You are working on a research idea from the Modal Research Exchange: GS1 kernel-baseline RL vs GRPO at an equal rollout budget on GSM8K Group-size-1 RL takes one rollout per prompt and builds its baseline from other prompts' rewards; we compare GSM8K accuracy with GRPO at the same rollout count. Goal: reproduce this on Spindle, Modal's Tinker-compatible training service, using the stock tinker SDK or tinker-cookbook pointed at TINKER_BASE_URL. Report real metrics only. If it cannot run on Spindle, say why. Selected experiment version: exp_32150a771c The saved experiment spec below defines the hypothesis, methods and success criterion. Launch mode: ordinary Purpose: Execute the selected saved experiment version. Source: https://modal-com.slack.com/archives/C0CEESNMDHS/p1793770868665239?thread_ts=1793770868.665239&cid=C0CEESNMDHS Triage: A GS1 kernel-baseline advantage is client-side code in the cookbook RL loop with the stock importance-sampling loss and AdamW, so it can be compared with GRPO on GSM8K on the shared Spindle. Experiment spec: { "title": "GS1 kernel-baseline RL vs GRPO at an equal rollout budget on GSM8K", "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.", "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.", "base_model": "Qwen/Qwen3.5-2B", "lora_rank": 32, "training_type": "rl", "dataset": { "name": "gsm8k", "hf_path": "openai/gsm8k", "split": "main/train", "n_examples": 2560 }, "optimizer": { "name": "adam", "lr": 4e-05 }, "steps": 40, "batch_size": 64, "max_seq_len": 1024, "eval": { "benchmark": "GSM8K test (openai/gsm8k main/test)", "metric": "exact-match accuracy, greedy, final checkpoint", "n_examples": 1319, "aggregation": null }, "primary_endpoint": "gsm8k_test_exact_match_accuracy_final_checkpoint", "arms": [ { "name": "grpo_gs8", "advantage": "group_mean", "group_size": 8, "groups_per_batch": 8, "lr": 4e-05, "seed": 0 }, { "name": "gs1_kernel", "advantage": "kernel_cross_prompt_baseline", "group_size": 1, "groups_per_batch": 64, "lr": 4e-05, "seed": 0 } ], "feasibility_required": true, "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.", "experiment_features": { "custom_loss": true, "custom_optimizer": false, "paired_comparison": true, "paired_training_arms": true, "checkpoint_restore": false, "warmup_adaptation": false, "generated_training_data": false, "learned_generator": false, "heldout_loss": false, "adaptive_selection": false }, "correctness_preflight": { "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", "paths": [ { "id": "gs1_kernel_advantage_loop", "kind": "loss", "baseline_revision": "UNRESOLVED: pin the tinker-cookbook commit SHA whose rl loop / math_rl group-mean advantage is the stock baseline", "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_loss": { "atol": 1e-06, "rtol": 1e-05 }, "baseline_gradient": { "atol": 1e-05, "rtol": 0.0001, "min_cosine": 0.99999 }, "reference_loss": { "atol": 1e-06, "rtol": 1e-05 }, "reference_gradient": { "atol": 1e-05, "rtol": 0.0001, "min_cosine": 0.99999 } } } ] }, "activation_gate": { "applicable": true, "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.", "checks": [ { "metric": "executed_group_size_treatment", "measures": "rollouts per distinct prompt in logged treatment batches", "one_of": [ "1" ] }, { "metric": "executed_group_size_baseline", "measures": "rollouts per distinct prompt in logged baseline batches", "one_of": [ "8" ] }, { "metric": "pilot_mean_reward", "measures": "mean training reward over the first 2 treatment steps", "min": 0.1, "max": 0.9 }, { "metric": "kernel_effective_neighbors_p10", "measures": "10th percentile over rollouts of (sum w)^2/sum w^2 for the kernel weights", "min": 8.0 }, { "metric": "frac_nonzero_advantage_treatment", "measures": "fraction of treatment rollouts with |advantage|>1e-6", "min": 0.5 }, { "metric": "baseline_reward_corr", "measures": "Pearson correlation of b_i with a held-out 8-sample per-prompt mean reward, over 64 pilot prompts", "min": 0.1 } ] }, "restoration_preflight": { "applicable": true, "reason": "Both trained arms start from the same base LoRA initialization within seed 0.", "fixed_batch_required": true, "groups": [ { "arms": [ { "name": "grpo_gs8", "optimizer": "adam", "optimizer_config": { "lr": 4e-05, "beta1": 0.9, "beta2": 0.95, "eps": 1e-08, "lora_rank": 32, "seed": 0 } }, { "name": "gs1_kernel", "optimizer": "adam", "optimizer_config": { "lr": 