TraceGRPO research update 20261009T141408Z (Review 21-23)

.gitignore

@@ -115,5 +115,7 @@ results/phase7/codebert_finetune/**
 !results/phase6/review21/**
 !results/phase6/review22/
 !results/phase6/review22/**
+!results/phase6/review23/
+!results/phase6/review23/**
 !results/phase6/scale_review21_step324_lr5e6/
 !results/phase6/scale_review21_step324_lr5e6/**

DECISIONS.md

@@ -2613,3 +2613,36 @@ Continuation directory: results/phase6/scale_review21_step324_lr5e6. The histori
 Accepted segment cost39.924725 GPU allocation wall-hours. Identified scale launches cost41.792836 hours including invalid v3/v4, leaving at most208.207164 hours under250. Early v1/v2 lack final-status records; the ledger is not a complete historical project-cost total. The CPU audit/tests consumed no GPU training hours. Next authorization is one40-hour segment including load/eval; conservative boundary checkpoint remains12 minutes before the supervisor's hard cap. Next scheduled full chronological VALID/checkpoint at step400, then600; best selection remains VALID P-Pass@1. Report at the next40-hour boundary or any tripwire. No automatic continuation past that boundary, no new experiment, no TEST.
 
 Review #21 launch confirmed at2026-10-09T12:51:08Z. Supervisor3904612 /worker3904667. The GPU worker restored324 scheduler steps and all144 Adam parameter states, next-update LR4.8651708513e-6; no Adam reset or warmup replay. Config SHA2568b4bdbd2a700cab50f70ade78b41c201be87bc8f08c99aa29f348fa442cd81a0; continuation manifest SHA256857fd234d80e105c08837207bad411e3d135295e2e90bb992e80e9d9f16a17c0. Source adapter SHA90ee18642fe528f3d83bdee39623bffc32cc523f65514676debfc6ceb751a421. See results/phase6/review21/launch_report.json. At startup324 optimizer steps were complete; no post-change result is yet available. Estimated step400 VALID milestone9–11 hours at historical throughput, not a guaranteed completion time. The resumed policy uses the live adapter without merging weights; phase budget supervises from process launch and holds after the authorized segment. The source config's init_adapter hash is historical SFT lineage; the effective continuation adapter is explicitly recorded as step324. In completion records, effective_learning_rate is the post-scheduler next-update LR. Training curves show the red seam after324. No HF mid-run snapshot or TEST action was performed.
+
+## Review #23 — continuation limit, experiment authorizations, and reward-record correction (2026-10-09)
+
+Review #23 authorizes the existing LR=5e-6 GRPO continuation only through optimizer step 400. Step 400 is the maximum; the supervisor/watchdog must stop before step 401, verify the full-state checkpoint, run the full chronological VALID evaluation after stopping, archive artifacts, and never resume automatically. The pre-registered Review #22 §4.5 predictions are to be tested against the continuation's sampled vulnerable rate, SAFE/VULNERABLE R2 behavior, nonempty statement emission, classification accuracy, and class balance. TEST remains sealed. E1 and E2 are authorized; E3 is authorized only after step 400, and must use its frozen pre-registered protocol, three seeds and specified ablations, 2021 TRAIN holdout for model selection, and 2022 chronological VALID for its one final evaluation. No E4 or further GRPO run is authorized. The private Hugging Face repository is scheduled for incremental snapshots every two hours; secrets and local auth stores remain excluded.
+
+### Corrected interpretation of the reward-induced class-prior tilt
+
+The earlier working hypothesis attributed the vulnerable-direction prior tilt to the SAFE empty-set R2 convention. Review #22's complete-group replay and fixed-output reward counterfactuals correct that interpretation. The production 1.5× vulnerable-label multiplier, combined with inverse-correctness difficulty weighting, produced the dominant observed prior tilt. In the post-reset complete-group replay, production weights yielded a net vulnerable-direction push of 6.62 reward-advantage units per group (95% CI 0.35 to 12.77), while removing the label multiplier reduced it to 1.39 (95% CI −3.62 to 6.26), and unit weights reduced it to 0.10 (95% CI −2.05 to 2.16). In the pre-reset replay, the corresponding production push was +6.98 per group; the direction and magnitude are consistent. Turning SAFE empty-set R2 off increased the vulnerable-direction push (+8.43 post-reset in the reported counterfactual), rather than removing it. Therefore the SAFE empty-set convention did not cause the directional prior tilt. It did, however, reward vacuous empty statement emission on SAFE examples and encouraged that output behavior.
+
+This revises the causal account; it does not rewrite or delete prior run artifacts. Future reward designs must begin with symmetric label weights, must not reward vacuous non-emission, and must verify the resulting within-group reward advantages before training. The underlying counterfactual is fixed-output analysis, not evidence that retraining with corrected rewards would improve validation discrimination.
+
+The Review #22 evidence further records that all 10,752 smoke rollouts were replayed; 18,722/18,722 non-failure reward records reproduced exactly, with 334 failure-floor records retained. These numbers establish reward implementation/replay consistency only, not model improvement.
+
+Pending authorized work at this entry: hard stop and full VALID evaluation at step 400; CPU E1 predicted-slice feasibility; matched-inference E2 row ROC-AUC and pair-cluster intervals; E3 only after the stop and the E1/E2 gates. No conclusion about publication framing is made here.
+
+### Review #23 scientific-record correction — reward-driven class-prior tilt
+
+The first interpretation attributed the vulnerable-direction prior shift to the SAFE empty-set R2 convention. That hypothesis is superseded by the stored reward counterfactual experiment (`results/phase6/review22/reward_counterfactuals.json`): with production settings, the vulnerable label multiplier of 1.5× combined with difficulty weighting produced the dominant measured shift (+6.62097 vulnerable-direction net advantage per complete post-reset group; pair/group bootstrap 95% interval [0.3514, 12.7713]). Removing the label multiplier reduced that estimate to +1.3885 (95% interval [−3.616, 6.264]); unit label weights yielded +0.1004 (95% interval [−2.047, 2.160]). Disabling SAFE-empty R2 did not remove the tilt; its measured tilt was +8.43, while it encouraged empty statement emission. Revised interpretation: the SAFE empty-set convention did not cause the directional class-prior tilt, although it encouraged vacuous non-emission. Future reward designs start with symmetric label weights, do not reward vacuous non-emission, and verify within-group advantages before training.
+
+### Review #23 progress snapshot — 2026-10-09 14:01 UTC
+
+The LR=5e-6 continuation is active at optimizer step 332; the independently running hard-stop watchdog is configured to verify step400 full-state artifacts and terminate before step401. Latest 20-step verifier failure rate is approximately1.1%, aggregate parse approximately98.66%, and entropy approximately0.53 nats; no current stop tripwire is reported. E1 CPU feasibility is running with pair-grouped five-fold out-of-fold localization; folds0–2 have completed without data leakage errors. This is an interim status, not a result. E2 waits for GPU release. Do not resume GRPO after its step400 report.
+
+Literature audit correction: primary arXiv pages verify R2Vul (arXiv:2504.04699; v2, 2025-08-07), Simoni et al. GRPO (arXiv:2507.03051; submitted 2025-07-03, under review), SecCodePRM (arXiv:2602.10418; submitted 2026-02-11), and that FocusVul (arXiv:2505.17460) was withdrawn by an author for fundamental methodological errors. The FocusVul claims are not treated as validated evidence. The audit file distinguishes bibliographic/method verification from unverified numerical claims and splits.
+
+
+### Review #23 E1 result — 2021 forward TRAIN holdout (2026-10-09)
+
+The CPU-only E1 predicted-slice feasibility experiment completed with pair-grouped five-fold OOF locator predictions, a pre-2021 TRAIN fit set (7,370 pairs), and a forward-2021 TRAIN holdout (862 pairs / 1,724 rows). The code verified pair/date partition accounting and zero train/holdout overlap; chronological VALID and TEST were not read. It tested fixed-radius windows around the top three diff-supervised locator lines, with a full-function classifier, random same-line-budget windows and slice-size-only control.
+
+The E1 gate is **no-go for Direction C**. Overall row AUC was 0.529 full vs 0.510 predicted slice (paired delta −0.0186, 95% pair-cluster CI [−0.0291, −0.0080]). On >100-line functions, AUC was 0.530 full vs 0.493 predicted slice (delta −0.0370, CI [−0.0581, −0.0172]); random same-size slices scored 0.527. On ≤30-line functions the predicted-slice loss was −0.0093, within the allowed −0.01, but this does not offset failure of the primary long-function gate, which required at least +0.03. No holdout tuning was done. Full details and limitations: `results/phase6/review23/e1_report.md` and `e1_predicted_slices.json`.
+
+Review #22 §4.4 says E3 runs after E1 and E2 pass their specified gates. Therefore E3 is held unless the controlling authorization explicitly supersedes this pre-registered dependency; a passing E2 alone does not erase the E1 no-go. E2 remains scheduled after the step-400 stop.

