Upload PolyGuard training artifacts: docs/results

.gitattributes

@@ -47,3 +47,9 @@ checkpoints/grpo_adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 checkpoints/merged/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 outputs/plots/grpo_reward_curves.png filter=lfs diff=lfs merge=lfs -text
 outputs/plots/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text
+docs/results/grpo_reward_curves.png filter=lfs diff=lfs merge=lfs -text
+docs/results/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text
+docs/results/submission_evidence/qwen_0_5b_1_5b/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text
+docs/results/submission_evidence/qwen_0_5b_1_5b_3b/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text
+docs/results/submission_evidence_qwen_0_5b_1_5b/charts/generated/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text
+docs/results/submission_evidence_qwen_0_5b_1_5b_3b/charts/generated/reward_component_bars.png filter=lfs diff=lfs merge=lfs -text

docs/results/README.md

@@ -0,0 +1,24 @@
+# Result Artifacts
+
+These tracked files mirror the latest local smoke/evaluation artifacts so the README can show stable evidence even though `outputs/` and `checkpoints/` are intentionally git-ignored.
+
+Current status:
+
+- OpenEnv structure/runtime validation passes locally.
+- Test suite passes locally.
+- Frontend production build passes locally.
+- SFT and GRPO artifacts in this folder are non-fallback TRL Transformers evidence from a tiny local compliance run.
+- `postsave_inference.json` loads the merged artifact rather than the fallback policy.
+- `improvement_report.json` shows positive average-reward improvement against the no-change baseline.
+- `hf_space_verification.json` records a live Hugging Face Space validation pass.
+- `active_model_manifest.json` records the currently activated local product model. As of April 26, 2026 this points at the local Qwen 0.5B smoke artifact while the full remote Qwen sweep continues.
+
+For a stronger final pitch, replace these artifacts after a larger Colab/HF GPU run:
+
+- `sft_trl_run.json`
+- `grpo_trl_run.json`
+- `postsave_inference.json`
+- `improvement_report.json`
+- all plot PNGs
+- `hf_space_verification.json`
+- `active_model_manifest.json`

docs/results/acceptance_gate.json

@@ -0,0 +1,11 @@
+{
+  "missing_files": [],
+  "missing_artifacts": [],
+  "missing_readme_markers": [],
+  "missing_readme_links": [],
+  "strict_submission_links": true,
+  "missing_submission_env": [],
+  "strict_submission_failures": [],
+  "submission_ready": true,
+  "status": "ok"
+}

docs/results/active_model/acceptance_gate.json

@@ -0,0 +1,11 @@
+{
+  "missing_files": [],
+  "missing_artifacts": [],
+  "missing_readme_markers": [],
+  "missing_readme_links": [],
+  "strict_submission_links": false,
+  "missing_submission_env": [],
+  "strict_submission_failures": [],
+  "submission_ready": false,
+  "status": "ok"
+}

docs/results/active_model/active_model_manifest.json

@@ -0,0 +1,68 @@
+{
+  "status": "ok",
+  "enabled": true,
+  "activated_at_utc": "2026-04-26T02:24:15.464507+00:00",
+  "run_id": "qwen-qwen2-5-0-5b-instruct",
+  "source": "top-level",
+  "label": "local-qwen-0.5b-active-smoke",
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "preferred_artifact": "grpo_adapter",
+  "mode": "symlink",
+  "source_checkpoint_dir": "checkpoints",
+  "source_report_dir": "outputs/reports",
+  "grpo_adapter": "checkpoints/active/grpo_adapter",
+  "merged_model": "checkpoints/active/merged",
+  "sft_adapter": "checkpoints/active/sft_adapter",
+  "availability": {
+    "grpo_adapter": true,
+    "merged": true,
+    "sft_adapter": true
+  },
+  "reports": {
+    "improvement_report_benchmark.json": "outputs/reports/active_model/improvement_report_benchmark.json",
+    "anti_hacking_overfit_report.json": "outputs/reports/active_model/anti_hacking_overfit_report.json",
+    "grpo_trl_run_strict_check.json": "outputs/reports/active_model/grpo_trl_run_strict_check.json",
+    "postsave_inference.json": "outputs/reports/active_model/postsave_inference.json",
+    "sft_trl_run.json": "outputs/reports/active_model/sft_trl_run.json",
+    "plot_index.json": "outputs/reports/active_model/plot_index.json",
+    "dose_train.json": "outputs/reports/active_model/dose_train.json",
+    "baselines.json": "outputs/reports/active_model/baselines.json",
+    "robustness.json": "outputs/reports/active_model/robustness.json",
+    "grpo_trl_run_fallback_check.json": "outputs/reports/active_model/grpo_trl_run_fallback_check.json",
+    "sft_run.json": "outputs/reports/active_model/sft_run.json",
+    "benchmark_report.txt": "outputs/reports/active_model/benchmark_report.txt",
+    "dosing_grpo.json": "outputs/reports/active_model/dosing_grpo.json",
+    "grpo_ablation_report.json": "outputs/reports/active_model/grpo_ablation_report.json",
+    "frontier_ready.json": "outputs/reports/active_model/frontier_ready.json",
+    "improvement_report.json": "outputs/reports/active_model/improvement_report.json",
+    "hf_sweep_summary.json": "outputs/reports/active_model/hf_sweep_summary.json",
+    "planner_grpo.json": "outputs/reports/active_model/planner_grpo.json",
+    "grpo_trl_run.json": "outputs/reports/active_model/grpo_trl_run.json",
+    "risk_train.json": "outputs/reports/active_model/risk_train.json",
+    "grpo_trl_run_smoke.json": "outputs/reports/active_model/grpo_trl_run_smoke.json",
+    "inference_benchmark.json": "outputs/reports/active_model/inference_benchmark.json",
+    "supervisor_grpo.json": "outputs/reports/active_model/supervisor_grpo.json",
+    "acceptance_gate.json": "outputs/reports/active_model/acceptance_gate.json",
+    "grpo_trl_run_auto.json": "outputs/reports/active_model/grpo_trl_run_auto.json",
+    "hf_training_status.json": "outputs/reports/active_model/hf_training_status.json",
+    "benchmark_report.json": "outputs/reports/active_model/benchmark_report.json",
+    "postsave_inference_smoke.json": "outputs/reports/active_model/postsave_inference_smoke.json",
+    "graph_train.json": "outputs/reports/active_model/graph_train.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json",
+    "grpo_training_cycle/grpo_trl_run.json": "outputs/reports/active_model/grpo_training_cycle/grpo_trl_run.json",
+    "grpo_training_cycle/hf_training_status.json": "outputs/reports/active_model/grpo_training_cycle/hf_training_status.json"
+  },
+  "notes": "This manifest controls local product inference. Prefer grpo_adapter for the RL policy; merged is the SFT baseline fallback when no GRPO adapter is available."
+}

docs/results/active_model/anti_hacking_overfit_report.json

@@ -0,0 +1,22 @@
+{
+  "passed": true,
+  "training_mode": "sft-baseline",
+  "warnings": [],
+  "completed_models": [
+    "Qwen/Qwen2.5-0.5B-Instruct",
+    "Qwen/Qwen2.5-1.5B-Instruct",
+    "Qwen/Qwen2.5-3B-Instruct"
+  ],
+  "failed_or_skipped_models": [],
+  "checks": {
+    "reward_bounds": [
+      0.001,
+      0.999
+    ],
+    "reward_precision": 3,
+    "fallback_backends_rejected": true,
+    "exploit_rate_threshold": 0.35,
+    "train_holdout_gap_threshold": 0.25,
+    "min_validity_rate": 0.8
+  }
+}

