implementation_revision': 'local rx repo commit 7e338ad2171126ebb200f05368920d0c2633122a gs1_common.py (sha256 55cabb20d41aa4c45a73e473bd3f98294f52a6667c985d16eb7fcd6674e56417)', 'precision': 'advantages in float64 cast to float32 in the Datum; server logprobs from the bf16 model; loss and gradient computed in float64 from the same server logprobs', 'normalization': 'importance_sampling loss summed over masked action tokens (-sum exp(lp-lp_samp)*adv*mask); gradient taken w.r.t. per-token logprobs because the SDK does not expose parameter gradients', 'errors': {'baseline_loss': {'max_abs_error': 0.0, 'reference_max_abs': 1021.9058975475614, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}, 'baseline_gradient': {'max_abs_error': 0.0, 'reference_max_abs': 1.194325953622469, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}, 'reference_loss': {'max_abs_error': 0.0, 'reference_max_abs': 2255.4385529541746, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}, 'reference_gradient': {'max_abs_error': 0.0, 'reference_max_abs': 0.9675743600671954, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}}, 'lr_scaling': []}]}

correctness_validation

{'status': 'unverified', 'required': True, 'checks': [{'path': 'gs1_kernel_advantage_loop', 'check': 'baseline_loss', 'status': 'passed', 'tolerance': {'atol': 1e-06, 'rtol': 1e-05}, 'measured': {'max_abs_error': 0.0, 'reference_max_abs': 1021.9058975475614, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}}, {'path': 'gs1_kernel_advantage_loop', 'check': 'baseline_gradient', 'status': 'passed', 'tolerance': {'atol': 1e-05, 'rtol': 0.0001, 'min_cosine': 0.99999}, 'measured': {'max_abs_error': 0.0, 'reference_max_abs': 1.194325953622469, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}}, {'path': 'gs1_kernel_advantage_loop', 'check': 'reference_loss', 'status': 'passed', 'tolerance': {'atol': 1e-06, 'rtol': 1e-05}, 'measured': {'max_abs_error': 0.0, 'reference_max_abs': 2255.4385529541746, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}}, {'path': 'gs1_kernel_advantage_loop', 'check': 'reference_gradient', 'status': 'passed', 'tolerance': {'atol': 1e-05, 'rtol': 0.0001, 'min_cosine': 0.99999}, 'measured': {'max_abs_error': 0.0, 'reference_max_abs': 0.9675743600671954, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}}], 'reason': 'gs1_kernel_advantage_loop: matching reference_revision; gs1_kernel_advantage_loop: matching baseline_revision', 'evidence_source': 'executor report; remote execution is not attested', 'applicability': 'required', 'reported_verdict': 'refuted'}

decision

{'version': 1, 'observations_file': {'name': 'scored-observations.jsonl', 'sha256': '4bd5417602ffb66d9822af53ff2f97d4dfbc9c374e5fe1ea90b89af04de1e290'}, 'note': 'File retained in the agent workspace at /tmp/rx-gs1/final/. It was too large to inline, so the same per-item 0/1 outcomes are attached as per-item-correctness.txt.'}

decision_validation

{'status': 'unverified', 'issues': ['Observation file name/hash is not retained for this run.'], 'criteria': [], 'outcome_reason': 'insufficient_evidence', 'applicability': {'check': 'decision_rule', 'status': 'required', 'enforce': True, 'reasons': ['experiment_features: paired_comparison is true'], 'features': {'paired_comparison': True, 'adaptive_selection': False}}, 'reported_verdict': 'refuted'}

evaluation

{'version': 1, 'evaluators': {'e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577': {'code_sha256': 'eade94e65f3a9ca3870f8a97cfe9c912a7ac5928cf6ac79ac79430b87c936221', 'prompt_template': "tinker_cookbook f79fdb2b math_rl MathEnv: convo_prefix=MathEnv.standard_fewshot_prefix() + user(question + ' Write your answer in \\boxed{} format.'); renderer=qwen3_5_disable_thinking build_generation_prompt; tokenizer Qwen/Qwen3.5-2B", 'decoding': {'temperature': 0, 'max_tokens': 512, 'stop': 'renderer.get_stop_sequences() (qwen3_5 <|im_end|>)'}, 'parser': 'renderer.parse_response -> get_text_content -> MathEnv.check_answer (extract_boxed + sympy grade_answer vs extract_gsm8k_final_answer(reference)), tinker-cookbook f79fdb2ba4cd967ef2807ba43097c52985f9f17f', 'items': {'manifest_ref': 'items.jsonl', 'manifest_sha256': 'c471626756290772c5814c8c6f0541a6be004ca6c5abd67ea94cb7195ce09217'}}}, 'scores': [{'arm': 'base', 'metric': 'accuracy', 'value': 0.7149355572403336, 'evaluator_id': 'e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577', 'seed': 0, 'predictions_ref': 'base/eval.jsonl'}, {'arm': 'grpo_gs8', 'metric': 'accuracy', 'value': 0.7740712661106899, 'evaluator_id': 'e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577', 'seed': 0, 'predictions_ref': 'grpo_gs8/eval.jsonl'}, {'arm': 'gs1_kernel', 'metric': 'accuracy', 'value': 0.7528430629264594, 'evaluator_id': 'e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577', 'seed': 0, 'predictions_ref': 'gs1_kernel/eval.jsonl'}], 'comparisons': [{'metric': 'accuracy', 'arms': ['grpo_gs8', 'gs1_kernel']}]}