RELEASE_INFO.json

@@ -8,7 +8,7 @@
   "snapshot_includes_uncommitted_working_tree_artifacts": true,
   "research_status": {
     "sft_gate": "failed; preserved reports",
-    "grpo_scale": "active, resumed from step 324; latest completed step 331; Review #23 hard stop step400 pending full VALID",
+    "grpo_scale": "active, resumed from step 324; latest completed step 334; Review #23 hard stop step400 pending full VALID",
     "review22": "CPU reward-counterfactual, classical threshold-free and oracle-slice feasibility artifacts present; literature verification and E1/E2/E3 remain in progress",
     "chronological_test": "sealed; not evaluated"
   },

results/hf_sync_20261009_review23_status.json

@@ -1,25 +1,25 @@
 {
-  "created_at_utc": "2026-10-09T13:52:36.945116+00:00",
+  "created_at_utc": "2026-10-09T14:14:08.370563+00:00",
   "repo_id": "Dhruv1000/tracegrpo",
   "repo_type": "model",
   "private": true,
   "base_remote_revision": "b92d1c2e46ee9f8c697ebc3f61664f30a095949c6",
   "base_snapshot_files_present_remote": 19790,
   "base_remote_extra_files_preserved": 29,
-  "active_run_step_observed": 331,
+  "active_run_step_observed": 334,
   "test_touched": false,
   "secrets_uploaded": false,
   "paths": [
     {
       "path": ".gitignore",
-      "sha256": "214b1c93dd835cf7788010eb8ad1ff97e9038946b9b4e7f59c2009086cfa4a9c",
-      "size_bytes": 4299,
+      "sha256": "6e3c1aef03813bf49cd39069fec909b603a1d90b6a3925886c1c7bc01b29f18b",
+      "size_bytes": 4353,
       "change": "updated"
     },
     {
       "path": "DECISIONS.md",
-      "sha256": "693043f5f7d32b86293df4048e0f9478614191f752e7bf7d1d46ceaf5847e55d",
-      "size_bytes": 174064,
+      "sha256": "09bd63a1729365f2d8ae68e10f2ec8be2b4fb3498c21c6db289f1bd724900d55",
+      "size_bytes": 181200,
       "change": "updated"
     },
     {
@@ -36,7 +36,7 @@
     },
     {
       "path": "RELEASE_INFO.json",
-      "sha256": "3202dfc80c66a6ca6cbcc755c205159c4c93461b847ac6b82779b0614716b336",
+      "sha256": "0afb4802018fed6a2dcc12cf7802acf2f410f511538fb9513610b18597107735",
       "size_bytes": 1450,
       "change": "updated"
     },
@@ -100,6 +100,36 @@
       "size_bytes": 60111,
       "change": "new"
     },
+    {
+      "path": "results/phase6/review23/e1_predicted_slices.json",
+      "sha256": "61b51cfe9b96e60a42dd0bfef111ab4c4ae1c7fafe02fb49ae93e82a03c84882",
+      "size_bytes": 6449,
+      "change": "new"
+    },
+    {
+      "path": "results/phase6/review23/e1_predicted_slices.log",
+      "sha256": "c115519ce7013b3a927d93bc204f2f0923b0cc1928233c63133215a36b037c4f",
+      "size_bytes": 6118,
+      "change": "new"
+    },
+    {
+      "path": "results/phase6/review23/e1_report.md",
+      "sha256": "041ad5e721c85ce6df0bce2c190db168a2c2415fa11028964d760e809b9796f2",
+      "size_bytes": 2651,
+      "change": "new"
+    },
+    {
+      "path": "results/phase6/review23/e2_preflight.json",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size_bytes": 0,
+      "change": "new"
+    },
+    {
+      "path": "results/phase6/review23/literature_audit.md",
+      "sha256": "9ad640b18da638e7967171a848b4bbeb2c6e2972856f8b0ce644d50fab834cf2",
+      "size_bytes": 21500,
+      "change": "new"
+    },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/best_checkpoint.json",
       "sha256": "a23a9160e1c9bb1ce73b48f24721c16b7d62f48ce4ba839d90a9ad74d82bdbef",
@@ -108,8 +138,8 @@
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/curves.json",
-      "sha256": "fd586fa5d49927330c76919a02498f9ede39d6cd1086a7cddfa68328e48e1a4e",
-      "size_bytes": 755579,
+      "sha256": "033cb41dff455e5130151569d6be4ed2ceec66ab29642abf30a35ae648e9b13e",
+      "size_bytes": 763221,
       "change": "new"
     },
     {
@@ -132,8 +162,8 @@
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/judge_outputs.jsonl",
-      "sha256": "d67d26a3a836bdc1540496178095b4f69ce2688d934a4d2b6d8a39b6d7fa94a1",
-      "size_bytes": 195343,
+      "sha256": "9d86edd85b871bda970cc870566d7615db4e4b602ac529b2b36d1ff3e417a127",
+      "size_bytes": 250988,
       "change": "new"
     },
     {
@@ -162,8 +192,8 @@
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/progress.json",
-      "sha256": "e718393f39ec35dfc43684f751b1814f3a3b21d4aaaedf78c12b75e793b4f939",
-      "size_bytes": 183,
+      "sha256": "b4b61ebd6fdd1e68fa016f281aec4de9e0625e411e71ed655fb2fa2277080ae1",
+      "size_bytes": 267,
       "change": "new"
     },
     {
@@ -180,8 +210,8 @@
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/rollout_failures.jsonl",
-      "sha256": "a54f2281b7076dc4c0e6e061275176b3ddbaf3c208402432b2235f4bec52e4a8",
-      "size_bytes": 386431,
+      "sha256": "55cff72a55267c2c372515b80c1967278c7a51c49b63c2c8d642df8535b12240",
+      "size_bytes": 482009,
       "change": "new"
     },
     {
@@ -192,32 +222,32 @@
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/sampled_generated_rollouts.jsonl",
-      "sha256": "af15ef187d2b822cb082372a53a4598e0b1904d0a60f4e4fc74d91b78071baa3",
-      "size_bytes": 1350213,
+      "sha256": "5d8d9cf17a0e24f10f562e74c9cda75fdc1134b88598c9da54d2554a0376a965",
+      "size_bytes": 1810667,
       "change": "new"
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/sampled_training_rollouts.jsonl",
-      "sha256": "dd417dc9a83b1cede38f0f8bf0f83a9f58ebe3a93fafbd5d493472958d8ed8c1",
-      "size_bytes": 1256771,
+      "sha256": "d531bab6a4e2b30bbc6064f50540c2459bcf4303c3589f58373aac93189c22da",
+      "size_bytes": 1865688,
       "change": "new"
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/sampled_verified_rollouts.jsonl",
-      "sha256": "f55cb1dc102e9642fb23795d4cfdb667917800ec0caf5ba92077232e2c8c4a05",
-      "size_bytes": 1354484,
+      "sha256": "491af5bc5e030a20e1bff4e6109019b6d86e6bed0fd800934d9e228e6b9ac678",
+      "size_bytes": 1844403,
       "change": "new"
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/scale.log",
-      "sha256": "9c7ec390e16962c2ff90e91d91a81f2c0c745c4ab5bc1bade24d45d68a68e761",
-      "size_bytes": 60797,
+      "sha256": "3b1c545a92516b45b692355889f9be241296b2d27d58a17bc5e3ec6beed09d04",
+      "size_bytes": 73571,
       "change": "new"
     },
     {
       "path": "results/phase6/scale_review21_step324_lr5e6/step_metrics.jsonl",
-      "sha256": "572690ba0b479bec0e19d9970805a02535073cf391b810ef6477a116b0171f66",
-      "size_bytes": 15602,
+      "sha256": "3518e7a904763b12130eb5f05c88ce46a482087719ef8850ea2165d2989f67a8",
+      "size_bytes": 22011,
       "change": "new"
     },
     {
@@ -246,8 +276,8 @@
     },
     {
       "path": "scripts/stop_scale_at_step400.py",
-      "sha256": "4e88becb11828db0d4b04906878e4b791c9f786b64390b47f50e7360f4fcb9b7",
-      "size_bytes": 4158,
+      "sha256": "643071ada3fa4d40c8a24ebd93ec0327a43e0326d006942daacc374e85116760",
+      "size_bytes": 4156,
       "change": "new"
     }
   ],
@@ -257,9 +287,5 @@
     "results/phase6/scale_review21_step324_lr5e6/checkpoints/step-000400/**",
     "Current per-step live adapter; it is mutable and not a resumable checkpoint."
   ],
-  "parent_remote_revision": "9ae66d3904a23842425046330ec1a5b82958d4bf",
-  "upload_commit": "1f71724e167a51f45b1ec43c732eadbbbcdac1ec",
-  "remote_revision_after": "1f71724e167a51f45b1ec43c732eadbbbcdac1ec",
-  "verified_remote_paths": 40,
-  "final_remote_revision": "1f71724e167a51f45b1ec43c732eadbbbcdac1ec"
+  "parent_remote_revision": "efb5fbefef76409596ea809aa76dfb4f1c51cb15"
 }