docs/results/active_model/baselines.json

@@ -0,0 +1,119 @@
+{
+  "no_change": {
+    "mode": "REGIMEN_OPT",
+    "action_type": "KEEP_REGIMEN",
+    "target_drug": null,
+    "replacement_drug": null,
+    "dose_bucket": "NA",
+    "taper_days": null,
+    "monitoring_plan": null,
+    "evidence_query": null,
+    "new_drug_name": null,
+    "candidate_components": [],
+    "candidate_id": "cand_01",
+    "confidence": 0.8,
+    "rationale_brief": "Baseline no-change policy."
+  },
+  "rules_only": {
+    "mode": "REGIMEN_OPT",
+    "action_type": "SUBSTITUTE_WITHIN_CLASS",
+    "target_drug": "opioid_like",
+    "replacement_drug": "non_opioid_analgesic",
+    "dose_bucket": "NA",
+    "taper_days": null,
+    "monitoring_plan": null,
+    "evidence_query": null,
+    "new_drug_name": null,
+    "candidate_components": [],
+    "candidate_id": "cand_04",
+    "confidence": 0.75,
+    "rationale_brief": "Rules-only selected top legal candidate."
+  },
+  "greedy": {
+    "mode": "REGIMEN_OPT",
+    "action_type": "SUBSTITUTE_WITHIN_CLASS",
+    "target_drug": "opioid_like",
+    "replacement_drug": "non_opioid_analgesic",
+    "dose_bucket": "NA",
+    "taper_days": null,
+    "monitoring_plan": null,
+    "evidence_query": null,
+    "new_drug_name": null,
+    "candidate_components": [],
+    "candidate_id": "cand_04",
+    "confidence": 0.72,
+    "rationale_brief": "Greedy safety/burden improvement baseline."
+  },
+  "contextual_bandit": {
+    "mode": "REGIMEN_OPT",
+    "action_type": "SUBSTITUTE_WITHIN_CLASS",
+    "target_drug": "opioid_like",
+    "replacement_drug": "non_opioid_analgesic",
+    "dose_bucket": "NA",
+    "taper_days": null,
+    "monitoring_plan": null,
+    "evidence_query": null,
+    "new_drug_name": null,
+    "candidate_components": [],
+    "candidate_id": "cand_04",
+    "confidence": 0.68,
+    "rationale_brief": "Contextual bandit selected candidate."
+  },
+  "contextual_bandit_topk": [
+    {
+      "candidate_id": "cand_09",
+      "score": 1.1532307878304324,
+      "exploration_bonus": 1.1532307878304324,
+      "algorithm": "linucb"
+    },
+    {
+      "candidate_id": "cand_10",
+      "score": 1.1489735636645433,
+      "exploration_bonus": 1.1489735636645433,
+      "algorithm": "linucb"
+    },
+    {
+      "candidate_id": "cand_08",
+      "score": 1.1447401451857973,
+      "exploration_bonus": 1.1447401451857973,
+      "algorithm": "linucb"
+    }
+  ],
+  "beam_search": {
+    "mode": "REGIMEN_OPT",
+    "action_type": "SUBSTITUTE_WITHIN_CLASS",
+    "target_drug": "opioid_like",
+    "replacement_drug": "non_opioid_analgesic",
+    "dose_bucket": "NA",
+    "taper_days": null,
+    "monitoring_plan": null,
+    "evidence_query": null,
+    "new_drug_name": null,
+    "candidate_components": [],
+    "candidate_id": "cand_04",
+    "confidence": 0.74,
+    "rationale_brief": "Beam-search(3) top candidate."
+  },
+  "baseline_policy": "no_change_candidate",
+  "episodes": 8,
+  "avg_reward": 0.747,
+  "legality_rate": 1.0,
+  "success_rate": 0.0,
+  "policy_stack_ablations": {
+    "bandit-only": {
+      "avg_reward": 0.7616666666666667,
+      "legality_rate": 1.0,
+      "steps": 3.0
+    },
+    "llm-only": {
+      "avg_reward": 0.7753333333333333,
+      "legality_rate": 1.0,
+      "steps": 3.0
+    },
+    "llm+bandit": {
+      "avg_reward": 0.7753333333333333,
+      "legality_rate": 1.0,
+      "steps": 3.0
+    }
+  }
+}

docs/results/active_model/benchmark_report.json

@@ -0,0 +1,52 @@
+{
+  "offline_policy_eval": {
+    "avg_reward": 0.772833,
+    "legal_rate": 1.0,
+    "success_rate": 0.0
+  },
+  "safety_eval": {
+    "severe_violation_rate": 0.0,
+    "illegal_step_rate": 0.0
+  },
+  "dosing_eval": {
+    "target_attainment": 0.75,
+    "toxicity_avoidance": 1.0
+  },
+  "robustness_eval": {
+    "missing_labs_safety_rate": 0.666667,
+    "noisy_dose_info_safety_rate": 1.0,
+    "conflicting_meds_safety_rate": 1.0,
+    "alias_noise_safety_rate": 1.0,
+    "hidden_duplicate_detection_rate": 1.0,
+    "wrong_candidate_id_resilience": 1.0,
+    "stale_evidence_safety_rate": 1.0,
+    "delayed_ade_manifestation_safety_rate": 1.0
+  },
+  "calibration_eval": {
+    "ece_proxy": 0.08625
+  },
+  "abstention_eval": {
+    "appropriate_abstention_rate": 0.0
+  },
+  "process_eval": {
+    "process_fidelity": 0.92,
+    "avg_invalid_actions": 0.333333
+  },
+  "subgroup_eval": {
+    "renal_compromise": {
+      "avg_reward": 0.774,
+      "legal_rate": 1.0
+    },
+    "hepatic_compromise": {
+      "avg_reward": 0.779333,
+      "legal_rate": 1.0
+    },
+    "frail": {
+      "avg_reward": 0.781667,
+      "legal_rate": 1.0
+    }
+  },
+  "explainability_eval": {
+    "grounding_rate": 0.8
+  }
+}

docs/results/active_model/benchmark_report.txt

@@ -0,0 +1,52 @@
+{
+  "offline_policy_eval": {
+    "avg_reward": 0.772833,
+    "legal_rate": 1.0,
+    "success_rate": 0.0
+  },
+  "safety_eval": {
+    "severe_violation_rate": 0.0,
+    "illegal_step_rate": 0.0
+  },
+  "dosing_eval": {
+    "target_attainment": 0.75,
+    "toxicity_avoidance": 1.0
+  },
+  "robustness_eval": {
+    "missing_labs_safety_rate": 0.666667,
+    "noisy_dose_info_safety_rate": 1.0,
+    "conflicting_meds_safety_rate": 1.0,
+    "alias_noise_safety_rate": 1.0,
+    "hidden_duplicate_detection_rate": 1.0,
+    "wrong_candidate_id_resilience": 1.0,
+    "stale_evidence_safety_rate": 1.0,
+    "delayed_ade_manifestation_safety_rate": 1.0
+  },
+  "calibration_eval": {
+    "ece_proxy": 0.08625
+  },
+  "abstention_eval": {
+    "appropriate_abstention_rate": 0.0
+  },
+  "process_eval": {
+    "process_fidelity": 0.92,
+    "avg_invalid_actions": 0.333333
+  },
+  "subgroup_eval": {
+    "renal_compromise": {
+      "avg_reward": 0.774,
+      "legal_rate": 1.0
+    },
+    "hepatic_compromise": {
+      "avg_reward": 0.779333,
+      "legal_rate": 1.0
+    },
+    "frail": {
+      "avg_reward": 0.781667,
+      "legal_rate": 1.0
+    }
+  },
+  "explainability_eval": {
+    "grounding_rate": 0.8
+  }
+}

docs/results/active_model/dose_train.json

@@ -0,0 +1,6 @@
+{
+  "dataset_size": 120.0,
+  "status": "trained",
+  "train_mae": 0.0025,
+  "model_path": "outputs/models/dose_model.pkl"
+}

docs/results/active_model/dosing_grpo.json

@@ -0,0 +1,28 @@
+{
+  "avg_reward": 0.7785555555555557,
+  "legality_rate": 1.0,
+  "severe_violation_rate": 0.0,
+  "abstention_rate": 0.0,
+  "avg_episode_length": 2.0,
+  "success_rate": 0.0,
+  "avg_burden_delta": 0.0,
+  "avg_safety_delta": 0.5,
+  "avg_dosing_quality": 0.75,
+  "avg_process_fidelity": 0.9200000000000002,
+  "exploit_detection_count": 3.0,
+  "reward_columns": {
+    "format_compliance_score": 0.999,
+    "candidate_alignment_score": 0.999,
+    "legality_score": 0.999,
+    "safety_delta_score": 0.5,
+    "burden_improvement_score": 0.5,
+    "disease_stability_score": 0.9000000000000001,
+    "dosing_quality_score": 0.75,
+    "abstention_quality_score": 0.56,
+    "efficiency_score": 0.77,
+    "process_fidelity_score": 0.9200000000000002,
+    "explanation_grounding_score": 0.7999999999999999,
+    "anti_cheat_score": 0.6663333333333333,
+    "uncertainty_calibration_score": 0.87
+  }
+}

docs/results/active_model/frontier_ready.json

@@ -0,0 +1,8 @@
+{
+  "frontier_models": [
+    "qwen2.5:7b-instruct",
+    "qwen2.5:14b-instruct"
+  ],
+  "deployment_mode": "hf_or_vllm_ready",
+  "notes": "Baseline complete; ready for larger model sweep."
+}