4e-05, "beta1": 0.9, "beta2": 0.95, "eps": 1e-08, "lora_rank": 32, "seed": 0 } } ] } ] }, "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.", "decision_rule": { "version": 1, "estimand": "Paired difference in GSM8K test exact-match accuracy (GS1 kernel minus GRPO) for these fixed seed-0 final checkpoints; larger is better", "unit": "fixed_checkpoint_items", "seeds": [ 0 ], "alpha": 0.05, "method": "paired_hoeffding", "criteria": [ { "id": "primary", "metric": "accuracy", "weights": { "treatment": 1, "baseline": -1 }, "margin": -0.03, "bounds": [ 0, 1 ] } ] }, "budget": { "gpu_hours": 2.0, "max_wall_minutes": 120 }, "generation_audit": { "required": false, "answer_comparison": "numeric_exact", "min_sample_size": 100, "max_label_errors": 0, "max_eval_overlaps": 0 }, "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.", "preflight": { "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.", "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.", "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." }, "feasibility": { "config": null, "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 }, "primary_endpoint": "gsm8k_test_exact_match_accuracy_final_checkpoint", "source": null, "measurements": null, "generation": null, "recovery": null }, "feasibility_check": { "status": "unknown", "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" ], "wall_seconds": null, "gpu_hours": null, "known_gpu_hours": 0, "phases": {}, "generation_counts": null, "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 } } }, "spec_preflight": { "status": "passed", "findings": [] } } How to work: - Run the experiment on Spindle with the stock tinker SDK (or tinker-cookbook) pointed at $TINKER_BASE_URL. Every request needs the headers Modal-Key: $MODAL_PROXY_TOKEN_ID and Modal-Secret: $MODAL_PROXY_TOKEN_SECRET, plus TINKER_API_KEY. These are already set as env vars in this session: call Spindle directly from here. Don't search the repo for them or wrap your client in a Modal app to get them. Never print their values. - Any Modal app you do need (a Spindle fork, a GPU eval) goes in the environment research-exchange-runs, named with run_bf5c66cfd2. Never deploy or run anything in the research-exchange environment (it is production), never mount the volume research-exchange-data, and never create Modal environments. - Keep the run small and inside the spec's budget. Real numbers only; never estimate or invent a metric. - Before any arm trains, run the spec's preflight probe (or, when it has none, create a training client with each arm's exact optimizer and config and take one optim_step on one example). If the shared Spindle rejects an arm's optimizer or option (for example an HTTP 400 at model creation), that arm needs a fork: fork as described below, or stop with status "failed" and outcome_reason "incomplete_execution". Never swap in a different optimizer silently. - Compare tuned arms: if the spec gives per-arm learning-rate grids, run each arm's grid under the same budget and report the selected LR per arm. A selected LR at the edge of its grid sets "grid_edge": true and keeps "confidence" weak. - Before a paired comparison, run the restoration preflight in the saved spec. Save initialization_provenance (version 1) with one entry per named arm: initialization (pinned base_checkpoint, measured adapter_sha256, source=server_response or local_hash, evidence_ref, before_first_update=true), restore_count, optimizer_checks, and initial_output. All arms/replays in an initialization group must load the same initial checkpoint/adapter. A matching seed, step-0 loss, or base accuracy does not establish matched initialization (LoRA B may be zero). optimizer_checks must cover creation (after_restore=0) and EVERY restore (1..restore_count), before_next_update=true, actual optimizer family and actual_config from a server_response plus evidence_ref. Compare against each arm's requested optimizer/config. Creating from saved state can silently fall back to AdamW: creating TriMuon first then loading weights is a workaround to verify, not proof. Stop on mismatch. Do not invent SDK inspection methods; if no server response exposes the active optimizer, use source=declared and actual=null: this stays unverified. On a fork, add a minimal observation of the concrete server optimizer after restoration; retain its code revision and