results/phase6/review23/e1_predicted_slices.json

@@ -0,0 +1,228 @@
+{
+  "experiment": "Review23 E1 CPU predicted-slice feasibility",
+  "config": {
+    "seed": 42,
+    "locator_crossfit_folds": 5,
+    "top_seed_lines": 3,
+    "window_radius": 2,
+    "max_function_lines": 4096,
+    "bootstrap_replicates": 5000,
+    "slice_rule": "union of +/-2 line windows around top 3 locator-ranked source lines",
+    "line_target": "diff-derived patch_lines_vuln on vulnerable pre-2021 TRAIN rows only"
+  },
+  "data": {
+    "pre2021_train_pairs": 7370,
+    "forward_2021_holdout_pairs": 862,
+    "undated_train_pairs_excluded": 0,
+    "train_split_sha256": "a365b114eecfb33fc9e08bda82e11c075530acb4db573cd8a4a29b9281503e4a",
+    "date_cache_sha256": "ef914835991298957438343e76f11a77f928d1df67e8636953788e0df6fcada7",
+    "corpus_sha256": "fda8efe71d0d74e9f0ca8d7b2f050481972363dbed80d364cc9271d83ea3435b",
+    "test_read": false,
+    "chronological_valid_read": false,
+    "pair_overlap": 0
+  },
+  "holdout_row_counts": {
+    "rows": 1724,
+    "safe": 862,
+    "vulnerable": 862
+  },
+  "metrics": {
+    "all": {
+      "full_function": {
+        "auc": 0.5286968201075575,
+        "ci95": [
+          0.5219116902902116,
+          0.5355949815892463
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5101016898057181,
+        "ci95": [
+          0.49969823455946083,
+          0.5207069668821765
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5160912408955594,
+        "ci95": [
+          0.4952023534272533,
+          0.5371888979656655
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.506307163505795,
+        "ci95": [
+          0.4939073258111229,
+          0.5184870949768788
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.018595130301839347,
+        "pair_cluster_ci95": [
+          -0.029051762076000846,
+          -0.00801984808436646
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": -0.005989551089841227,
+        "pair_cluster_ci95": [
+          -0.02866280462529812,
+          0.01693727087494148
+        ]
+      }
+    },
+    "le30": {
+      "full_function": {
+        "auc": 0.5282517424665467,
+        "ci95": [
+          0.5063586165466457,
+          0.551413312195885
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5189654815825939,
+        "ci95": [
+          0.49156191779319947,
+          0.5460544888545237
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5176728506132242,
+        "ci95": [
+          0.4849043981847654,
+          0.5521557073375465
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.5108684411904615,
+        "ci95": [
+          0.47588881926648136,
+          0.5453234919152746
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.00928626088395279,
+        "pair_cluster_ci95": [
+          -0.03602488693436783,
+          0.016895064431930343
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": 0.0012926309693697124,
+        "pair_cluster_ci95": [
+          -0.037650924240432974,
+          0.041127893441307524
+        ]
+      }
+    },
+    "31_100": {
+      "full_function": {
+        "auc": 0.5258550350173387,
+        "ci95": [
+          0.5137995388466453,
+          0.5393698644336296
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5131807982593322,
+        "ci95": [
+          0.495928373554597,
+          0.5302597693992738
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5064153124362548,
+        "ci95": [
+          0.4737825782235385,
+          0.5374951316079581
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.49492078602026246,
+        "ci95": [
+          0.474269333571956,
+          0.5166331962378459
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.012674236758006496,
+        "pair_cluster_ci95": [
+          -0.030460532942386927,
+          0.004576771939678252
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": 0.006765485823077344,
+        "pair_cluster_ci95": [
+          -0.031342862961154905,
+          0.04408230571559027
+        ]
+      }
+    },
+    "gt100": {
+      "full_function": {
+        "auc": 0.5304826925380285,
+        "ci95": [
+          0.5181359856164629,
+          0.5446499653824551
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.49346478620194034,
+        "ci95": [
+          0.47780355463058044,
+          0.5088132885817919
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5273835189843095,
+        "ci95": [
+          0.4862254833800285,
+          0.569569595301933
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.5030542579949695,
+        "ci95": [
+          0.48346118322915166,
+          0.5233351833527657
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.03701790633608815,
+        "pair_cluster_ci95": [
+          -0.058117530441811956,
+          -0.017156310405371064
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": -0.03391873278236912,
+        "pair_cluster_ci95": [
+          -0.07894993751168335,
+          0.012249716489872381
+        ]
+      }
+    }
+  },
+  "criteria": {
+    "go": ">100-line predicted-slice minus full-function AUC >= .03 and pair-cluster CI excludes 0; <=30-line loss <= .01",
+    "no_go": ">100-line improvement <= .01",
+    "review22_e1_gate_applied": true,
+    "gate_outcome": "no_go_for_direction_C",
+    "decision": "Predicted slicing does not improve >100-line AUC; the pre-registered E1 success gate is not met. Per Review22 \u00a74.4, E3 remains gated on E1 and E2 passing; do not launch E3 unless governing authorization explicitly supersedes that condition.",
+    "observed": {
+      "gt100_predicted_minus_full_auc": -0.03701790633608815,
+      "gt100_pair_cluster_ci95": [
+        -0.058117530441811956,
+        -0.017156310405371064
+      ],
+      "le30_predicted_minus_full_auc": -0.00928626088395279,
+      "overall_predicted_minus_full_auc": -0.018595130301839347
+    }
+  },
+  "limitations": [
+    "The line locator's diff supervision is available only on vulnerable functions; it is applied to both candidate class members.",
+    "Predicted slices are fixed windows, not data-flow closures; dependency-aware expansion is not evaluated in this CPU feasibility pass.",
+    "The random-window control is matched to predicted-slice line count, not exact token count."
+  ]
+}

