You are working on a research idea from the Modal Research Exchange: Self-generated math tasks vs GSM8K train for LoRA RL on Qwen3.5-2B Qwen3.5-2B writes its own checkable math word problems for RL training; we compare GSM8K test accuracy against the same RL run on real GSM8K problems. 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_c2bf4b4ffc 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.enterprise.slack.com/archives/C0AEUHUHZL4/p1792592566946059 Triage: Fits math_rl LoRA RL on a served Qwen model. The task generator is the frozen base model, sampled through the sampling client with code-computed labels. The generated set plugs in as a client-side dataset in the cookbook RL loop. Experiment spec: { "title": "Self-generated math tasks vs GSM8K train for LoRA RL on Qwen3.5-2B", "hypothesis": "Two things should hold on the 1,319 GSM8K test problems: (1) LoRA RL on 640 audited tasks written by frozen Qwen3.5-2B reaches GSM8K test exact-match accuracy no more than 3 points below the same RL on 640 GSM8K train problems; (2) it recovers more than half of that run's gain over the untrained base model.", "method": "The linked repo (browser-use/rl-env-use) does web-mining and system-prompt hillclimbing with no RL weight training. This experiment borrows only the environment-generation premise and tests it directly with RL.\n\nGeneration (custom, client-side):\n- Frozen base Qwen/Qwen3.5-2B, through the sampling client at T=1.0, is prompted with one fixed instruction: 'Write one grade-school math word problem with a single integer answer, then a short Python function solve() that returns the answer.'\n- Each solve() runs in a sandbox with a 2 s timeout. The label is its integer return value.\n- Filters, fixed before any training:\n - parse and execution succeed;\n - the problem text does not contain the answer;\n - the problem is 20 to 120 words long;\n - normalized exact match against all GSM8K test questions finds no overlap;\n - the base model solves the problem from text alone in 1 to 6 of 8 samples at max_tokens=512, so the task is learnable and not trivial.\n- The first 640 accepted tasks, in generation order, form the training set. No selection is made on eval results.\n\nGeneration audit, before any training:\n- Draw 100 random accepted tasks.\n- Each is re-solved blind from the problem text alone by Qwen/Qwen3.5-397B-A17B, which is used for sampling only. Record solver identity and worked solutions.\n- Compare answers numeric_exact.\n- Any label error, or any eval overlap, stops the run for repair and re-audit. Do not train.\n- Save the generated/accepted/rejected/execution-error counts and the data and audit manifests as evidence.\n\nTraining:\n- Both RL arms start from fresh LoRA rank 32 on Qwen3.5-2B and use the identical stock math_rl loop:\n - group_size=8, groups_per_batch=16, lr=4e-5 (AdamW), max_tokens=512, 40 steps, seed 0;\n - the same renderer and the same answer-extraction exact-match reward.\n- Baseline arm: 640 GSM8K train problems (first 640 after a fixed seed-0 shuffle).\n- Treatment arm: the 640 generated tasks. This is the only difference between the arms.\n- A third, untrained arm: base model, eval only.\n\nEval:\n- All three arms on the full GSM8K test split (1,319 problems), greedy decoding, max_tokens=512, the same prompt and extractor.\n- Results are paired per item.\n\nPreflight: no custom loss or optimizer. The only custom code is the dataset builder and the generation/audit script.", "base_model": "Qwen/Qwen3.5-2B", "lora_rank": 32, "training_type": "rl", "dataset": { "name": "gsm8k (baseline train / eval) and self-generated tasks (treatment train)", "hf_path": "openai/gsm8k", "split": "main/train for baseline, main/test for eval", "n_examples": 640 }, "optimizer": { "name": "adam", "lr": 4e-05 }, "steps": 40, "batch_size": 16, "max_seq_len": 1024, "eval": { "benchmark": "GSM8K test (openai/gsm8k main/test)", "metric": "exact-match accuracy, greedy, max_tokens=512", "n_examples": 1319, "aggregation": null }, "primary_endpoint": "gsm8k_test_exact_match_accuracy_greedy_1319", "arms": [ { "name": "base", "trained": false, "data": "none", "seed": 0 }, { "name": "rl_gsm8k", "trained": true, "data": "640 GSM8K train problems", "optimizer": "adam", "lr": 4e-05, "steps": 40, "seed": 0 }, { "name": "rl_selfgen", "trained": true, "data": "640 audited Qwen3.5-2B-generated tasks", "optimizer": "adam", "lr": 4e-05, "steps": 40, "seed": 0 } ], "feasibility_required": true, "feasibility": null, "baseline": "rl_gsm8k is the same stock