response. Check a create-save-restore fixed-gradient update against the uninterrupted client before a sweep. initial_output records available=true, batch_sha256 (token IDs/targets/masks), output_sha256 (all full-precision logits/logprobs, not an averaged loss), evidence_ref and before_first_update=true; compare across the paired group. If capture is unavailable, record available=false with a reason. For exact replay, require it in the spec. Retain the evidence before cleanup. Publish the provenance while running; full training waits for verified checks. - Initialization proof may use measured adapter_sha256, or an immutable FULL initialization checkpoint with source=server_response, checkpoint_uri, checkpoint_immutable=true, checkpoint_scope=full_initialization, loaded_checkpoint_uri (from the actual successful load response), load_succeeded=true, evidence_ref and before_first_update=true. Keep base_checkpoint in both proof types. A declared URI or base-model-only load does not prove matching adapters. Compare all available hashes and URI identities; mixed proofs need a shared identity. A lone resumed training arm may form one group only when checkpoint_restore=true, paired_training_arms=false, and it records at least one actual restore. Exclude frozen, untrained baselines from optimizer restoration groups. - If the spec's base_model is not served on Spindle, use the closest served model of the same family (e.g. Qwen/Qwen3.5-2B for a smaller Qwen3.5) and say so in the summary; a missing model is not a reason to stop. - If the shared Spindle lacks something the experiment needs (an optimizer, loss, or feature), you may fork Spindle (https://github.com/modal-projects/spindle), add it, and deploy your fork following its README as a throwaway Modal app named spindle-fork-run_bf5c66cfd2 in the research-exchange-runs environment. Call your fork with Modal-Key: $RUNS_PROXY_TOKEN_ID and Modal-Secret: $RUNS_PROXY_TOKEN_SECRET (the shared Spindle token gets 401 on research-exchange-runs); they're also set as env vars here. Run every arm on the fork so the comparison is fair. Never push to, redeploy, reconfigure, or stop the shared Spindle deployment, and never open PRs or issues on the Spindle repo; if the change looks worth upstreaming, propose it in your report. Include the fork's diff and app name in your report, and when you are done stop spindle-fork-run_bf5c66cfd2 and every app your fork deployed (Spindle deploys separate trainer, lora and inference apps; give them names containing run_bf5c66cfd2), with `modal app stop -e research-exchange-runs `, then check `modal app list -e research-exchange-runs` shows none of yours still deployed. Stop nothing else. - Spindle errors like "trainer capacity exhausted" or 429/503 are usually temporary because other runs share it: retry with backoff (e.g. 1, 2, 4 minutes) and update your structured output while you wait. If no comparison training step has run yet and it is still refused after about 10 minutes, don't keep waiting: release the Spindle clients this session created where the SDK lets you (yours only), then finish with "status": "failed", no "verdict", "outcome_reason": "incomplete_execution" and "capacity_block". The exchange may re-queue the run and start a fresh session later, within the spec's wall-time budget. If comparison steps already ran, keep retrying for up to 30 minutes so finished work isn't lost, then stop and report "capacity_block" with the real step count. - Release probe and setup clients you no longer need before starting the paired arms. - Clean up after yourself before you finish: stop every Modal app, sandbox, or other resource you started for this run (by exact name), and nothing else. Use the structured output tool to keep your output updated with exactly this JSON shape, and set "status" to "succeeded" or "failed" when you are done: {"status": "running" | "succeeded" | "failed", "summary": "<= 2 plain sentences: what ran and what it showed", "metrics": {"": , ..., "arms": {"": {"": }}}, "completed_arms": [""], "verdict": "supported" | "refuted", "confidence": "strong" | "weak", "effect": , "margin": , "ci": [, ], "seeds": , "grid_edge": true | false, "reasoning": "2-4 sentences: what the evidence showed, the effect against the margin, what limits confidence", "outcome_reason": "incomplete_execution", "capacity_block": {"error": "", "comparison_steps_completed": , "released_clients": true | false}} Include "capacity_block" only when Spindle capacity stopped the run. Put each arm's numbers under "arms", keyed by the arm's name, not in flat keys like "_". Update it while you work, not only at the end: post the base eval as soon as you have it, and each arm's metrics under "arms" with the arm added to "completed_arms" as soon as that arm finishes. Keep every number you posted earlier in later updates; if the session ends early the finished arms still count as partial results. Leave "verdict" out until you are done. A verdict is a direction plus a confidence; there is no "inconclusive". Once the comparison measured the effect, "verdict" follows the effect against the margin: "supported" when it meets the success criterion, "refuted" otherwise (ties, effects inside the tie band, mixed criteria). "confidence" is "strong" only when the interval or at least two seeds settle the margin and nothing limits the result; "weak" otherwise, with the limit named in "reasoning" (one seed, a wide interval, a grid-edge LR, a changed experiment, a check that did not pass). Finishing execution alone does not support the hypothesis. If the comparison never measured the effect (stopped early, a gate failed before training, an arm could not run), finish with "status": "failed", no "verdict" and "outcome_reason": "incomplete_execution": that is shown as no result, not a verdict. Use "succeeded" when execution completed; "failed" when it stopped early. Changes from the spec: the output also has "deviations": [{"field": "model" | "objective" | "effective_batch" | "steps" | "dataset" | "evaluator" | "endpoint", "original": "", "effective": "", "reason": ""}], "original_hypothesis_tested": true | false, "tested_instead": "" Use [] when you run the spec as written. effective_batch is sequences per optimizer step after accumulation. Before you train with any change (a substituted model, a different loss or RL objective, a smaller batch, another dataset or split, a different grader or eval set, a different decision metric or step), add it to "deviations" with status "running" first, then continue. A change you only report at the end is flagged as unrecorded. If the change means the run no longer tests the spec's hypothesis, set "original_hypothesis_tested": false and fill "tested_instead". Changing the objective (e.g. GRPO to REINFORCE) or the endpoint always counts as not testing it: the exchange then records that your verdict judges the changed experiment, which lowers its confidence. Correctness preflight is required before ANY full sweep using a new/custom/forked loss or optimizer, even if discovered after launch. First reproduce the stock baseline THROUGH the new path, with treatment disabled; then compare the enabled implementation to the pinned treatment reference. Use the same fixed inputs, weights, optimizer state, masks, normalization, precision and seeds. Loss paths need both loss AND full gradient errors; optimizer paths need update errors. Similar gradient norms alone do not prove parity. Check packed versus unpacked layouts on meaningful fixed inputs, including stale inputs when relevant. Reset state independently for each LR probe, measure update norms (per factor when applicable), and verify the prespecified LR scaling. STOP before full training/sweeps if any check is missing or fails. If the spec has no accepted preflight plan for a discovered custom path, report it as required/unverified and request a spec revision; do not pick a loose tolerance after observing errors. Activation and checkpoint provenance are separate gates. Add correctness_preflight alongside status/summary/metrics/verdict in structured output (or RESULT): {"required": true, "paths": [{"id": "", "fixed_input": {"sha256": "<64 hex>", "description": ""}, "baseline_revision": "", "reference_revision": "", "implementation_revision": "", "precision": "", "normalization": "", "errors": {"": {"max_abs_error": , "reference_max_abs": , "relative_l2_error": , "cosine_similarity": }}, "lr_scaling": [{"learning_rate": , "update_norm": }, ...]