results/phase6/review23/e1_predicted_slices.log

@@ -0,0 +1,222 @@
+{"fold": 0, "locator_train_pairs": 5896, "oof_rows": 2948}
+{"fold": 1, "locator_train_pairs": 5896, "oof_rows": 2948}
+{"fold": 2, "locator_train_pairs": 5896, "oof_rows": 2948}
+{"fold": 3, "locator_train_pairs": 5896, "oof_rows": 2948}
+{"fold": 4, "locator_train_pairs": 5896, "oof_rows": 2948}
+{
+  "experiment": "Review23 E1 CPU predicted-slice feasibility",
+  "config": {
+    "seed": 42,
+    "locator_crossfit_folds": 5,
+    "top_seed_lines": 3,
+    "window_radius": 2,
+    "max_function_lines": 4096,
+    "bootstrap_replicates": 5000,
+    "slice_rule": "union of +/-2 line windows around top 3 locator-ranked source lines",
+    "line_target": "diff-derived patch_lines_vuln on vulnerable pre-2021 TRAIN rows only"
+  },
+  "data": {
+    "pre2021_train_pairs": 7370,
+    "forward_2021_holdout_pairs": 862,
+    "undated_train_pairs_excluded": 0,
+    "train_split_sha256": "a365b114eecfb33fc9e08bda82e11c075530acb4db573cd8a4a29b9281503e4a",
+    "date_cache_sha256": "ef914835991298957438343e76f11a77f928d1df67e8636953788e0df6fcada7",
+    "corpus_sha256": "fda8efe71d0d74e9f0ca8d7b2f050481972363dbed80d364cc9271d83ea3435b",
+    "test_read": false,
+    "chronological_valid_read": false,
+    "pair_overlap": 0
+  },
+  "holdout_row_counts": {
+    "rows": 1724,
+    "safe": 862,
+    "vulnerable": 862
+  },
+  "metrics": {
+    "all": {
+      "full_function": {
+        "auc": 0.5286968201075575,
+        "ci95": [
+          0.5219116902902116,
+          0.5355949815892463
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5101016898057181,
+        "ci95": [
+          0.49969823455946083,
+          0.5207069668821765
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5160912408955594,
+        "ci95": [
+          0.4952023534272533,
+          0.5371888979656655
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.506307163505795,
+        "ci95": [
+          0.4939073258111229,
+          0.5184870949768788
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.018595130301839347,
+        "pair_cluster_ci95": [
+          -0.029051762076000846,
+          -0.00801984808436646
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": -0.005989551089841227,
+        "pair_cluster_ci95": [
+          -0.02866280462529812,
+          0.01693727087494148
+        ]
+      }
+    },
+    "le30": {
+      "full_function": {
+        "auc": 0.5282517424665467,
+        "ci95": [
+          0.5063586165466457,
+          0.551413312195885
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5189654815825939,
+        "ci95": [
+          0.49156191779319947,
+          0.5460544888545237
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5176728506132242,
+        "ci95": [
+          0.4849043981847654,
+          0.5521557073375465
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.5108684411904615,
+        "ci95": [
+          0.47588881926648136,
+          0.5453234919152746
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.00928626088395279,
+        "pair_cluster_ci95": [
+          -0.03602488693436783,
+          0.016895064431930343
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": 0.0012926309693697124,
+        "pair_cluster_ci95": [
+          -0.037650924240432974,
+          0.041127893441307524
+        ]
+      }
+    },
+    "31_100": {
+      "full_function": {
+        "auc": 0.5258550350173387,
+        "ci95": [
+          0.5137995388466453,
+          0.5393698644336296
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.5131807982593322,
+        "ci95": [
+          0.495928373554597,
+          0.5302597693992738
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5064153124362548,
+        "ci95": [
+          0.4737825782235385,
+          0.5374951316079581
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.49492078602026246,
+        "ci95": [
+          0.474269333571956,
+          0.5166331962378459
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.012674236758006496,
+        "pair_cluster_ci95": [
+          -0.030460532942386927,
+          0.004576771939678252
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": 0.006765485823077344,
+        "pair_cluster_ci95": [
+          -0.031342862961154905,
+          0.04408230571559027
+        ]
+      }
+    },
+    "gt100": {
+      "full_function": {
+        "auc": 0.5304826925380285,
+        "ci95": [
+          0.5181359856164629,
+          0.5446499653824551
+        ]
+      },
+      "predicted_slice_oof": {
+        "auc": 0.49346478620194034,
+        "ci95": [
+          0.47780355463058044,
+          0.5088132885817919
+        ]
+      },
+      "random_same_size_slice": {
+        "auc": 0.5273835189843095,
+        "ci95": [
+          0.4862254833800285,
+          0.569569595301933
+        ]
+      },
+      "slice_size_only": {
+        "auc": 0.5030542579949695,
+        "ci95": [
+          0.48346118322915166,
+          0.5233351833527657
+        ]
+      },
+      "predicted_minus_full": {
+        "mean_delta": -0.03701790633608815,
+        "pair_cluster_ci95": [
+          -0.058117530441811956,
+          -0.017156310405371064
+        ]
+      },
+      "predicted_minus_random": {
+        "mean_delta": -0.03391873278236912,
+        "pair_cluster_ci95": [
+          -0.07894993751168335,
+          0.012249716489872381
+        ]
+      }
+    }
+  },
+  "criteria": {
+    "go": ">100-line predicted-slice minus full-function AUC >= .03 and pair-cluster CI excludes 0; <=30-line loss <= .01",
+    "no_go": ">100-line improvement <= .01",
+    "review22_e1_gate_applied": false
+  },
+  "limitations": [
+    "The line locator's diff supervision is available only on vulnerable functions; it is applied to both candidate class members.",
+    "Predicted slices are fixed windows, not data-flow closures; dependency-aware expansion is not evaluated in this CPU feasibility pass.",
+    "The random-window control is matched to predicted-slice line count, not exact token count."
+  ]
+}

results/phase6/review23/e1_report.md

@@ -0,0 +1,28 @@
+# Review #23 E1 — Predicted-slice feasibility (CPU-only)
+
+**Status: no-go for the pre-registered slicing direction.** This is a negative result, not a reason to tune the locator on the 2021 holdout.
+
+## Protocol and leakage controls
+
+- Locator trained on vulnerable examples from pre-2021 chronological TRAIN (7,370 pairs). The forward 2021 TRAIN holdout contains 862 pairs / 1,724 rows (862 SAFE, 862 VULNERABLE).
+- Five-fold cross-fitting grouped by pair produced out-of-fold predicted slices for classifier training. Both rows of each pair remain in one fold. The final locator fit uses pre-2021 vulnerable rows only.
+- Evaluated full functions, predicted slices, random same-size slices, and slice-size-only scores. Pair-cluster bootstrap: 5,000 replicates, seed 42.
+- The script verifies split accounting and reports no overlap. It did not read chronological VALID or TEST. Corpus, split and date-cache SHA-256 values are in `e1_predicted_slices.json`.
+- This tests fixed-radius windows around three locator-ranked lines. It does not test dependency-closure slices; the random-window control matches line count, not token count.
+
+## Results
+
+| 2021 TRAIN holdout stratum | Full function AUC (95% pair CI) | Predicted slice AUC (95% pair CI) | Random same-size AUC | Predicted − full (paired 95% CI) |
+|---|---:|---:|---:|---:|
+| All rows | 0.529 [0.522, 0.536] | 0.510 [0.500, 0.521] | 0.516 | −0.0186 [−0.0291, −0.0080] |
+| ≤30 lines | 0.528 [0.506, 0.551] | 0.519 [0.492, 0.546] | 0.518 | −0.0093 [−0.0360, 0.0169] |
+| 31–100 lines | 0.526 [0.514, 0.539] | 0.513 [0.496, 0.530] | 0.506 | −0.0127 [−0.0305, 0.0046] |
+| >100 lines | 0.530 [0.518, 0.545] | 0.493 [0.478, 0.509] | 0.527 | −0.0370 [−0.0581, −0.0172] |
+
+The >100-line result contradicts the predicted benefit: the candidate loses 3.7 AUC points versus full functions, and the pair-cluster interval excludes zero in the negative direction. The ≤30-line loss is −0.0093, within the pre-registered maximum 0.01 loss. Overall, predicted slices are also worse than full functions; they do not beat the random-window control.
+
+## Gate application
+
+Review #22 §4.3 success required >100-line predicted-slice improvement ≥+0.03 with paired CI excluding zero, and ≤30-line loss no worse than −0.01. The first condition fails decisively; therefore E1 is **no-go for Direction C**. Review #22 §4.4 conditions E3 on E1 and E2 passing their gates. Under that pre-registration, E3 must remain on hold even if E2 passes, unless an explicit governing ruling changes that dependency.
+
+No hyperparameters or thresholds were changed after observing the holdout.