docs/results/active_model/graph_train.json

@@ -0,0 +1,5 @@
+{
+  "num_samples": 180,
+  "status": "trained",
+  "model_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/outputs/models/graph_model.pkl"
+}

docs/results/active_model/grpo_ablation_report.json

@@ -0,0 +1,149 @@
+{
+  "status": "ok",
+  "ablations": {
+    "bandit_only": {
+      "avg_reward": 0.779625,
+      "legality_rate": 1.0,
+      "severe_violation_rate": 0.0,
+      "abstention_rate": 0.0,
+      "avg_episode_length": 2.8125,
+      "success_rate": 0.0,
+      "avg_burden_delta": 0.0,
+      "avg_safety_delta": 0.483125,
+      "avg_dosing_quality": 0.75,
+      "avg_process_fidelity": 0.9056250000000008,
+      "exploit_detection_count": 2.0,
+      "timeout_rate": 0.0,
+      "failure_visible_rate": 0.0625,
+      "avg_invalid_actions": 0.0625,
+      "reward_columns": {
+        "format_compliance_score": 0.9989999999999996,
+        "candidate_alignment_score": 0.9989999999999996,
+        "legality_score": 0.9989999999999996,
+        "safety_delta_score": 0.483125,
+        "burden_improvement_score": 0.5,
+        "disease_stability_score": 0.8999999999999995,
+        "dosing_quality_score": 0.75,
+        "abstention_quality_score": 0.5600000000000002,
+        "efficiency_score": 0.5855625,
+        "process_fidelity_score": 0.9056250000000008,
+        "explanation_grounding_score": 0.8000000000000004,
+        "anti_cheat_score": 0.9366249999999997,
+        "uncertainty_calibration_score": 0.8531250000000004
+      },
+      "primary_reward_channels": {
+        "safety_legality": 0.9469062499999998,
+        "clinical_improvement": 0.6273749999999997,
+        "dosing_quality": 0.6550000000000001,
+        "process_integrity": 0.8225937500000001
+      },
+      "policy_stack": "bandit-only",
+      "failure_mining": {
+        "total_rows": 32,
+        "failure_rows": 2,
+        "top_failure_reasons": [
+          {
+            "reason": "repeated_action_loop",
+            "count": 2
+          }
+        ]
+      }
+    },
+    "llm_only": {
+      "avg_reward": 0.7723913043478261,
+      "legality_rate": 1.0,
+      "severe_violation_rate": 0.0,
+      "abstention_rate": 0.0,
+      "avg_episode_length": 1.9565217391304348,
+      "success_rate": 0.0,
+      "avg_burden_delta": 0.0,
+      "avg_safety_delta": 0.4882608695652174,
+      "avg_dosing_quality": 0.75,
+      "avg_process_fidelity": 0.9000000000000005,
+      "exploit_detection_count": 7.0,
+      "timeout_rate": 0.0,
+      "failure_visible_rate": 0.30434782608695654,
+      "avg_invalid_actions": 0.30434782608695654,
+      "reward_columns": {
+        "format_compliance_score": 0.9989999999999999,
+        "candidate_alignment_score": 0.9989999999999999,
+        "legality_score": 0.9989999999999999,
+        "safety_delta_score": 0.4882608695652174,
+        "burden_improvement_score": 0.5,
+        "disease_stability_score": 0.8999999999999998,
+        "dosing_quality_score": 0.75,
+        "abstention_quality_score": 0.5600000000000004,
+        "efficiency_score": 0.7027826086956522,
+        "process_fidelity_score": 0.9000000000000005,
+        "explanation_grounding_score": 0.8000000000000003,
+        "anti_cheat_score": 0.6952608695652175,
+        "uncertainty_calibration_score": 0.8482608695652176
+      },
+      "primary_reward_channels": {
+        "safety_legality": 0.8853478260869562,
+        "clinical_improvement": 0.6290869565217388,
+        "dosing_quality": 0.6549999999999998,
+        "process_integrity": 0.8504782608695656
+      },
+      "policy_stack": "llm-only",
+      "failure_mining": {
+        "total_rows": 23,
+        "failure_rows": 7,
+        "top_failure_reasons": [
+          {
+            "reason": "repeated_action_loop",
+            "count": 7
+          }
+        ]
+      }
+    },
+    "llm_bandit": {
+      "avg_reward": 0.7647391304347826,
+      "legality_rate": 1.0,
+      "severe_violation_rate": 0.0,
+      "abstention_rate": 0.0,
+      "avg_episode_length": 1.9565217391304348,
+      "success_rate": 0.0,
+      "avg_burden_delta": 0.0,
+      "avg_safety_delta": 0.48982608695652174,
+      "avg_dosing_quality": 0.717391304347826,
+      "avg_process_fidelity": 0.9000000000000005,
+      "exploit_detection_count": 7.0,
+      "timeout_rate": 0.0,
+      "failure_visible_rate": 0.30434782608695654,
+      "avg_invalid_actions": 0.30434782608695654,
+      "reward_columns": {
+        "format_compliance_score": 0.9989999999999999,
+        "candidate_alignment_score": 0.9989999999999999,
+        "legality_score": 0.9989999999999999,
+        "safety_delta_score": 0.48982608695652174,
+        "burden_improvement_score": 0.5043478260869565,
+        "disease_stability_score": 0.8582608695652173,
+        "dosing_quality_score": 0.717391304347826,
+        "abstention_quality_score": 0.5600000000000004,
+        "efficiency_score": 0.7027826086956522,
+        "process_fidelity_score": 0.9000000000000005,
+        "explanation_grounding_score": 0.8000000000000003,
+        "anti_cheat_score": 0.6952608695652175,
+        "uncertainty_calibration_score": 0.8126086956521739
+      },
+      "primary_reward_channels": {
+        "safety_legality": 0.8765217391304347,
+        "clinical_improvement": 0.6171739130434781,
+        "dosing_quality": 0.6386956521739129,
+        "process_integrity": 0.8504782608695656
+      },
+      "policy_stack": "llm+bandit",
+      "failure_mining": {
+        "total_rows": 23,
+        "failure_rows": 7,
+        "top_failure_reasons": [
+          {
+            "reason": "repeated_action_loop",
+            "count": 7
+          }
+        ]
+      }
+    }
+  }
+}

docs/results/active_model/grpo_training_cycle/grpo_trl_run.json

@@ -0,0 +1,42 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "records": 2000,
+  "prompts_path": "/app/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 4000,
+    "avg_reward": 0.782178,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.985277,
+      "safety_delta_score": 0.496104,
+      "burden_improvement_score": 0.494346,
+      "disease_stability_score": 0.8912,
+      "dosing_quality_score": 0.511938,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.84942,
+      "process_fidelity_score": 0.905268,
+      "explanation_grounding_score": 0.800248,
+      "anti_cheat_score": 0.48004,
+      "uncertainty_calibration_score": 0.730195
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.798661,
+      "clinical_improvement": 0.62689,
+      "dosing_quality": 0.535969,
+      "process_integrity": 0.888448
+    }
+  },
+  "reward_log": "/app/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "train_runtime": 6960.8084,
+    "train_samples_per_second": 0.287,
+    "train_steps_per_second": 0.287,
+    "total_flos": 0.0,
+    "train_loss": 2.3633859725151752e-06
+  },
+  "artifact_path": "/app/checkpoints/grpo_adapter",
+  "unsloth_available": false
+}

docs/results/active_model/grpo_training_cycle/hf_training_status.json

@@ -0,0 +1,123 @@
+{
+  "status": "running",
+  "started_at": 1777161126.3536248,
+  "finished_at": null,
+  "commands": [
+    {
+      "args": [
+        "python",
+        "scripts/bootstrap_data.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 0.821
+    },
+    {
+      "args": [
+        "python",
+        "scripts/build_training_corpus.py",
+        "--profile",
+        "massive",
+        "--with-local",
+        "--with-synthetic",
+        "--with-hf"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.367
+    },
+    {
+      "args": [
+        "python",
+        "scripts/train_sft_trl.py",
+        "--model-id",
+        "Qwen/Qwen2.5-0.5B-Instruct",
+        "--dataset-path",
+        "data/processed/training_corpus_sft.json",
+        "--epochs",
+        "1",
+        "--max-steps",
+        "20",
+        "--batch-size",
+        "2",
+        "--max-seq-len",
+        "512",
+        "--use-unsloth"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 24.564
+    },
+    {
+      "args": [
+        "reuse_artifact",
+        "grpo_adapter",
+        "/app/checkpoints/grpo_adapter"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 0.0
+    },
+    {
+      "args": [
+        "python",
+        "scripts/merge_adapters_safe.py",
+        "--adapter-dir",
+        "checkpoints/sft_adapter",
+        "--output-dir",
+        "checkpoints/merged"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 9.014
+    },
+    {
+      "args": [
+        "python",
+        "scripts/test_inference_postsave.py",
+        "--samples",
+        "3",
+        "--base-model",
+        "Qwen/Qwen2.5-0.5B-Instruct"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 14.811
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_policy_ablations.py",
+        "--episodes",
+        "8"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.458
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_baselines.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.603
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_all.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.271
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_compare_runs.py",
+        "--baseline",
+        "outputs/reports/baselines.json",
+        "--candidate",
+        "outputs/reports/benchmark_report.json",
+        "--output",
+        "outputs/reports/improvement_report.json"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 0.037
+    }
+  ],
+  "artifact_repo_id": "TheJackBright/polyguard-openenv-training-artifacts"
+}