}]}. For stock-only execution use {"required": false, "paths": []}. Keep probe scripts and inputs as report artifacts. max_abs_error is max(abs(actual-reference)); relative_l2_error is norm(actual-reference)/norm(reference). Acceptance requires max_abs_error <= atol + rtol*reference_max_abs and any prespecified min_cosine. LR scaling checks the relative change of update_norm/LR against the first probe. These are measured errors, not self-declared pass flags. Small standard-library helpers parity_errors and lr_scaling_error live in research_exchange.correctness; `python -m research_exchange.correctness spec.json preflight.json` validates evidence and exits nonzero unless passed/not_applicable. Gate full-sweep launch on that exit code when available. The Exchange validates/persists reported checks and withholds definitive judgments when required checks fail or are missing. It cannot prevent autonomous training calls or attest reported remote measurements. Before full training, follow the spec's activation_gate: measure every prespecified diagnostic in a short pilot using the exact full-run model, loss path, dtype, lag and updates per batch. Stop before full training when a required diagnostic is missing or outside its accepted range; do not start full training. Do not change ranges after measuring. Stopping full training does not end the investigation. First use saved evidence to distinguish a code or instrumentation defect, a wrong reference/variant, and an inactive treatment regime. Check deterministic inputs that must activate and must not activate, including gradients and boundary decisions. Preserve the actual loss inputs and masks, with hashes. For policy constraints, record sampler checkpoint/version and immutable rollout logprobs, current logprobs before each update, token identities, advantage signs, and loss-token denominators. Log signed and absolute drift tails (p90/p99/max), threshold exceedance counts and removed-gradient counts; zero-advantage and padding tokens must not count as treatment engagement. Distinguish optimizer calls, disjoint minibatches, epochs over reused data, and the number of intervening updates since sampling. Continue offline diagnosis after a gate failure. Calibration training may continue only under a recorded, source-justified diagnostic plan with fixed settings, selection rule, update and wall-time caps within the remaining authorized budget. Keep calibration data separate from held-out outcome evaluation. If the accepted spec has no such plan, retain a concrete revised spec/deviation and cheapest next diagnostic; do not execute changed training settings under the old spec. Never lower a threshold or relax a gate merely to pass it. In the RESULT summary, state the verified or suspected cause, evidence gaps, fixes and cheapest next diagnostic. Include activation_calibration in RESULT: {"values": {"": }, "note": ""}. Preserve it separately from scalar metrics. The server validates reported calibration, not remote execution; a missing or failing required check withholds a definitive scientific judgment. Freeze the evaluator before scoring any arm. Report an evaluation object alongside metrics: { "version": 1, "evaluators": { "": { "code_sha256": "", "prompt_template": "", "decoding": {"temperature": 0, "max_tokens": 512, "stop": []}, "parser": "", "items": {"ids": [""]} }}, "scores": [{"arm": "baseline", "metric": "accuracy", "value": 0.8, "evaluator_id": "", "seed": 0, "predictions_ref": "predictions.jsonl"}], "comparisons": [{"metric": "accuracy", "arms": ["baseline", "treatment"]}] } evaluator_id is hashlib.sha256(json.dumps(provenance, sort_keys=True, separators=(",", ":"), ensure_ascii=True, allow_nan=False).encode()).hexdigest(). Include seed per score when applicable. For large item lists use items {"manifest_ref": "items.jsonl", "manifest_sha256": ""} instead of ids; references are evidence file names. Include evaluator_id, metric, arm and item_id in each prediction record. Keep the scored model/checkpoint separate from evaluator identity. Any grader, parser, prompt, decoding or item change requires re-evaluating EVERY compared arm/seed and the base model with the new evaluator, including already completed arms. Until then report no verdict. Keep old scores identified with their original evaluator; never relabel them. When spec.decision_rule is present, report decision {version:1, observations:[ {item_id:"",arm:"treatment",seed:0,metric:"accuracy",value:1, evaluator_id:""}, ...]