results/phase6/review23/literature_audit.md

@@ -0,0 +1,53 @@
+# Review #23 Literature and Novelty Audit
+
+**Checked:** 2026-10-09. Every status below is based on an original arXiv record, publisher/conference page, or author-maintained artifact. “Preprint” does not imply peer review. Reported results use each paper's own data split, prompt, metrics, and model; they are not directly comparable to TraceGRPO's 2022 chronological VALID.
+
+## Closest prior work
+
+| Work and verified bibliographic status | Dataset and evaluation actually reported | Contribution and exact overlap with TraceGRPO | Still untested here |
+|---|---|---|---|
+| [VULPO: Context-Aware Vulnerability Detection via On-Policy LLM Optimization](https://arxiv.org/abs/2511.11896), Youpeng Li, Fuxun Yu, Weiliang Qi, Xinda Wang; submitted 2025-11-14, v3 2026-05-26; arXiv preprint (venue not verified). | ContextVul combines PrimeVul and SecVulEval; chronological 8:1:1 split (15,760/1,970/1,970); Qwen3 policy, vLLM inference; primary P-Pass@1/P-Pass@8 and LLM-judge classification. | Closest method prior. Two-stage SFT then on-policy GRPO; multidimensional verdict/localization/causal-reasoning reward, label-level and sample-level difficulty weighting, judge-based reward/evaluation. This overlaps TraceGRPO's backbone family, paired function setting, SFT→GRPO, gated auxiliary rewards, and class/difficulty weighting substantially. It also reports teacher reasoning vetted using CVE and patch metadata, which is a potential rationale-quality concern; this is not proof of their results being invalid. | A common-data, common-prompt, threshold-free comparison; robustness under pair-cluster row-AUC; whether their weighting yields a vulnerable prior shift under our exact rollout/reward distribution; and whether any advantage persists under an independent human audit. Do not claim novelty for this general recipe. |
+| [Graph Is the Verifier: Agentic Reinforcement Learning for Interprocedural Vulnerability Detection](https://arxiv.org/abs/2607.26656), Yikun Li, Ting Zhang, Jiakun Liu, Jinfeng Jiang, Yuheng Yieh, Yixin Yang, Wen Bin Leow, Yide Yin, Yintong Huo, Eng Lieh Ouh, Lwin Khin Shar, David Lo; submitted 2026-07-29; arXiv preprint (venue not verified). | PrimeVul paired examples; strict repository-level 8:1:1 split, 1,122 functions/275 repositories after dedup and Joern filtering; external TitanVul_OOD (648 functions/217 repositories); P-C pair correctness, accuracy, class imbalance. Qwen2.5-Coder-7B agent queries Joern CPG. | Closely overlaps CPG-grounded evidence reward and SFT warm start + GRPO. It directly demonstrates outcome-only reward drifting toward both-pair-vulnerable predictions (31.7% vs 19.0% with composite reward). Its principal distinction is interactive interprocedural CPG queries and exact persistent node-ID evidence verification, not TraceGRPO's cached intra-function localization/DFG matching. | Replication on a chronological split with the same classical baselines; efficacy of intra-function traces when no interprocedural context is available; whether our current DFG/line rewards improve row AUC. |
+| [Vulnerability Detection with Code Language Models: How Far Are We? (PrimeVul)](https://arxiv.org/abs/2403.18624), Yangruibo Ding, Yanjun Fu, Omniyyah Ibrahim, Chawin Sitawarin, Xinyun Chen, Basel Alomair, David Wagner, Baishakhi Ray, Yizheng Chen; submitted 2024-03-27, v2 2024-07-10; published ICSE 2025. | PrimeVul C/C++ benchmark; deduplication, label methods and chronological splits; proposes VD-S and paired evaluation. The paper also tests class weighting and contrastive learning, reporting marginal improvements in its own setup. | Defines the data/paired-evaluation lineage relevant to TraceGRPO and documents duplication, label-noise and evaluation pitfalls. Our SecVulEval work is related but uses a distinct source dataset. | Whether traces/rewards improve threshold-free discrimination on this corpus; pair-cluster intervals and long-function strata. |
+| [SecVulEval: Benchmarking LLMs for Real-World C/C++ Vulnerability Detection](https://arxiv.org/abs/2505.19828), Md Basim Uddin Ahmed, Nima Shiri Harzevili, Jiho Shin, Hung Viet Pham, Song Wang; arXiv submitted 2025-05-26. Updated conference title [SecVulEval: Context-Aware Benchmarking of LLMs for Vulnerability Detection](https://2026.aiwareconf.org/details/aiware-2026-benchmark---dataset-track/11/SecVulEval-Context-Aware-Benchmarking-of-LLMs-for-Vulnerability-Detection), AIware 2026 Benchmark & Dataset Track; DOI 10.1145/3805760.3814932. | 25,440 C/C++ functions, 5,867 CVEs, 1999–2024; statement-level labels and contextual data; standalone and multi-agent LLM evaluation; statement F1 (best agent 23.83%). | One of TraceGRPO's data sources and the source of line-level supervision. We independently repaired the project's known file-absolute line mapping per local audit. | Human validation of diff-derived labels in this project; cross-paper comparability of line-level scores; whether model-generated traces faithfully identify these lines. |
+| [SecCodePRM: A Process Reward Model for Code Security](https://arxiv.org/abs/2602.10418), Weichen Yu, Ravi Mangal, Yinyi Luo, Kai Hu, Jingxuan He, Corina S. Pasareanu, Matt Fredrikson; v1 submitted 2026-02-11; arXiv preprint, venue unverified (ICML appears as a keyword only). | Qwen2.5-Coder-7B-Instruct plus classification head; step boundaries are blank-line separators with AST-aware merging. The method labels changed vulnerable steps from vulnerable/fixed alignment and labels fixed-function steps safe; trains weighted step-level cross entropy, then aggregates low-safety steps for function detection and continuation ranking. Its table reports PrimeVul paired train/test sizes of 7,305/854 pairs. Reported full-code paired PrimeVul results: 96.83% accuracy and 96.73 F1; pair table gives 93.66 P-C, 0 P-V, 6.34 P-B, 0 P-R. The paper also reports full-code SVEN only on four CWEs, and partial-code results across SVEN, PrimeVul, ReposVul, PreciseBugs. It excludes BigVul and PrimeVul-unpaired from training because of a roughly 3× vulnerable/safe token-length gap. The source does not establish that the paired PrimeVul test result uses a chronological or repository-disjoint split; treat metrics as the authors' reported result, not directly comparable to our 2022 chronological VALID. | Directly establishes that dense segment-level security scoring and risk-weighted aggregation are prior art; those generic ideas are not novel. TraceGRPO differs in generating structured evidence/trace and optimizing a policy against frozen cached verifiers, rather than classifying every code segment and ranking candidate code continuations. | Matched chronological row-AUC and pair-cluster uncertainty; controls for repository/time leakage and residual patch/length shortcuts; whether structured emitted evidence is faithful; and whether reward-trained policy updates improve discrimination beyond a step-scoring classifier. |