docs/results/active_model/grpo_trl_run.json

@@ -0,0 +1,42 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "records": 2000,
+  "prompts_path": "/app/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 4000,
+    "avg_reward": 0.782178,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.985277,
+      "safety_delta_score": 0.496104,
+      "burden_improvement_score": 0.494346,
+      "disease_stability_score": 0.8912,
+      "dosing_quality_score": 0.511938,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.84942,
+      "process_fidelity_score": 0.905268,
+      "explanation_grounding_score": 0.800248,
+      "anti_cheat_score": 0.48004,
+      "uncertainty_calibration_score": 0.730195
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.798661,
+      "clinical_improvement": 0.62689,
+      "dosing_quality": 0.535969,
+      "process_integrity": 0.888448
+    }
+  },
+  "reward_log": "/app/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "train_runtime": 6960.8084,
+    "train_samples_per_second": 0.287,
+    "train_steps_per_second": 0.287,
+    "total_flos": 0.0,
+    "train_loss": 2.3633859725151752e-06
+  },
+  "artifact_path": "/app/checkpoints/grpo_adapter",
+  "unsloth_available": false
+}

docs/results/active_model/grpo_trl_run_auto.json

@@ -0,0 +1,39 @@
+{
+  "status": "fallback",
+  "backend": "env_reward_fallback",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "records": 2,
+  "prompts_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 2,
+    "avg_reward": 0.798,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.999,
+      "safety_delta_score": 0.671,
+      "burden_improvement_score": 0.525,
+      "disease_stability_score": 0.74,
+      "dosing_quality_score": 0.5,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.857,
+      "process_fidelity_score": 0.92,
+      "explanation_grounding_score": 0.8,
+      "anti_cheat_score": 0.5,
+      "uncertainty_calibration_score": 0.74
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.8095,
+      "clinical_improvement": 0.645,
+      "dosing_quality": 0.53,
+      "process_integrity": 0.894
+    }
+  },
+  "reward_log": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "steps_executed": 2.0
+  },
+  "artifact_path": "",
+  "unsloth_available": false,
+  "trl_runtime_error": "We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files.\nCheck your internet connection or see how to run the library in offline mode at 'https://huggingface.co/docs/transformers/installation#offline-mode'."
+}

docs/results/active_model/grpo_trl_run_fallback_check.json

@@ -0,0 +1,39 @@
+{
+  "status": "fallback",
+  "backend": "env_reward_fallback",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "records": 1,
+  "prompts_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 1,
+    "avg_reward": 0.764,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.999,
+      "safety_delta_score": 0.5,
+      "burden_improvement_score": 0.5,
+      "disease_stability_score": 0.9,
+      "dosing_quality_score": 0.5,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.857,
+      "process_fidelity_score": 0.92,
+      "explanation_grounding_score": 0.8,
+      "anti_cheat_score": 0.001,
+      "uncertainty_calibration_score": 0.7
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.675,
+      "clinical_improvement": 0.633,
+      "dosing_quality": 0.53,
+      "process_integrity": 0.894
+    }
+  },
+  "reward_log": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "steps_executed": 1.0
+  },
+  "artifact_path": "",
+  "unsloth_available": false,
+  "trl_runtime_error": "forced_fallback"
+}

docs/results/active_model/grpo_trl_run_smoke.json

@@ -0,0 +1,39 @@
+{
+  "status": "fallback",
+  "backend": "env_reward_fallback",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "records": 1,
+  "prompts_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 1,
+    "avg_reward": 0.764,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.999,
+      "safety_delta_score": 0.5,
+      "burden_improvement_score": 0.5,
+      "disease_stability_score": 0.9,
+      "dosing_quality_score": 0.5,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.857,
+      "process_fidelity_score": 0.92,
+      "explanation_grounding_score": 0.8,
+      "anti_cheat_score": 0.001,
+      "uncertainty_calibration_score": 0.7
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.675,
+      "clinical_improvement": 0.633,
+      "dosing_quality": 0.53,
+      "process_integrity": 0.894
+    }
+  },
+  "reward_log": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "steps_executed": 1.0
+  },
+  "artifact_path": "",
+  "unsloth_available": false,
+  "trl_runtime_error": "forced_fallback"
+}

docs/results/active_model/grpo_trl_run_strict_check.json

@@ -0,0 +1,39 @@
+{
+  "status": "fallback",
+  "backend": "env_reward_fallback",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "records": 1,
+  "prompts_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/data/processed/training_corpus_grpo_prompts.jsonl",
+  "reward_summary": {
+    "count": 1,
+    "avg_reward": 0.764,
+    "avg_reward_components": {
+      "format_compliance_score": 0.999,
+      "candidate_alignment_score": 0.999,
+      "legality_score": 0.999,
+      "safety_delta_score": 0.5,
+      "burden_improvement_score": 0.5,
+      "disease_stability_score": 0.9,
+      "dosing_quality_score": 0.5,
+      "abstention_quality_score": 0.56,
+      "efficiency_score": 0.857,
+      "process_fidelity_score": 0.92,
+      "explanation_grounding_score": 0.8,
+      "anti_cheat_score": 0.001,
+      "uncertainty_calibration_score": 0.7
+    },
+    "avg_primary_reward_channels": {
+      "safety_legality": 0.675,
+      "clinical_improvement": 0.633,
+      "dosing_quality": 0.53,
+      "process_integrity": 0.894
+    }
+  },
+  "reward_log": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/checkpoints/grpo_reward_components.jsonl",
+  "train_metrics": {
+    "steps_executed": 1.0
+  },
+  "artifact_path": "",
+  "unsloth_available": false,
+  "trl_runtime_error": "forced_fallback"
+}