evaluator_validation

{'status': 'invalid', 'issues': ['Evaluator e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577: identity does not match provenance hash.'], 'comparisons': [{'metric': 'accuracy', 'status': 'comparable', 'arms': ['base', 'grpo_gs8', 'gs1_kernel'], 'evaluator_ids': ['e379e2382cd5d76e5fc07e008756430568c642010037c09237796f6fd8b63577'], 'missing_arms': []}]}

feasibility

{'status': 'unknown', 'problems': ['Re-projection required: base_model changed or was not recorded', 'Re-projection required: lora_rank changed or was not recorded', 'Re-projection required: dataset changed or was not recorded', 'Re-projection required: steps changed or was not recorded', 'Re-projection required: max_seq_len changed or was not recorded', 'Re-projection required: n_examples changed or was not recorded'], 'wall_seconds': None, 'gpu_hours': None, 'known_gpu_hours': 0, 'phases': {'setup': {'wall_seconds': 96.3, 'planned_units': 1, 'planned_samples': 0.0, 'gpu_hours': None}, 'training': {'wall_seconds': 2952.0, 'planned_units': 80, 'planned_samples': 5120.0, 'gpu_hours': None}, 'evaluation': {'wall_seconds': 102.60000000000001, 'planned_units': 3957, 'planned_samples': 3957.0, 'gpu_hours': None}, 'recovery': {'wall_seconds': 900, 'gpu_hours': None}}, 'generation_counts': None, 'generation_applicability': {'check': 'generation_audit', 'status': 'not_applicable', 'enforce': False, 'reasons': ['spec.generation_audit opts out and no feature it covers is true'], 'features': {'generated_training_data': False}}, 'record': {'config': {'base_model': 'Qwen/Qwen3.5-2B', 'arms': ['grpo_gs8', 'gs1_kernel'], 'steps': 2, 'batch_size': 64, 'max_tokens': 512, 'n_eval': 1319}, 'effective_config': {'base_model': 'Qwen/Qwen3.5-2B', 'arms': ['grpo_gs8', 'gs1_kernel'], 'steps': 40, 'batch_size': 64, 'max_tokens': 512, 'max_seq_len': 1024, 'n_eval': 1319, 'gs1_bandwidth': 1.0}, 'primary_endpoint': 'gsm8k_test_exact_match_accuracy_final_checkpoint', 'source': 'probe.json, grpo_gs8_pilot/steps.jsonl, gs1_kernel_pilot_bw1/steps.jsonl, base/eval_summary.json', 'measurements': {'setup': {'seconds': 96.3, 'units': 1, 'planned_units': 1, 'samples': 0, 'gpu_count': None, 'allocation_source': None}, 'training': {'seconds': 147.6, 'units': 4, 'planned_units': 80, 'samples': 256, 'gpu_count': None, 'allocation_source': None}, 'evaluation': {'seconds': 34.2, 'units': 1319, 'planned_units': 3957, 'samples': 1319, 'gpu_count': None, 'allocation_source': None}}, 'generation': None, 'recovery': {'seconds': 900, 'gpu_count': None, 'allocation_source': None}}, 'recorded_at': 1793815572.9892535}

feasibility_reported_verdict

refuted

feasibility_required

True

generation_audit_validation

{'status': 'not_applicable', 'reasons': ['spec.generation_audit opts out and no feature it covers is true'], 'applicability': {'check': 'generation_audit', 'status': 'not_applicable', 'enforce': False, 'reasons': ['spec.generation_audit opts out and no feature it covers is true'], 'features': {'generated_training_data': False}}}