+| [Do Fine-Tuned LLMs Understand Vulnerabilities? An Investigation into the Semantic Trap](https://arxiv.org/abs/2601.22655), Feiyang Huang, Yuqiang Sun, Fan Zhang, Ziqi Yang, Han Liu, Yang Liu; submitted 2026-01-30, v3 2026-05-21; arXiv preprint (venue not verified). | TrapEval uses vulnerable–patched V2P and vulnerable–unrelated-normal V2N pairs, plus semantic perturbations and CodeBLEU gap analysis; evaluates five LLMs with and without reasoning SFT. | Direct evidence that fine-tuned models can exploit pair/textual-gap cues and remain fragile to semantics-preserving changes; substantially overlaps our pair-shortcut concern. | Whether TraceGRPO has these same perturbation failures, especially at row AUC and across long-function strata; E3 alone does not test them. |
+| [Top Score on the Wrong Exam: On Benchmarking in Machine Learning for Vulnerability Detection](https://doi.org/10.1145/3728887), Niklas Risse, Jing Liu, Marcel Böhme; ISSTA 2025 / Proc. ACM Softw. Eng., DOI verified. | Survey and analyses of popular ML4VD datasets; manual context audit and lexical/word-count baselines; examines whether function labels are decidable without program context. | Foundational warning about ill-defined function-level labels, contextual dependence, and spurious correlations. It directly motivates our leakage, size, source, and pair-shortcut controls; a high row metric alone is insufficient. | Whether our held-out SecVulEval cases contain sufficient function-local evidence; human audit and contextual sufficiency analysis remain needed. |
+
+## Line localization and graph/program-analysis priors
+
+| Work | Verified source/status | Contribution and overlap | Remaining gap relevant to E3 |
+|---|---|---|---|
+| LineVul | [Fu & Tantithamthavorn, MSR 2022, DOI 10.1145/3524842.3528452](https://doi.org/10.1145/3524842.3528452) | Transformer-based line-level vulnerability prediction, with function and line ranking. Establishes that line-level neural localization is prior art. | Different benchmark/split and metrics; does not establish paired chronological row-AUC gains from evidence-first scoring. |
+| LineVD | [Hin, Kan, Chen, Babar, arXiv:2203.05181](https://arxiv.org/abs/2203.05181) | Statement-level node classification using code tokens plus control/data dependency GNNs; direct prior for line/statement scoring. | E3's line-head architecture and pair-balanced time split need explicit baseline comparison; “line scoring” alone is not novel. |
+| IVDetect | [Li, Wang & Nguyen, arXiv:2106.10478](https://arxiv.org/abs/2106.10478) | PDG-based detector with vulnerability-relevant statements and dependency context; prior art for evidence subgraphs and localization. | Does not address the present LLM post-training failure or matched forward-time pair-AUC protocol. |
+| DeepDFA | [Steenhoek, Gao & Le, arXiv:2212.08108](https://arxiv.org/abs/2212.08108) | Data-flow-analysis-inspired graph learning that encodes abstract dataflow semantics; evaluates BigVul and DbgBench and hybridizes with an LLM. | E3 is not a substitute for actual graph-semantic modeling; our line head only uses line representations and diff targets. Compare baselines before novelty claims. |
+| LLMxCPG | Ahmed Lekssays, Hamza Mouhcine, Khang Tran, Ting Yu, Issa Khalil; [USENIX Security 2025](https://www.usenix.org/conference/usenixsecurity25/presentation/lekssays). | Builds CPG-guided slices for LLM vulnerability detection; authors report reducing input size 67.84–90.93% while preserving context, and 15–40% F1 gains over their baselines. Direct prior for E1's slice-and-classify idea. | E1's predicted local windows and OOF training differ in slice construction; we still need same-split AUC evaluation and must avoid a “first slicing” claim. |
+| Trace Gadgets: Minimizing Code Context for Machine Learning-Based Vulnerability Prediction | Felix Mächtle et al.; [arXiv:2504.13676](https://arxiv.org/abs/2504.13676), v2 2025-12-16; published ACM AsiaCCS 2026, DOI 10.1145/3779208.3785375. | Minimal, path-complete vulnerability context; reports unseen-data evaluation and comparisons with CodeQL. Direct prior for E1's context-minimization motivation. | E1's fixed windows are not path-complete and do not reproduce this representation; results are not temporally comparable. |
+| SliceLocator: Locating Vulnerable Statements with Graph-based Detectors | Baijun Cheng et al.; [arXiv:2401.02737](https://arxiv.org/abs/2401.02737), v4 revised 2025-02-23; venue not verified. | Ranks candidate taint-flow paths with graph-detector scores; reports about 87% accuracy for vulnerability-triggering statement localization across six common C/C++ vulnerability categories. | Localization is prior art; that metric is statement localization, not row AUC, and must not be compared as such. |
+| AEGIS: From Clues to Verdicts -- Graph-Guided Deep Vulnerability Reasoning via Dialectics and Meta-Auditing | Sen Fang, Weiyuan Ding, Zhezhen Cao, Zhou Yang, Bowen Xu; [arXiv:2603.20637](https://arxiv.org/abs/2603.20637), submitted 2026-03-21; venue not verified. | Builds per-variable dependency chains with on-demand repository CPG slicing, then runs competing verifier arguments and an independent auditor; reports PrimeVul P-C results. | Strong overlap with grounded evidence and reasoning verification. It is agentic and repository-context-dependent rather than a single line-score classifier; numbers remain incomparable without aligned splits. |
+
+## Verified additions and corrections from the Review #22 literature list
+
+| Work | Primary-source verification | Contribution and protocol facts verified | Overlap / caveat |
+|---|---|---|---|
+| R2Vul: Learning to Reason about Software Vulnerabilities with Reinforcement Learning and Structured Reasoning Distillation | [arXiv:2504.04699](https://arxiv.org/abs/2504.04699), Martin Weyssow, Chengran Yang, Junkai Chen, Ratnadira Widyasari, Ting Zhang, Huihui Huang, Huu Hung Nguyen, Yan Naing Tun, Tan Bui, Yikun Li, Ang Han Wei, Frank Liauw, Eng Lieh Ouh, Lwin Khin Shar, David Lo; submitted 2025-04-07, v2 revised 2025-08-07; venue not stated on the record. | RLAIF plus structured reasoning distillation, explicitly rewarding well-founded rather than merely plausible vulnerability explanations; 18,000 multilingual preference examples across C#, JavaScript, Java, Python, C; evaluates against static analyzers, LLMs and tuning methods. | Strong overlap with reasoning + feedback; exact dataset split and numeric result tables need PDF audit before quantitative citation. |
+| Improving LLM Reasoning for Vulnerability Detection via Group Relative Policy Optimization | [arXiv:2507.03051](https://arxiv.org/abs/2507.03051), v1 submitted 2025-07-03; Marco Simoni, Aleksandar Fontana, Giulio Rossolini, Andrea Saracino; record says Under Review, venue unverified. | GRPO with structured rule-based rewards and redefined advantage/reward signals from BigVul, DiverseVul and CleanVul annotations; investigates generalization, reasoning and SFT comparisons. | Directly means vulnerability-detection GRPO is prior art. Exact split, metrics and implementation details require full-text inspection before numeric comparisons. |