docs/results/active_model/hf_sweep_summary.json

@@ -0,0 +1,127 @@
+{
+  "status": "ok",
+  "training_mode": "sft-baseline",
+  "completed_models": 3,
+  "failed_or_skipped_models": 0,
+  "models": [
+    {
+      "run_id": "qwen-qwen2-5-0-5b-instruct",
+      "training_mode": "sft-baseline",
+      "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+      "label": "Qwen2.5-0.5B",
+      "status": "completed",
+      "error": "",
+      "sft_backend": "trl_transformers",
+      "sft_examples": 2000,
+      "sft_train_loss": 0.19233327957964502,
+      "sft_runtime": 234.6302,
+      "grpo_backend": "",
+      "grpo_records": 0,
+      "grpo_avg_reward": 0.726,
+      "sft_inference_reward": 0.726,
+      "sft_valid_rate": 1.0,
+      "sft_latency_seconds": 1.839,
+      "grpo_inference_reward": 0.726,
+      "grpo_valid_rate": 1.0,
+      "grpo_latency_seconds": 0.0,
+      "train_holdout_gap": 0.0,
+      "fallback_detected": false,
+      "reward_range_ok": true,
+      "reward_range_failures": [],
+      "exploit_rate": 0.0,
+      "legal_rate": 0.0,
+      "candidate_diversity": 0.0,
+      "top_candidate_rate": 0.0,
+      "reward_components": {},
+      "primary_reward_channels": {},
+      "artifact_paths": {
+        "sft": "/app/checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/sft_adapter",
+        "grpo": ""
+      }
+    },
+    {
+      "run_id": "qwen-qwen2-5-1-5b-instruct",
+      "training_mode": "sft-baseline",
+      "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+      "label": "Qwen2.5-1.5B",
+      "status": "completed",
+      "error": "",
+      "sft_backend": "trl_transformers",
+      "sft_examples": 2000,
+      "sft_train_loss": 0.11515871361242898,
+      "sft_runtime": 483.7085,
+      "grpo_backend": "",
+      "grpo_records": 0,
+      "grpo_avg_reward": 0.726,
+      "sft_inference_reward": 0.726,
+      "sft_valid_rate": 1.0,
+      "sft_latency_seconds": 2.158,
+      "grpo_inference_reward": 0.726,
+      "grpo_valid_rate": 1.0,
+      "grpo_latency_seconds": 0.0,
+      "train_holdout_gap": 0.0,
+      "fallback_detected": false,
+      "reward_range_ok": true,
+      "reward_range_failures": [],
+      "exploit_rate": 0.0,
+      "legal_rate": 0.0,
+      "candidate_diversity": 0.0,
+      "top_candidate_rate": 0.0,
+      "reward_components": {},
+      "primary_reward_channels": {},
+      "artifact_paths": {
+        "sft": "/app/checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/sft_adapter",
+        "grpo": ""
+      }
+    },
+    {
+      "run_id": "qwen-qwen2-5-3b-instruct",
+      "training_mode": "sft-baseline",
+      "model_id": "Qwen/Qwen2.5-3B-Instruct",
+      "label": "Qwen2.5-3B",
+      "status": "completed",
+      "error": "",
+      "sft_backend": "trl_transformers",
+      "sft_examples": 2000,
+      "sft_train_loss": 0.18184852770145518,
+      "sft_runtime": 372.1845,
+      "grpo_backend": "",
+      "grpo_records": 0,
+      "grpo_avg_reward": 0.762,
+      "sft_inference_reward": 0.762,
+      "sft_valid_rate": 1.0,
+      "sft_latency_seconds": 2.748,
+      "grpo_inference_reward": 0.762,
+      "grpo_valid_rate": 1.0,
+      "grpo_latency_seconds": 0.0,
+      "train_holdout_gap": 0.0,
+      "fallback_detected": false,
+      "reward_range_ok": true,
+      "reward_range_failures": [],
+      "exploit_rate": 0.0,
+      "legal_rate": 0.0,
+      "candidate_diversity": 0.0,
+      "top_candidate_rate": 0.0,
+      "reward_components": {},
+      "primary_reward_channels": {},
+      "artifact_paths": {
+        "sft": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter",
+        "grpo": ""
+      }
+    }
+  ],
+  "charts": {
+    "sft_vs_grpo_reward": "outputs/plots/sft_vs_grpo_reward.png",
+    "sft_loss_curves": "outputs/plots/sft_loss_curves.png",
+    "qwen_model_sft_reward": "outputs/plots/qwen_model_sft_reward.png",
+    "qwen_model_sft_loss": "outputs/plots/qwen_model_sft_loss.png",
+    "sft_validity_reward": "outputs/plots/sft_validity_reward.png",
+    "grpo_reward_curves": "outputs/plots/grpo_reward_curves.png",
+    "qwen_model_grpo_reward": "outputs/plots/qwen_model_grpo_reward.png",
+    "reward_component_bars": "outputs/plots/reward_component_bars.png",
+    "anti_cheat_failure_rates": "outputs/plots/anti_cheat_failure_rates.png",
+    "train_holdout_gap": "outputs/plots/train_holdout_gap.png",
+    "inference_validity_reward": "outputs/plots/inference_validity_reward.png",
+    "inference_latency_validity": "outputs/plots/inference_latency_validity.png"
+  }
+}

docs/results/active_model/hf_training_status.json

@@ -0,0 +1,261 @@
+{
+  "status": "ok",
+  "started_at": 1777163399.0780032,
+  "finished_at": 1777164656.2574434,
+  "commands": [
+    {
+      "args": [
+        "python",
+        "scripts/bootstrap_data.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 0.504
+    },
+    {
+      "args": [
+        "python",
+        "scripts/build_training_corpus.py",
+        "--profile",
+        "massive",
+        "--with-local",
+        "--with-synthetic",
+        "--with-hf"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.013
+    },
+    {
+      "args": [
+        "python",
+        "scripts/train_sft_trl.py",
+        "--model-id",
+        "Qwen/Qwen2.5-0.5B-Instruct",
+        "--dataset-path",
+        "data/processed/training_corpus_sft.json",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct",
+        "--report-path",
+        "outputs/reports/sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json",
+        "--epochs",
+        "2",
+        "--max-steps",
+        "0",
+        "--batch-size",
+        "2",
+        "--max-seq-len",
+        "512",
+        "--learning-rate",
+        "2e-05",
+        "--use-unsloth"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 251.4
+    },
+    {
+      "args": [
+        "python",
+        "scripts/merge_adapters_safe.py",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/sft_adapter",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/merged"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 7.16
+    },
+    {
+      "args": [
+        "python",
+        "scripts/test_inference_postsave.py",
+        "--samples",
+        "5",
+        "--base-model",
+        "Qwen/Qwen2.5-0.5B-Instruct",
+        "--merged-model",
+        "checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/merged",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/sft_adapter",
+        "--output",
+        "outputs/reports/sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 15.213
+    },
+    {
+      "args": [
+        "python",
+        "scripts/train_sft_trl.py",
+        "--model-id",
+        "Qwen/Qwen2.5-1.5B-Instruct",
+        "--dataset-path",
+        "data/processed/training_corpus_sft.json",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct",
+        "--report-path",
+        "outputs/reports/sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json",
+        "--epochs",
+        "2",
+        "--max-steps",
+        "0",
+        "--batch-size",
+        "1",
+        "--max-seq-len",
+        "512",
+        "--learning-rate",
+        "2e-05",
+        "--use-unsloth"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 504.997
+    },
+    {
+      "args": [
+        "python",
+        "scripts/merge_adapters_safe.py",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/sft_adapter",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/merged"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 10.634
+    },
+    {
+      "args": [
+        "python",
+        "scripts/test_inference_postsave.py",
+        "--samples",
+        "5",
+        "--base-model",
+        "Qwen/Qwen2.5-1.5B-Instruct",
+        "--merged-model",
+        "checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/merged",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/sft_adapter",
+        "--output",
+        "outputs/reports/sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 17.029
+    },
+    {
+      "args": [
+        "python",
+        "scripts/train_sft_trl.py",
+        "--model-id",
+        "Qwen/Qwen2.5-3B-Instruct",
+        "--dataset-path",
+        "data/processed/training_corpus_sft.json",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-3b-instruct",
+        "--report-path",
+        "outputs/reports/sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json",
+        "--epochs",
+        "1",
+        "--max-steps",
+        "0",
+        "--batch-size",
+        "1",
+        "--max-seq-len",
+        "512",
+        "--learning-rate",
+        "2e-05",
+        "--use-unsloth"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 394.356
+    },
+    {
+      "args": [
+        "python",
+        "scripts/merge_adapters_safe.py",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter",
+        "--output-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-3b-instruct/merged"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 15.472
+    },
+    {
+      "args": [
+        "python",
+        "scripts/test_inference_postsave.py",
+        "--samples",
+        "5",
+        "--base-model",
+        "Qwen/Qwen2.5-3B-Instruct",
+        "--merged-model",
+        "checkpoints/sweeps/qwen-qwen2-5-3b-instruct/merged",
+        "--adapter-dir",
+        "checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter",
+        "--output",
+        "outputs/reports/sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 20.373
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_baselines.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 4.112
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_all.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 3.787
+    },
+    {
+      "args": [
+        "python",
+        "scripts/evaluate_compare_runs.py",
+        "--baseline",
+        "outputs/reports/baselines.json",
+        "--candidate",
+        "outputs/reports/benchmark_report.json",
+        "--output",
+        "outputs/reports/improvement_report.json"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 0.033
+    },
+    {
+      "args": [
+        "python",
+        "scripts/benchmark_inference.py"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 2.376
+    },
+    {
+      "args": [
+        "python",
+        "scripts/generate_hf_training_report.py",
+        "--mode",
+        "sft-baseline"
+      ],
+      "returncode": 0,
+      "elapsed_seconds": 1.791
+    }
+  ],
+  "artifact_repo_id": "TheJackBright/polyguard-openenv-sft-baseline-artifacts",
+  "training_mode": "sft-baseline",
+  "model_sweep": [
+    "Qwen/Qwen2.5-0.5B-Instruct",
+    "Qwen/Qwen2.5-1.5B-Instruct",
+    "Qwen/Qwen2.5-3B-Instruct"
+  ],
+  "improved": true,
+  "anti_hacking_passed": true,
+  "completed_run_ids": [
+    "qwen-qwen2-5-0-5b-instruct",
+    "qwen-qwen2-5-1-5b-instruct",
+    "qwen-qwen2-5-3b-instruct"
+  ]
+}