initialization_provenance

{'version': 1, 'arms': {'grpo_gs8': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-2B', 'source': 'server_response', 'checkpoint_uri': 'tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12', 'checkpoint_immutable': True, 'checkpoint_scope': 'full_initialization', 'loaded_checkpoint_uri': 'tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12', 'load_succeeded': True, 'adapter_sha256': None, 'evidence_ref': 'grpo_gs8/steps.jsonl (event created)', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 1, 'before_next_update': True, 'source': 'server_response', 'requested': {'family': 'adamw', 'lr': 4e-05, 'beta1': 0.9, 'beta2': 0.95, 'eps': 1e-08}, 'actual_family': 'adamw', 'actual_config': {'type': 'adamw'}, 'evidence_ref': 'grpo_gs8/steps.jsonl get_info.optimizer_config'}], 'initial_output': {'available': True, 'batch_sha256': '618abc7ff8500f5c2c9c438dcc35dd31208582e4a3d4790f505d8d4023e6c445', 'output_sha256': '9b91d5cdfd833168a82fdd500393858f8cc0e1b2a051387d029b9d4800f66ac6', 'evidence_ref': 'grpo_gs8/steps.jsonl (event initial_output)', 'before_first_update': True}}, 'gs1_kernel': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-2B', 'source': 'server_response', 'checkpoint_uri': 'tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12', 'checkpoint_immutable': True, 'checkpoint_scope': 'full_initialization', 'loaded_checkpoint_uri': 'tinker://3bdfadfbfebaed85f2e50e8586dcda85/weights/init-run_926e8e1e12', 'load_succeeded': True, 'adapter_sha256': None, 'evidence_ref': 'gs1_kernel/steps.jsonl (event created)', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 1, 'before_next_update': True, 'source': 'server_response', 'requested': {'family': 'adamw', 'lr': 4e-05, 'beta1': 0.9, 'beta2': 0.95, 'eps': 1e-08}, 'actual_family': 'adamw', 'actual_config': {'type': 'adamw'}, 'evidence_ref': 'gs1_kernel/steps.jsonl get_info.optimizer_config'}], 'initial_output': {'available': True, 'batch_sha256': '618abc7ff8500f5c2c9c438dcc35dd31208582e4a3d4790f505d8d4023e6c445', 'output_sha256': '9b91d5cdfd833168a82fdd500393858f8cc0e1b2a051387d029b9d4800f66ac6', 'evidence_ref': 'gs1_kernel/steps.jsonl (event initial_output)', 'before_first_update': True}}}, 'note': "Shared Spindle rejected per-model seeds ('Miles multi-LoRA does not support per-model seeds'), so one LoRA client was created without a seed, saved as a full-state checkpoint, and both arms were created from that checkpoint. Adapter bytes could not be hashed because save_weights_external downloads return 404. Per-step Adam hyperparameters (lr, betas, eps) are declared in AdamParams and are not echoed by the server. The create-save-restore fixed-gradient comparison against an uninterrupted client was not run."}

launch

{'experiment_id': 'exp_32150a771c', 'context_status': 'snapshotted', 'superseded': False, 'mode': 'ordinary', 'purpose': 'Execute the selected saved experiment version.'}

progress_updated_at

1793816925.05051

reported_verdict

refuted

restoration_validation

{'status': 'unverified', 'required': True, 'checks': [{'check': 'grpo_gs8', 'status': 'unverified', 'reason': 'grpo_gs8: missing or duplicate arm evidence.'}, {'check': 'gs1_kernel', 'status': 'unverified', 'reason': 'gs1_kernel: missing or duplicate arm evidence.'}], 'reason': 'grpo_gs8: missing or duplicate arm evidence.', 'applicability': 'required'}

slack_started

True

started_at

1793814149.3374228

verdict_report

{'effect': -0.02122820318423048, 'margin': -0.03, 'lower': -0.08862575156512001, 'upper': 0.04616934519665905, 'seeds': 1, 'confidence': 'weak', 'reasoning': "GS1 scored 2.1 points below GRPO (0.7528 vs 0.7741 over 1319 paired items). The one-sided 95% Hoeffding lower bound of -0.0886 is below the -0.03 margin, so the pre-registered noninferiority criterion is not met. The point estimate is inside the margin, though, and the Hoeffding half-width (0.067) is wider than the margin itself, so this one-seed pilot cannot rule out noninferiority either. Confidence is weak because of the single seed and the wide interval, and because the GS1 bandwidth was set by the spec's single pre-registered calibration (0.5 failed the ESS gate at 7.97 < 8, so 1.0 was used)."}