+| Learning to Focus: Context Extraction for Efficient Code Vulnerability Detection with Language Models (FocusVul) | [arXiv:2505.17460](https://arxiv.org/abs/2505.17460), v3, withdrawn by an author; arXiv states “Due to fundamental errors in the methodology, I have to withdraw this paper.” | The withdrawn manuscript described commit-supervised sensitive-region selection and dependency/execution-flow context extraction. | Do not cite performance claims as validated evidence or use it as a published-prior comparison. It can only be mentioned as a withdrawn preprint, with the withdrawal explicit. |
+| SecCodePRM: A Process Reward Model for Code Security | [arXiv:2602.10418](https://arxiv.org/abs/2602.10418), submitted 2026-02-11; Weichen Yu et al.; arXiv preprint, venue unverified. | Step-level context-aware security scoring, supervision from static analyzers and experts; risk-sensitive aggregation for full/partial-code detection and ranking generated continuations. | Process-level security scoring is prior art. Dataset and benchmark-specific numerical claims still require full-text table/split audit; no direct apples-to-apples claim is justified. |
+| LLMxCPG | [Official USENIX Security 2025 page](https://www.usenix.org/conference/usenixsecurity25/presentation/lekssays). | CPG-guided code slicing for LLM vulnerability analysis; official page is a primary publication source. | Confirms context slicing is established. Exact reported reduction/F1 ranges should be cited only after checking the paper itself, not inferred from the talk page. |
+| Trace Gadgets: Minimizing Code Context for Machine Learning-Based Vulnerability Prediction | Felix Mächtle, Nils Loose, Tim Schulz, Florian Sieck, Jan-Niclas Serr, Ralf Möller, Thomas Eisenbarth; [arXiv:2504.13676](https://arxiv.org/abs/2504.13676), v2 revised 2025-12-16; published ACM AsiaCCS 2026, DOI 10.1145/3779208.3785375. | Minimal, path-complete vulnerability context; reports unseen-data evaluation and comparisons with CodeQL. | E1's fixed windows are not path-complete; metrics and splits are not temporally comparable. |
+| SliceLocator: Locating Vulnerable Statements with Graph-based Detectors | Baijun Cheng, Kailong Wang, Cuiyun Gao, Xiapu Luo, Li Li, Yao Guo, Xiangqun Chen, Haoyu Wang; [arXiv:2401.02737](https://arxiv.org/abs/2401.02737), submitted 2024-01-05, v4 revised 2025-02-23; venue not verified. | Ranks candidate taint-flow paths with graph-detector scores; reports about 87% accuracy for vulnerability-triggering statement localization across six common C/C++ vulnerability categories. | Localization is prior art; that metric is statement localization, not row AUC, and must not be compared as such. |
+| AEGIS: From Clues to Verdicts—Graph-Guided Deep Vulnerability Reasoning via Dialectics and Meta-Auditing | [arXiv:2603.20637](https://arxiv.org/abs/2603.20637), submitted 2026-03-21; venue not verified. | Graph-guided clue extraction, competing arguments and a separate audit agent. | Overlaps evidence-grounded reasoning and consistency auditing. Numeric benchmark claims are not yet cleared for citation without full-text table/protocol verification. |
+| DAPO: An Open-Source LLM Reinforcement Learning System at Scale | [arXiv:2503.14476](https://arxiv.org/abs/2503.14476), submitted 2025-03-18; venue not verified here. | Proposes decoupled clipping, dynamic sampling, token-level policy-gradient loss, and overlong-sequence reward shaping; evaluates math reasoning using Qwen2.5-32B on AIME 2024. | These are training-stability/sample-efficiency techniques, not a remedy for biased labels or invalid localization rewards. A DAPO port would change several variables and is not authorized by Review #23. |
+| Understanding R1-Zero-Like Training: A Critical Perspective (Dr. GRPO) | [arXiv:2503.20783](https://arxiv.org/abs/2503.20783); ICML 2025 AI for Math Workshop Best Paper Honorable Mention (author-maintained project record). | Analyzes base-model priors and GRPO's response-length/difficulty weighting; Dr. GRPO removes length normalization and adjusts group scaling. Experiments concern math reasoning, including Qwen2.5-based runs. | Relevant algorithmic caveat for token-length weighting, but it does not explain the observed class-prior shift in this project's stored counterfactuals. Any optimizer change would require a separately authorized controlled experiment. |
+| Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning | [arXiv:2410.08146](https://arxiv.org/abs/2410.08146), submitted 2024-10-10; venue not verified here. | Defines process advantage verifiers to estimate how a reasoning step changes the probability of eventual correctness under a distinct prover; studies verifier-guided search and dense-reward online RL on general reasoning benchmarks. | Motivates progress-based credit assignment instead of matching trace overlap. Transfer to code security would require a calibrated vulnerability verifier and leakage-safe labels; it does not validate current R3. |
+| LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking | [arXiv:2604.15149](https://arxiv.org/abs/2604.15149), submitted 2026-04-16; venue not verified here. | On inductive rule-learning tasks, shows RLVR can favor example-specific outputs that satisfy extensional checks while failing isomorphic tasks; proposes Isomorphic Perturbation Testing and demonstrates verifier changes affect shortcut learning. | Strong methodological precedent for perturbation-based anti-hacking evaluation. For vulnerability detection, the analogue would preserve vulnerability semantics while changing superficial patch/length/context cues, then compare detector predictions and verifier scores. |
+
+The withdrawn status of FocusVul is verified and should remain prominent. SecCodePRM's method, reported PrimeVul paired counts and table values have now been checked against the original arXiv HTML; its split provenance is still not established in the accessible protocol, so those numbers must not be presented as a matched comparison. This audit does not verify every numeric result in every work: unverified items are labeled in the table and must not enter a publication draft as established facts.
+
+## Novelty implications (interim; no publication framing decision)
+
+1. A generic “Qwen3-4B + cold-start SFT + GRPO with label/localization/reasoning rewards” contribution is not novel in light of VULPO and Simoni et al. Structured reasoning feedback is also present in R2Vul. Graph-grounded verifiable reward is not novel in light of VulAgentRL. Vulnerability segment/process reward is not novel in light of SecCodePRM. Line-level localization and PDG/DFG evidence are established in LineVul, LineVD, IVDetect, and DeepDFA.
+2. Potentially defensible work lies in a controlled failure/mechanism study: fixed-output counterfactuals that isolate label weighting from empty-set conventions; threshold-free, pair-clustered chronological evaluation; and testing an evidence-first line-scoring detector under long-function strata against matched classical and shortcut controls. These remain empirical claims to be earned by E1–E3, not novelty claims established by this literature audit.
+3. SecCodePRM's reported PrimeVul numbers make split/protocol documentation especially important. Its result cannot be contrasted with ours until the exact data version, split, deduplication, and prompt are matched.