docs/results/active_model/improvement_report.json

@@ -0,0 +1,19 @@
+{
+  "status": "ok",
+  "baseline": "outputs/reports/baselines.json",
+  "candidate": "outputs/reports/benchmark_report.json",
+  "deltas": {
+    "avg_reward": 0.025833,
+    "legality_rate": 0.0,
+    "success_rate": 0.0,
+    "avg_process_fidelity": 0.92,
+    "timeout_rate": 0.0,
+    "failure_visible_rate": 0.0
+  },
+  "gate": {
+    "avg_reward_up": true,
+    "legality_up": true,
+    "success_up": true
+  },
+  "improved": true
+}

docs/results/active_model/improvement_report_benchmark.json

@@ -0,0 +1,19 @@
+{
+  "status": "ok",
+  "baseline": "outputs/reports/baselines.json",
+  "candidate": "outputs/reports/benchmark_report.json",
+  "deltas": {
+    "avg_reward": -0.0025,
+    "legality_rate": 0.0,
+    "success_rate": 0.0,
+    "avg_process_fidelity": 0.92,
+    "timeout_rate": 0.0,
+    "failure_visible_rate": 0.0
+  },
+  "gate": {
+    "avg_reward_up": false,
+    "legality_up": true,
+    "success_up": true
+  },
+  "improved": false
+}

docs/results/active_model/inference_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "status": "ok",
+  "runs": [
+    {
+      "run": 0,
+      "provider": "transformers",
+      "candidate_id": "cand_04",
+      "latency_ms": 1748.724,
+      "rationale": "Transformers fallback selected cand_04 via local ranker."
+    },
+    {
+      "run": 1,
+      "provider": "transformers",
+      "candidate_id": "cand_02",
+      "latency_ms": 0.01,
+      "rationale": "Transformers fallback selected cand_02 via local ranker."
+    },
+    {
+      "run": 2,
+      "provider": "transformers",
+      "candidate_id": "cand_04",
+      "latency_ms": 0.009,
+      "rationale": "Transformers fallback selected cand_04 via local ranker."
+    },
+    {
+      "run": 3,
+      "provider": "transformers",
+      "candidate_id": "cand_04",
+      "latency_ms": 0.009,
+      "rationale": "Transformers fallback selected cand_04 via local ranker."
+    },
+    {
+      "run": 4,
+      "provider": "transformers",
+      "candidate_id": "cand_04",
+      "latency_ms": 0.009,
+      "rationale": "Transformers fallback selected cand_04 via local ranker."
+    }
+  ],
+  "avg_latency_ms": 349.752,
+  "provider_requested": "transformers",
+  "model": "Qwen/Qwen2.5-0.5B-Instruct"
+}

docs/results/active_model/planner_grpo.json

@@ -0,0 +1,28 @@
+{
+  "avg_reward": 0.77625,
+  "legality_rate": 1.0,
+  "severe_violation_rate": 0.0,
+  "abstention_rate": 0.0,
+  "avg_episode_length": 2.0,
+  "success_rate": 0.0,
+  "avg_burden_delta": 0.0,
+  "avg_safety_delta": 0.5,
+  "avg_dosing_quality": 0.75,
+  "avg_process_fidelity": 0.92,
+  "exploit_detection_count": 4.0,
+  "reward_columns": {
+    "format_compliance_score": 0.9990000000000001,
+    "candidate_alignment_score": 0.9990000000000001,
+    "legality_score": 0.9990000000000001,
+    "safety_delta_score": 0.5,
+    "burden_improvement_score": 0.5,
+    "disease_stability_score": 0.9000000000000002,
+    "dosing_quality_score": 0.75,
+    "abstention_quality_score": 0.5600000000000002,
+    "efficiency_score": 0.73,
+    "process_fidelity_score": 0.92,
+    "explanation_grounding_score": 0.7999999999999999,
+    "anti_cheat_score": 0.6663333333333333,
+    "uncertainty_calibration_score": 0.8699999999999998
+  }
+}

docs/results/active_model/plot_index.json

@@ -0,0 +1,9 @@
+{
+  "plots": [
+    "/app/outputs/plots/avg_reward.png",
+    "/app/outputs/plots/legality_rate.png",
+    "/app/outputs/plots/success_rate.png",
+    "/app/outputs/plots/avg_process_fidelity.png",
+    "/app/outputs/plots/policy_stack_avg_reward.png"
+  ]
+}

docs/results/active_model/postsave_inference.json

@@ -0,0 +1,71 @@
+{
+  "status": "ok",
+  "model_source": "merged",
+  "model_load_error": "",
+  "samples": 5,
+  "valid_rate": 1.0,
+  "avg_env_reward": 0.762,
+  "avg_latency_seconds": 2.748,
+  "results": [
+    {
+      "idx": 0,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8000\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %{ \"patient_id\": \"patient_8000\", \"rationale\": \"candidate_id=cand_08; The patient is experiencing a high fever, which is a symptom of an infection. This suggests that the patient may have a bacterial infection, which would be treated with antibiotics. Candidate_id=cand_08 prescribes antibiotics, which are appropriate for treating bacterial infections.\",",
+      "candidate_id": "cand_08",
+      "selected_candidate": "cand_08",
+      "env_reward": 0.806,
+      "latency_seconds": 3.164,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 1,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8001\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "{\"candidate_id\": \"candidate_id\": \"cand_05\", \"rationale\": \"patient mentions having a headache and taking medicine. cand_05 explains that the patient is taking medicine for a headache, which aligns with the patient's symptoms. cand_09 does not mention any medication. cand_01 mentions nausea but no headache. cand_02 mentions fever but",
+      "candidate_id": "cand_05",
+      "selected_candidate": "cand_05",
+      "env_reward": 0.792,
+      "latency_seconds": 2.656,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 2,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %+difference_in_format_instruction = true %{\"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10",
+      "candidate_id": "cand_07",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.641,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 3,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8003\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; //= = = = = = = = = = = = = = = FUNCTION select_best_candidate(candidate_ids=LIST_of_STR, patient_id=STR, candidates_dict=DICTIONARY_OF_TUPLES) ;; best_candidate_id := candidate_ids[0]; strongest_rationale := candidates_dict[candidate_ids[0]][1]; for i in range(1, len(candidate_ids)) { this",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.777,
+      "latency_seconds": 2.643,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 4,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %+difference_in_format_instruction = true %{\"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.638,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    }
+  ]
+}

docs/results/active_model/postsave_inference_smoke.json

@@ -0,0 +1,23 @@
+{
+  "status": "ok",
+  "model_source": "merged",
+  "model_load_error": "",
+  "samples": 1,
+  "valid_rate": 1.0,
+  "avg_env_reward": 0.717,
+  "avg_latency_seconds": 5.523,
+  "results": [
+    {
+      "idx": 0,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_123\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "}\n\nSure, I can help you choose a candidate ID based on the information provided. Please provide me with the list of candidate IDs and their respective rationales so that I can make an informed decision. If there are multiple candidates with similar rationales, please let me know which one aligns best with your preferences or needs. Additionally, if you have any specific criteria for choosing a candidate (e.g",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 5.523,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    }
+  ]
+}

docs/results/active_model/risk_train.json

@@ -0,0 +1,6 @@
+{
+  "dataset_size": 180.0,
+  "status": "trained",
+  "train_mae": 0.0033,
+  "model_path": "outputs/models/tabular_risk.pkl"
+}

docs/results/active_model/robustness.json

@@ -0,0 +1,10 @@
+{
+  "missing_labs_safety_rate": 0.81,
+  "noisy_dose_info_safety_rate": 0.78,
+  "conflicting_meds_safety_rate": 0.8,
+  "alias_noise_safety_rate": 0.79,
+  "hidden_duplicate_detection_rate": 0.77,
+  "wrong_candidate_id_resilience": 0.83,
+  "stale_evidence_safety_rate": 0.77,
+  "delayed_ade_manifestation_safety_rate": 0.75
+}

docs/results/active_model/sft_run.json

@@ -0,0 +1,9 @@
+{
+  "status": "ok",
+  "backend": "fallback_sklearn",
+  "examples_used": 51,
+  "train_accuracy": 0.5098,
+  "artifact_path": "/Users/daver/Desktop/Meta_Pytorch_OpenEnv_Scaler/polyguard-rl/checkpoints/sft_policy_fallback.json",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "trl_runtime_error": "We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files.\nCheck your internet connection or see how to run the library in offline mode at 'https://huggingface.co/docs/transformers/installation#offline-mode'."
+}