math_rl LoRA RL on 640 GSM8K train problems, with identical model, rank, lr, group size, batch, steps, seed, renderer, reward and eval. The untrained base model is evaluated as a reference for gain recovery.", "experiment_features": { "custom_loss": false, "custom_optimizer": false, "paired_comparison": true, "paired_training_arms": true, "checkpoint_restore": false, "warmup_adaptation": false, "generated_training_data": true, "learned_generator": false, "heldout_loss": false, "adaptive_selection": false }, "correctness_preflight": { "required": false, "reason": "Both arms use the unchanged math_rl importance-sampling loss and the stock AdamW backend. Only the training-problem source differs, and that is client-side data code with no custom loss or optimizer." }, "activation_gate": { "applicable": true, "reason": "The treatment is a change of training data. The pilot confirms three things: (1) each arm actually executes on its intended data; (2) generated tasks give non-degenerate GRPO groups, so RL has signal (all-0 or all-1 groups give zero advantage); (3) the generated set is not contaminated with eval items.", "checks": [ { "metric": "treatment_fraction_batches_from_generated_set", "measures": "share of rl_selfgen training problems whose ids are in the audited generated manifest", "min": 1.0, "max": 1.0 }, { "metric": "baseline_fraction_batches_from_gsm8k_train", "measures": "share of rl_gsm8k training problems whose ids are in the GSM8K train manifest", "min": 1.0, "max": 1.0 }, { "metric": "selfgen_frac_groups_nonzero_reward_variance_pilot", "measures": "fraction of rl_selfgen groups in the first 2 pilot steps with mixed rewards", "min": 0.3 }, { "metric": "selfgen_mean_train_reward_pilot", "measures": "mean reward on rl_selfgen pilot steps", "min": 0.15, "max": 0.85 }, { "metric": "generated_eval_exact_overlaps", "measures": "normalized exact matches between accepted tasks and the GSM8K test questions", "min": 0.0, "max": 0.0 } ] }, "restoration_preflight": { "applicable": true, "reason": "Two RL arms are trained from a shared starting point. Initial identity must be shown by matching logprobs on a fixed tokenized batch, from both freshly created LoRA clients, before the first update. The base model name alone is insufficient. The active optimizer and its config are verified after creation.", "fixed_batch_required": true, "groups": [ { "arms": [ { "name": "rl_gsm8k", "optimizer": "adam", "optimizer_config": { "learning_rate": 4e-05, "beta1": 0.9, "beta2": 0.95, "eps": 1e-08, "lora_rank": 32, "seed": 0 } }, { "name": "rl_selfgen", "optimizer": "adam", "optimizer_config": { "learning_rate": 4e-05, "beta1": 0.9, "beta2": 0.95, "eps": 1e-08, "lora_rank": 32, "seed": 0 } } ] } ] }, "success_criterion": "Validated if both criteria hold, with simultaneous 95% paired Hoeffding intervals over the 1,319 GSM8K test items:\n(1) rl_selfgen minus rl_gsm8k accuracy has a lower bound above -0.03 (noninferiority);\n(2) rl_selfgen - 0.5*rl_gsm8k - 0.5*base has a lower bound above 0 (more than half the gain recovered).\nFailed if either upper bound falls below its margin. Otherwise inconclusive. A one-seed pilot may be inconclusive, and it says nothing about training-seed variance.", "decision_rule": { "version": 1, "estimand": "GSM8K test greedy exact-match accuracy of fixed seed-0 checkpoints of Qwen3.5-2B LoRA RL on self-generated tasks, relative to RL on GSM8K train and to the untrained base; larger is better", "unit": "fixed_checkpoint_items", "seeds": [ 0 ], "alpha": 0.05, "method": "paired_hoeffding", "criteria": [ { "id": "noninferiority_vs_gsm8k_rl", "metric": "accuracy", "weights": { "rl_selfgen": 1, "rl_gsm8k": -1 }, "margin": -0.03, "bounds": [ 0, 1 ] }, { "id": "recovered_half_gain", "metric": "accuracy", "weights": { "rl_selfgen": 1, "rl_gsm8k": -0.5, "base": -0.5 }, "margin": 0.0, "bounds": [ 0, 1 ] } ] }, "budget": { "gpu_hours": 2.0, "max_wall_minutes": 120 }, "generation_audit": { "required": true, "answer_comparison": "numeric_exact", "min_sample_size": 100, "max_label_errors": 0, "max_eval_overlaps": 0 }, "command": "Step 1, generation, filtering and audit (custom script):\n python gen_tasks.py --base_url $TINKER_BASE_URL --generator Qwen/Qwen3.5-2B --n_accept 640 --filter_samples 8 --keep_band 1-6 --overlap_ref openai/gsm8k:main/test --out /tmp/rx-selfgen/tasks.jsonl\n This uses only tinker.ServiceClient(base_url=os.environ['TINKER_BASE_URL']).create_sampling_client plus sandboxed exec of solve().