} for every arm, item and scheduled seed. For large evaluations, avoid putting all rows in structured output. Attach a UTF-8 JSONL file with those same scored rows (or a JSON array / object with observations), retain it via evidence.files as kind metrics, then report only decision {version:1, observations_file:{name:"scored-observations.jsonl",sha256:""}}. Use exactly one of observations or observations_file. The reference is a retained file name and hash, never a path or URL. The server reads only this run's verified immutable evidence copy, capped at 10 MiB and 100000 observations; missing/corrupt files stay unverified. Upload a new immutable file name/hash for each update. Keep selection metadata in the decision object. No scores, item IDs, seeds or evaluator IDs are inferred. Values are finite measured outcomes in the endpoint's original units; never clip or normalize them. For bounded accuracy, include failures as zero, not just successful items. Report unrounded per-arm/seed score means so they reproduce the paired observations. Record evaluation provenance and per-arm/seed scores with explicit items.ids (hashed manifest-only item lists are not yet resolved by the decision validator). Persist the raw predictions through evidence too. The server recomputes paired effects and uncertainty; it does not trust an agent-supplied interval or parse prose. Never add a seed after seeing the gap or treat test items/repeated evals as independent training seeds. For adaptive selection, observations must use only confirmation items, and decision.selection must contain frozen_before_confirmation:true. For kind weight_edit also report budgets:{:{values_touched:N,proposals:N}} for selected and random-control arms. Report "confidence": "strong" only when every criterion's interval settles the prespecified margin; otherwise "weak". Retain review evidence in structured output under "evidence": {"version":1,"files":[{"name":"run.py","kind":"script","content":""}, {"name":"effective-config.json","kind":"effective_config","content":""}], "checkpoints":[{"arm":"baseline","step":100,"uri":"tinker://training-id/weights/checkpoint-name"}], "per_arm_metrics":{"baseline":{"accuracy":0.5}}} Include files for the executed script, effective config, code revision (kind code_revision) and any patch (kind code_patch), per-item predictions (kind predictions), and per-arm metrics. Include generated_tasks when training on generated data. Use JSONL predictions with item_id, arm, prediction and reference in each row; retain seed, checkpoint, metric and evaluator_id metadata when available. Invalid outputs must remain rows. For larger files attach them in Devin and set source_url to the HTTPS /attachments/... URL instead of content; never use a session page URL, VM path or arbitrary external URL. Optional sha256 must match the original bytes. The web server copies accepted files into this run's storage and exposes reviewer downloads. Post evidence as stages finish, before cleanup. File names are immutable: use a new name for each stage/checkpoint/update. Each file is capped at 10 MiB, with 64 files / 40 MiB total per run. Never include .env, credentials, auth headers, secret values or credential-bearing URLs. Read credentials from environment variables in scripts. Checkpoint URIs are references, not copied weights; omit credentials and signed URL query strings. Missing evidence is shown as unknown or partial, never as verified. Do not invent an artifact or checkpoint that does not exist. Feasibility is a measured projection, separate from scientific metrics and actual GPU usage. After resolving the served model, run a representative timing pilot before the full comparison. Publish feasibility in structured output before continuing; repeat it after changing model, arms, context, batch, steps, dataset size or evaluation count. Include fork/setup time, training AND held-out evaluation, generation AND filtering samples, measured yield, and a recovery reserve. Never scale timing by parameter count or assume concurrency is free. Measure the actual schedule. Preserve the primary endpoint when downsizing: drop optional secondary evals, extra seeds or LR points first, then shorten training with an explicit deviation; never replace held-out eval by train loss. If the primary comparison cannot fit, stop with no verdict. Use this additional structured field (null means no measurements yet): "feasibility": {"config": , "effective_config": , "primary_endpoint": , "source": , "measurements": { "setup": {"seconds": , "units": 