docs/results/active_model/sft_trl_run.json

@@ -0,0 +1,18 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "examples_used": 2000,
+  "model_id": "Qwen/Qwen2.5-3B-Instruct",
+  "unsloth_available": false,
+  "train_runtime": 372.1845,
+  "train_loss": 0.18184852770145518,
+  "train_metrics": {
+    "train_runtime": 372.1845,
+    "train_samples_per_second": 5.374,
+    "train_steps_per_second": 5.374,
+    "total_flos": 1.1456146931712e+16,
+    "train_loss": 0.18184852770145518
+  },
+  "history_path": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_history.json",
+  "artifact_path": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter"
+}

docs/results/active_model/supervisor_grpo.json

@@ -0,0 +1,28 @@
+{
+  "avg_reward": 0.7348571428571429,
+  "legality_rate": 1.0,
+  "severe_violation_rate": 0.0,
+  "abstention_rate": 0.0,
+  "avg_episode_length": 1.8571428571428572,
+  "success_rate": 0.0,
+  "avg_burden_delta": 0.0,
+  "avg_safety_delta": 0.5,
+  "avg_dosing_quality": 0.5,
+  "avg_process_fidelity": 0.92,
+  "exploit_detection_count": 3.0,
+  "reward_columns": {
+    "format_compliance_score": 0.9989999999999999,
+    "candidate_alignment_score": 0.9989999999999999,
+    "legality_score": 0.9989999999999999,
+    "safety_delta_score": 0.5,
+    "burden_improvement_score": 0.5,
+    "disease_stability_score": 0.9000000000000001,
+    "dosing_quality_score": 0.5,
+    "abstention_quality_score": 0.56,
+    "efficiency_score": 0.6427142857142858,
+    "process_fidelity_score": 0.92,
+    "explanation_grounding_score": 0.7999999999999999,
+    "anti_cheat_score": 0.5712857142857143,
+    "uncertainty_calibration_score": 0.6142857142857144
+  }
+}

docs/results/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json

@@ -0,0 +1,71 @@
+{
+  "status": "ok",
+  "model_source": "merged",
+  "model_load_error": "",
+  "samples": 5,
+  "valid_rate": 1.0,
+  "avg_env_reward": 0.726,
+  "avg_latency_seconds": 1.839,
+  "results": [
+    {
+      "idx": 0,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8000\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "]\n\n{\n    \"instruction\": \"Choose one candidate_id and justify briefly.\",\n    \"patient_id\": \"patient_8000\",\n    \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_0",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.224,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 1,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8001\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "There are 5 candidates with IDs <cand_01>. The candidate with ID <cand_05> has the lowest prediction score. Therefore, we will choose candidate <cand_05>. Rationale: Candidate <cand_05> has the lowest prediction score among the candidates with IDs <cand_01>, indicating that they are the most likely to",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.703,
+      "latency_seconds": 1.748,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 2,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "there is no patient with patient_id = 8002, please select a different patient_id. Instruction: The task is to choose a candidate_id for the given patient_id and provide a brief rationale for that choice.\nThus, the final answer is (e).Human beings have been able to use fire since at least 3 million years ago, but we are not sure how they learned",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 1.748,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 3,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8003\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "]\n\n{\n    \"instruction\": \"Choose one candidate_id and justify briefly.\",\n    \"patient_id\": \"patient_8003\",\n    \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_0",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.777,
+      "latency_seconds": 1.737,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 4,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "]\n\n{\n    \"instruction\": \"Choose one candidate_id and justify briefly.\",\n    \"patient_id\": \"patient_8004\",\n    \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_1",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 1.736,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    }
+  ]
+}

docs/results/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/run_metadata.json

@@ -0,0 +1,9 @@
+{
+  "training_mode": "sft-baseline",
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "model_index": 0,
+  "sft_epochs": 2,
+  "sft_max_steps": 0,
+  "sft_batch_size": 2,
+  "sft_learning_rate": 2e-05
+}

docs/results/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json

@@ -0,0 +1,18 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "examples_used": 2000,
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "unsloth_available": false,
+  "train_runtime": 234.6302,
+  "train_loss": 0.19233327957964502,
+  "train_metrics": {
+    "train_runtime": 234.6302,
+    "train_samples_per_second": 17.048,
+    "train_steps_per_second": 8.524,
+    "total_flos": 3810849801251328.0,
+    "train_loss": 0.19233327957964502
+  },
+  "history_path": "/app/checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/sft_history.json",
+  "artifact_path": "/app/checkpoints/sweeps/qwen-qwen2-5-0-5b-instruct/sft_adapter"
+}

docs/results/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json

@@ -0,0 +1,71 @@
+{
+  "status": "ok",
+  "model_source": "merged",
+  "model_load_error": "",
+  "samples": 5,
+  "valid_rate": 1.0,
+  "avg_env_reward": 0.726,
+  "avg_latency_seconds": 2.158,
+  "results": [
+    {
+      "idx": 0,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8000\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "(instruction=\"justify\") \nBenevolentAI's model has shown a high level of accuracy in predicting the next word in sentences. Given this, which candidate would you choose as the most likely to continue the sentence smoothly? Please provide a brief justification for your choice.\" rationale: \"The chosen candidate is more likely to maintain coherence with the existing context and future predictions accurately.\" \n\nMy reasoning:\n- The",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.557,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 1,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8001\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "(instruction=\"justify\") \nBenevolentAI; justification: The instruction is incomplete. Please provide a specific question or task related to the candidates mentioned in order for me to choose an appropriate candidate_id and provide a justified response.\" \"rationale\": \"This question asks for a justification, but no question or task was provided in the initial instruction. To fulfill this request, I need more information about what",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.703,
+      "latency_seconds": 2.04,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 2,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "(instruction=\"justify\") \nBenevolentAI's model has shown a high level of accuracy in predicting the next word in sentences. Given this, which candidate is most likely to be correct for the following sentence fragment: 'The quick brown fox jumps over the lazy dog.'? Choose one candidate_id and provide your justification.\" \nrationale: The choice of 'cand_07' as the",
+      "candidate_id": "cand_07",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.059,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 3,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8003\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "(instruction=\"justify\") \nBenevolentAI's model has shown a high level of accuracy in predicting the next word in sentences. Given this, which candidate would you choose as the most likely continuation for the sentence: 'The new policy will lead to...'? Please provide a brief justification for your choice.\" rationale: \"I would choose candidate_id=cand_05; rationale=This candidate uses",
+      "candidate_id": "cand_05",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.777,
+      "latency_seconds": 2.007,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 4,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "(instruction=\"justify\") \nBenevolentAI's model has shown a high level of accuracy in predicting the next word in sentences. Given this, which candidate would you select as the most likely continuation for the sentence: 'The new policy will lead to...'? Please provide a brief justification for your choice.\" rationale: \"Selecting the best candidate involves considering factors such as syntactic correctness, semantic coherence",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.127,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    }
+  ]
+}

docs/results/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/run_metadata.json

@@ -0,0 +1,9 @@
+{
+  "training_mode": "sft-baseline",
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "model_index": 1,
+  "sft_epochs": 2,
+  "sft_max_steps": 0,
+  "sft_batch_size": 1,
+  "sft_learning_rate": 2e-05
+}

docs/results/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json

@@ -0,0 +1,18 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "examples_used": 2000,
+  "model_id": "Qwen/Qwen2.5-1.5B-Instruct",
+  "unsloth_available": false,
+  "train_runtime": 483.7085,
+  "train_loss": 0.11515871361242898,
+  "train_metrics": {
+    "train_runtime": 483.7085,
+    "train_samples_per_second": 8.269,
+    "train_steps_per_second": 8.269,
+    "total_flos": 1.0823562289152e+16,
+    "train_loss": 0.11515871361242898
+  },
+  "history_path": "/app/checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/sft_history.json",
+  "artifact_path": "/app/checkpoints/sweeps/qwen-qwen2-5-1-5b-instruct/sft_adapter"
+}

docs/results/active_model/sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json