\n python audit_tasks.py --base_url $TINKER_BASE_URL --solver Qwen/Qwen3.5-397B-A17B --sample 100 --compare numeric_exact --max_label_errors 0 --out /tmp/rx-selfgen/audit.json\n The run stops if the audit fails.\n\nStep 2, timing pilot: 2 steps per arm plus 100 eval items. Publish feasibility before continuing.\n\nStep 3, baseline:\n python -m tinker_cookbook.recipes.math_rl.train model_name=Qwen/Qwen3.5-2B base_url=$TINKER_BASE_URL env=gsm8k group_size=8 groups_per_batch=16 learning_rate=4e-5 lora_rank=32 max_tokens=512 max_steps=40 seed=0 log_path=/tmp/rx-run/rl_gsm8k\n Restrict to the 640-problem seed-0 shuffled subset via the dataset builder.\n\nStep 4, treatment:\n The same command with a custom client-side dataset builder (custom part) that serves /tmp/rx-selfgen/tasks.jsonl through the same math env and reward, with log_path=/tmp/rx-run/rl_selfgen.\n\nStep 5, eval: sample greedily from base and both save_weights_for_sampler checkpoints on all 1,319 GSM8K test problems with max_tokens=512, score exact match per item, and apply the paired decision rule.", "feasibility_check": { "status": "unknown", "problems": [ "Measured feasibility not recorded" ], "wall_seconds": null, "gpu_hours": null, "known_gpu_hours": 0, "phases": {} } } 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_4125735d30. 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 makes the verdict "inconclusive". - 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_4125735d30 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_4125735d30 and every app your fork deployed (Spindle deploys separate trainer, lora and inference apps; give them names containing run_4125735d30), 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", "verdict": "inconclusive", "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" | "inconclusive", "outcome_reason": "supported" | "refuted" | "tie" | "mixed_criteria" | "insufficient_evidence" | "incomplete_execution", "capacity_block": {"error": "", "comparison_steps_completed": , "released_clients": true | false}} Include "capacity_block" only when Spindle capacity stopped the run. A finished result must include a valid "verdict". Finishing execution alone does not support the hypothesis. 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. The research exchange saves each update, so if the session ends early the finished arms still count as partial results. Leave "verdict" out until you are done. Set "verdict" when you are done: "supported" or "refuted" only if every arm ran and the comparison judged the success criterion; "inconclusive" for a tie, mixed criteria, insufficient evidence, or incomplete comparison. Use "succeeded" when execution completed, including ties and mixed scientific outcomes; "failed" when execution stopped unsuccessfully. Optional "outcome_reason" describes the science separately from execution status: "supported"/"refuted" agree with the corresponding verdict; the other reasons require "inconclusive". "tie" means the completed comparison falls in the spec's tie band; "mixed_criteria" means some success criteria pass and others fail; "insufficient_evidence" means the evidence cannot settle the hypothesis (for example, uncertainty is too large); "incomplete_execution" means required execution or evaluation did not finish. Report the reason only with a final result. Never infer a tie or support just because execution succeeded. 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 the hypothesis as inconclusive and keeps your verdict for the changed experiment. 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. Treatment activation gate: when the spec's "activation_gate" is applicable, first run a short pilot of the treatment arm in the exact regime of the full run (same model, loss path, batch, updates per batch, lag, dtype). Measure every check's metric and record it in "activation_calibration": {"values": {"": }, "note": ""}, before any full training. Keep the spec's accepted ranges fixed. If a check is missing or outside its range, do not start full training; report the evidence and an inconclusive verdict. The exchange validates reported calibration; it cannot prevent autonomous training calls. 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 inconclusive. 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 supported/refuted only when every criterion's confidence interval settles the prespecified margin; otherwise inconclusive. 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 and report inconclusive. 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.