1, "planned_units": 1, "samples": 0, "gpu_count": null, "allocation_source": null}, "training": {"seconds": , "units": , "planned_units": , "samples": , "gpu_count": null, "allocation_source": null}, "evaluation": {"seconds": , "units": , "planned_units": , "samples": , "gpu_count": null, "allocation_source": null}}, "generation": null, "recovery": {"seconds": , "gpu_count": null, "allocation_source": null}}. For generated tasks replace generation:null by {"attempted": , "accepted": , "target_accepted": , "filter_samples_per_candidate": } and add generation and filter measurements (units=candidates and filter rollouts respectively). Include an audit measurement for independent task re-solves when the shared generation_audit applicability is enforced. generated_training_data=true overrides a legacy required=false flag; unresolved new feature declarations need resolution and cannot skip generation/filter/audit sizing. Record GPU counts only from explicit allocation evidence, including all trainer/inference GPUs allocated for the timed phase; null for shared Spindle when allocation is unknown. CPU-only phases may explicitly use 0. Keep projected wall seconds, projected GPU-hours and observed usage separate. Do not invent timings or GPU-hours to fill this contract. Resolve applicability with experiment_features.check_applicability(spec,"generation_audit"). An explicit generated_training_data=true requires the audit. False can establish non-applicability when no explicit contract requires the check; missing feature evidence stays unknown. For training on generated tasks, complete generation_audit BEFORE training. Running generated Python only checks execution; it does not verify labels. Save all candidates as JSONL in evidence.files kind generated_tasks: item_id (unique string), prompt, answer, disposition (accepted/rejected/execution_error). Retain generated code and execution errors when available. Save the FULL frozen evaluation set as JSONL (item_id,prompt), kind generated_tasks; this is for overlap checking, never training. Freeze both files before sampling. Report generation_audit with version=1, stage="before_training", dataset_file, evaluation_file, audit_file (a new immutable .json name), sample_seed (integer), counts={generated,accepted,rejected,execution_errors}, samples=[{item_id,independent_answer,resolver_id,method:"independent_resolve",reasoning,blind_to_label:true}], label_error_count, overlap_count. Select min(accepted_count, spec.generation_audit.min_sample_size or 100) accepted tasks deterministically with research_exchange.generation_audit.select_sample(tasks,size,seed), which sorts by SHA256 of the JSON [seed,item_id]. Re-solve those problems from their TEXT without seeing the stored answer or generated code (human or a separate solver invocation); record solver identity and worked solution. Never reuse the generator's Python as independent evidence. Report observed disagreements and the sample denominator; answer agreement with a declared solution does not establish semantic truth. New specs must explicitly set generation_audit.answer_comparison to numeric_exact for strict scalar numbers/fractions (zero tolerance), or categorical_exact for nonempty string labels. Categorical comparison is exact and case-sensitive, with no whitespace normalization or synonym mapping. Choose from the label contract, never the experiment/dataset name; missing or unsupported strategies remain unverified. Use answer_agreement(label, independent_answer, strategy) with that declared strategy, and evaluation_overlap for NFKC/casefold/whitespace-normalized exact prompt overlap against the full held-out set. This does not detect paraphrases or semantic leakage. Report duplicate and overlap counts, and persist the dataset and audit manifest through evidence before cleanup. Failed/missing audits block a definitive scientific verdict. Call generation_audit.validate(report, spec.generation_audit, tasks=tasks, evaluation=evaluation, require_comparison=True) and require status="passed" before the first optimizer update. Do not invent passing re-solve evidence to get past this check. If any sampled label is wrong or overlaps evaluation, repair/reject affected tasks and freeze NEW dataset and audit file versions, then repeat the audit before training. Do not train on known bad labels as-is.