@@ -0,0 +1,71 @@
+{
+  "status": "ok",
+  "model_source": "merged",
+  "model_load_error": "",
+  "samples": 5,
+  "valid_rate": 1.0,
+  "avg_env_reward": 0.762,
+  "avg_latency_seconds": 2.748,
+  "results": [
+    {
+      "idx": 0,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8000\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %{ \"patient_id\": \"patient_8000\", \"rationale\": \"candidate_id=cand_08; The patient is experiencing a high fever, which is a symptom of an infection. This suggests that the patient may have a bacterial infection, which would be treated with antibiotics. Candidate_id=cand_08 prescribes antibiotics, which are appropriate for treating bacterial infections.\",",
+      "candidate_id": "cand_08",
+      "selected_candidate": "cand_08",
+      "env_reward": 0.806,
+      "latency_seconds": 3.164,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 1,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8001\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "{\"candidate_id\": \"candidate_id\": \"cand_05\", \"rationale\": \"patient mentions having a headache and taking medicine. cand_05 explains that the patient is taking medicine for a headache, which aligns with the patient's symptoms. cand_09 does not mention any medication. cand_01 mentions nausea but no headache. cand_02 mentions fever but",
+      "candidate_id": "cand_05",
+      "selected_candidate": "cand_05",
+      "env_reward": 0.792,
+      "latency_seconds": 2.656,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 2,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %+difference_in_format_instruction = true %{\"patient_id\": \"patient_8002\", \"candidate_ids\": [\"cand_07\", \"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10",
+      "candidate_id": "cand_07",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.641,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    },
+    {
+      "idx": 3,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8003\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_05\", \"cand_08\", \"cand_09\", \"cand_10\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; //= = = = = = = = = = = = = = = FUNCTION select_best_candidate(candidate_ids=LIST_of_STR, patient_id=STR, candidates_dict=DICTIONARY_OF_TUPLES) ;; best_candidate_id := candidate_ids[0]; strongest_rationale := candidates_dict[candidate_ids[0]][1]; for i in range(1, len(candidate_ids)) { this",
+      "candidate_id": null,
+      "selected_candidate": "cand_01",
+      "env_reward": 0.777,
+      "latency_seconds": 2.643,
+      "done": false,
+      "valid": true,
+      "termination_reason": "ongoing"
+    },
+    {
+      "idx": 4,
+      "prompt": "{\"instruction\": \"Choose one candidate_id and justify briefly.\", \"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06\"], \"format\": \"candidate_id=<cand_xx>; rationale=<text>\"}",
+      "completion": "; %+difference_in_format_instruction = true %{\"patient_id\": \"patient_8004\", \"candidate_ids\": [\"cand_01\", \"cand_02\", \"cand_03\", \"cand_04\", \"cand_08\", \"cand_09\", \"cand_10\", \"cand_06",
+      "candidate_id": "cand_01",
+      "selected_candidate": "cand_01",
+      "env_reward": 0.717,
+      "latency_seconds": 2.638,
+      "done": true,
+      "valid": true,
+      "termination_reason": "exploit_detection"
+    }
+  ]
+}

docs/results/active_model/sweeps/qwen-qwen2-5-3b-instruct/run_metadata.json

@@ -0,0 +1,9 @@
+{
+  "training_mode": "sft-baseline",
+  "model_id": "Qwen/Qwen2.5-3B-Instruct",
+  "model_index": 2,
+  "sft_epochs": 1,
+  "sft_max_steps": 0,
+  "sft_batch_size": 1,
+  "sft_learning_rate": 2e-05
+}

docs/results/active_model/sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json

@@ -0,0 +1,18 @@
+{
+  "status": "ok",
+  "backend": "trl_transformers",
+  "examples_used": 2000,
+  "model_id": "Qwen/Qwen2.5-3B-Instruct",
+  "unsloth_available": false,
+  "train_runtime": 372.1845,
+  "train_loss": 0.18184852770145518,
+  "train_metrics": {
+    "train_runtime": 372.1845,
+    "train_samples_per_second": 5.374,
+    "train_steps_per_second": 5.374,
+    "total_flos": 1.1456146931712e+16,
+    "train_loss": 0.18184852770145518
+  },
+  "history_path": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_history.json",
+  "artifact_path": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter"
+}

docs/results/active_model_manifest.json

@@ -0,0 +1,68 @@
+{
+  "status": "ok",
+  "enabled": true,
+  "activated_at_utc": "2026-04-26T02:24:15.464507+00:00",
+  "run_id": "qwen-qwen2-5-0-5b-instruct",
+  "source": "top-level",
+  "label": "local-qwen-0.5b-active-smoke",
+  "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "preferred_artifact": "grpo_adapter",
+  "mode": "symlink",
+  "source_checkpoint_dir": "checkpoints",
+  "source_report_dir": "outputs/reports",
+  "grpo_adapter": "checkpoints/active/grpo_adapter",
+  "merged_model": "checkpoints/active/merged",
+  "sft_adapter": "checkpoints/active/sft_adapter",
+  "availability": {
+    "grpo_adapter": true,
+    "merged": true,
+    "sft_adapter": true
+  },
+  "reports": {
+    "improvement_report_benchmark.json": "outputs/reports/active_model/improvement_report_benchmark.json",
+    "anti_hacking_overfit_report.json": "outputs/reports/active_model/anti_hacking_overfit_report.json",
+    "grpo_trl_run_strict_check.json": "outputs/reports/active_model/grpo_trl_run_strict_check.json",
+    "postsave_inference.json": "outputs/reports/active_model/postsave_inference.json",
+    "sft_trl_run.json": "outputs/reports/active_model/sft_trl_run.json",
+    "plot_index.json": "outputs/reports/active_model/plot_index.json",
+    "dose_train.json": "outputs/reports/active_model/dose_train.json",
+    "baselines.json": "outputs/reports/active_model/baselines.json",
+    "robustness.json": "outputs/reports/active_model/robustness.json",
+    "grpo_trl_run_fallback_check.json": "outputs/reports/active_model/grpo_trl_run_fallback_check.json",
+    "sft_run.json": "outputs/reports/active_model/sft_run.json",
+    "benchmark_report.txt": "outputs/reports/active_model/benchmark_report.txt",
+    "dosing_grpo.json": "outputs/reports/active_model/dosing_grpo.json",
+    "grpo_ablation_report.json": "outputs/reports/active_model/grpo_ablation_report.json",
+    "frontier_ready.json": "outputs/reports/active_model/frontier_ready.json",
+    "improvement_report.json": "outputs/reports/active_model/improvement_report.json",
+    "hf_sweep_summary.json": "outputs/reports/active_model/hf_sweep_summary.json",
+    "planner_grpo.json": "outputs/reports/active_model/planner_grpo.json",
+    "grpo_trl_run.json": "outputs/reports/active_model/grpo_trl_run.json",
+    "risk_train.json": "outputs/reports/active_model/risk_train.json",
+    "grpo_trl_run_smoke.json": "outputs/reports/active_model/grpo_trl_run_smoke.json",
+    "inference_benchmark.json": "outputs/reports/active_model/inference_benchmark.json",
+    "supervisor_grpo.json": "outputs/reports/active_model/supervisor_grpo.json",
+    "acceptance_gate.json": "outputs/reports/active_model/acceptance_gate.json",
+    "grpo_trl_run_auto.json": "outputs/reports/active_model/grpo_trl_run_auto.json",
+    "hf_training_status.json": "outputs/reports/active_model/hf_training_status.json",
+    "benchmark_report.json": "outputs/reports/active_model/benchmark_report.json",
+    "postsave_inference_smoke.json": "outputs/reports/active_model/postsave_inference_smoke.json",
+    "graph_train.json": "outputs/reports/active_model/graph_train.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-0-5b-instruct/postsave_inference_sft.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-3b-instruct/postsave_inference_sft.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/sft_trl_run.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/run_metadata.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/run_metadata.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/sft_history.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/sft_history.json",
+    "sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json": "outputs/reports/active_model/sweeps/qwen-qwen2-5-1-5b-instruct/postsave_inference_sft.json",
+    "grpo_training_cycle/grpo_trl_run.json": "outputs/reports/active_model/grpo_training_cycle/grpo_trl_run.json",
+    "grpo_training_cycle/hf_training_status.json": "outputs/reports/active_model/grpo_training_cycle/hf_training_status.json"
+  },
+  "notes": "This manifest controls local product inference. Prefer grpo_adapter for the RL policy; merged is the SFT baseline fallback when no GRPO adapter is available."
+}

docs/results/anti_hacking_overfit_report.json

@@ -0,0 +1,22 @@
+{
+  "passed": false,
+  "training_mode": "full",
+  "warnings": [
+    "Qwen2.5-3B:high_exploit_rate"
+  ],
+  "completed_models": [
+    "Qwen/Qwen2.5-3B-Instruct"
+  ],
+  "failed_or_skipped_models": [],
+  "checks": {
+    "reward_bounds": [
+      0.001,
+      0.999
+    ],
+    "reward_precision": 3,
+    "fallback_backends_rejected": true,
+    "exploit_rate_threshold": 0.35,
+    "train_holdout_gap_threshold": 0.25,
+    "min_validity_rate": 0.8
+  }
+}