results/phase6/scale_review21_step324_lr5e6/curves.json

@@ -20803,5 +20803,225 @@
     "entropy_nats": 0.5321434676768195,
     "entropy_completion_tokens": 72407,
     "completed_at": "2026-10-09T13:50:11.607610+00:00"
+  },
+  {
+    "optimizer_step": 332,
+    "adapter_step": 331,
+    "n_rollouts": 512,
+    "groups": 64,
+    "zero_groups": 0,
+    "zero_variance_fraction": 0.0,
+    "parse_rate": 0.9921875,
+    "verifier_failures": 4,
+    "step_failure_rate": 0.0078125,
+    "label_accuracy": 0.533203125,
+    "reward_component_means": {
+      "R1": 0.07421875,
+      "R2": 0.24101022533402044,
+      "R3": 0.029607451763140388,
+      "R4": 0.48046875,
+      "R5": -0.042578125
+    },
+    "completion_tokens": {
+      "mean": 136.076171875,
+      "p50": 127.0,
+      "p95": 259.9,
+      "max": 646
+    },
+    "failure_causes": {
+      "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 2,
+      "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 11, column 43": 1,
+      "parser_failure:contract_v2: ParseError: mismatched tag: line 2, column 34": 1
+    },
+    "generated_at": "2026-10-09T13:52:45.304479+00:00",
+    "aggregate_parse_rate": 0.9866517633908874,
+    "verifier_window_failure_rate": 0.0111328125,
+    "verifier_window_failures": 114,
+    "verifier_window_rollouts": 10240,
+    "canary_in_failure_window": false,
+    "canaries_in_failure_window": 0,
+    "all_generated_rollouts_including_canary": 169985,
+    "dgrpo6b_failure_decision": {
+      "source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)",
+      "rolling_failure_rate": 0.0111328125,
+      "rolling_window_steps": 20,
+      "nonparser_rolling_rate": 9.765625e-05,
+      "completed_block_wilson_lowers": [
+        0.018117845054430304,
+        0.016752423051967957,
+        0.013668988573899227,
+        0.014120917591715385,
+        0.013307864394505934,
+        0.011239402594480075,
+        0.009631719614628635,
+        0.008920953010197434,
+        0.010970713591566107,
+        0.012047080241744917,
+        0.009631719614628635,
+        0.009631719614628635,
+        0.008743666775120913,
+        0.007244158561402301,
+        0.009542737334915601,
+        0.009098405671651343
+      ],
+      "warning": false,
+      "stop": false,
+      "stop_reason": null
+    },
+    "learning_rate_change_step": 325,
+    "base_learning_rate": 5e-06,
+    "learning_rate_source": "D-GRPO-15 / Review #21",
+    "effective_learning_rate": 4.858279533144357e-06,
+    "entropy_nats": 0.5333679221763303,
+    "entropy_completion_tokens": 69671,
+    "completed_at": "2026-10-09T13:56:32.691748+00:00"
+  },
+  {
+    "optimizer_step": 333,
+    "adapter_step": 332,
+    "n_rollouts": 512,
+    "groups": 64,
+    "zero_groups": 0,
+    "zero_variance_fraction": 0.0,
+    "parse_rate": 0.98828125,
+    "verifier_failures": 6,
+    "step_failure_rate": 0.01171875,
+    "label_accuracy": 0.45703125,
+    "reward_component_means": {
+      "R1": -0.07421875,
+      "R2": 0.21056271525827627,
+      "R3": 0.020094086021505374,
+      "R4": 0.423828125,
+      "R5": -0.062890625
+    },
+    "completion_tokens": {
+      "mean": 137.5703125,
+      "p50": 127.0,
+      "p95": 275.9,
+      "max": 754
+    },
+    "failure_causes": {
+      "parser_failure:contract_v2: ParseError: mismatched tag: line 16, column 2": 1,
+      "parser_failure:contract_v2: ValueError: 'CWE-overflow' does not match '^(CWE-[0-9]+|NONE)$'": 1,
+      "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 2,
+      "parser_failure:contract_v2: ParseError: mismatched tag: line 20, column 2": 1,
+      "parser_failure:contract_v2: ValueError: 'CWE-59)' does not match '^(CWE-[0-9]+|NONE)$'": 1
+    },
+    "generated_at": "2026-10-09T14:00:45.381089+00:00",
+    "aggregate_parse_rate": 0.9866566567153674,
+    "verifier_window_failure_rate": 0.0115234375,
+    "verifier_window_failures": 118,
+    "verifier_window_rollouts": 10240,
+    "canary_in_failure_window": false,
+    "canaries_in_failure_window": 0,
+    "all_generated_rollouts_including_canary": 170497,
+    "dgrpo6b_failure_decision": {
+      "source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)",
+      "rolling_failure_rate": 0.0115234375,
+      "rolling_window_steps": 20,
+      "nonparser_rolling_rate": 9.765625e-05,
+      "completed_block_wilson_lowers": [
+        0.018117845054430304,
+        0.016752423051967957,
+        0.013668988573899227,
+        0.014120917591715385,
+        0.013307864394505934,
+        0.011239402594480075,
+        0.009631719614628635,
+        0.008920953010197434,
+        0.010970713591566107,
+        0.012047080241744917,
+        0.009631719614628635,
+        0.009631719614628635,
+        0.008743666775120913,
+        0.007244158561402301,
+        0.009542737334915601,
+        0.009098405671651343
+      ],
+      "warning": false,
+      "stop": false,
+      "stop_reason": null
+    },
+    "learning_rate_change_step": 325,
+    "base_learning_rate": 5e-06,
+    "learning_rate_source": "D-GRPO-15 / Review #21",
+    "effective_learning_rate": 4.857406392712601e-06,
+    "entropy_nats": 0.5398191982927094,
+    "entropy_completion_tokens": 70436,
+    "completed_at": "2026-10-09T14:04:44.345485+00:00"
+  },
+  {
+    "optimizer_step": 334,
+    "adapter_step": 333,
+    "n_rollouts": 512,
+    "groups": 64,
+    "zero_groups": 0,
+    "zero_variance_fraction": 0.0,
+    "parse_rate": 0.98828125,
+    "verifier_failures": 6,
+    "step_failure_rate": 0.01171875,
+    "label_accuracy": 0.486328125,
+    "reward_component_means": {
+      "R1": -0.015625,
+      "R2": 0.2463194344420451,
+      "R3": 0.024276377217553685,
+      "R4": 0.47265625,
+      "R5": -0.058007812500000006
+    },
+    "completion_tokens": {
+      "mean": 142.55859375,
+      "p50": 125.0,
+      "p95": 288.34999999999997,
+      "max": 1073
+    },
+    "failure_causes": {
+      "parser_failure:contract_v2: ValueError: 'CWE-125,CWE-119' does not match '^(CWE-[0-9]+|NONE)$'": 1,
+      "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 2, column 59": 1,
+      "parser_failure:contract_v2: ValueError: '' should be non-empty": 2,
+      "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 5, column 114": 1,
+      "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 1
+    },
+    "generated_at": "2026-10-09T14:07:48.246021+00:00",
+    "aggregate_parse_rate": 0.9866615207386745,
+    "verifier_window_failure_rate": 0.01142578125,
+    "verifier_window_failures": 117,
+    "verifier_window_rollouts": 10240,
+    "canary_in_failure_window": false,
+    "canaries_in_failure_window": 0,
+    "all_generated_rollouts_including_canary": 171009,
+    "dgrpo6b_failure_decision": {
+      "source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)",
+      "rolling_failure_rate": 0.01142578125,
+      "rolling_window_steps": 20,
+      "nonparser_rolling_rate": 9.765625e-05,
+      "completed_block_wilson_lowers": [
+        0.018117845054430304,
+        0.016752423051967957,
+        0.013668988573899227,
+        0.014120917591715385,
+        0.013307864394505934,
+        0.011239402594480075,
+        0.009631719614628635,
+        0.008920953010197434,
+        0.010970713591566107,
+        0.012047080241744917,
+        0.009631719614628635,
+        0.009631719614628635,
+        0.008743666775120913,
+        0.007244158561402301,
+        0.009542737334915601,
+        0.009098405671651343
+      ],
+      "warning": false,
+      "stop": false,
+      "stop_reason": null
+    },
+    "learning_rate_change_step": 325,
+    "base_learning_rate": 5e-06,
+    "learning_rate_source": "D-GRPO-15 / Review #21",
+    "effective_learning_rate": 4.856530649774567e-06,
+    "entropy_nats": 0.5097311908977622,
+    "entropy_completion_tokens": 72990,
+    "completed_at": "2026-10-09T14:11:51.503238+00:00"
   }
 ]

results/phase6/scale_review21_step324_lr5e6/judge_outputs.jsonl

@@ -205,3 +205,62 @@
 {"prompt_id": "secvuleval:CVE-2014-2309:c88507fbad8055297c1d1e21e599f46960cbee39:256:bcd7aa12a3aea734", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340389, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:32.758218+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
 {"prompt_id": "primevul:CVE-2019-3558:c5d6e07588cd03061bc54d451a7fa6e84883d62b:197137:f014f6566ce41a58", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340399, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:33.800216+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
 {"prompt_id": "secvuleval:CVE-2013-1943:fa3d315a4ce2c0891cdde262562e710d95fba19e:3870:96188ec0996f945f", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340403, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:34.124090+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8952:82939d7999dfc1f1998c4b1c12e2f19edbdff272:5083:fe23d379c177e2fe", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340436, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:36.952127+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2012-3552:f6d8bd051c391c1c0458a30b2a7abcd939329259:3816:290b4448e79e726d", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340487, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:39.791740+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2023-42801:f57bd745b4cbed577ea654fad4701bea4d38b44c:18795:9e524c0d73dfcbae", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340513, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:41.648277+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2017-5970:34b2cef20f19c87999fff3da4071e66937db9644:2273:9a1f785d81154097", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340517, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:41.881629+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-1066:cabfb3680f78981d26c078a26e5c748531257ebb:2848:644349604e19d98e", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340546, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:44.895350+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-1066:cabfb3680f78981d26c078a26e5c748531257ebb:2848:644349604e19d98e", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 331, "generation_seed": 334340553, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:52:45.247886+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-27675:073d0552ead5bfc7a3a9c01de590e924f11b5dd2:840:3e360f0b049fa5a8", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350058, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:53.131219+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2011-2491:0b760113a3a155269a3fba93a409c640031dd68f:3883:71e48a5c2659e40a", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350068, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:53.578984+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2021-3753:2287a51ba822384834dafc1c798453375d1107c7:338:f51d22f2c2262d03", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350084, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:54.475041+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-26759:3e990547e064b8fca916b23a0f7d6fe8c63c7f6b:16412:55dc4f31147ab55f", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350095, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:56.131001+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:6cf97230cd5f36b7665099083272595c55d72be7:201386:e50b5109c97ce17f", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350107, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:57.159204+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:6cf97230cd5f36b7665099083272595c55d72be7:201386:e50b5109c97ce17f", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350109, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:57.249440+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2021-43114:274dc14aed1eb9b3350029d1063578a6b9c77b54:17822:95df2aa0c4ad1521", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350144, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:59.125989+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2016-4581:5ec0811d30378ae104f250bfc9b3640242d81e3f:5240:cd1755e8aec8109f", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350156, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T13:59:59.868850+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2016-4997:ce683e5f9d045e5d67d1312a42b359cb2ab2a13c:5194:e3ef6d66f37f897e", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350167, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:00.550624+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-15211:1970c2158b1ffa416d159d03c3370b9a462aee35:12802:479633d6a06f7923", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350248, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:07.487736+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-22873:430ea27accd4d4ffddc946c9402e7c9064835a18:16193:392ab972b26b8a26", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350287, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:13.276969+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2012-2891:116d0963cadfbf55ef2ec3d13781987c4d80517a:21070:defaae510b88731d", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350398, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:32.056637+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:3619c170461a3107a258d1fd2d00ed4832adb1b1:210954:79295cfb80a0e8f6", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350439, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:37.009197+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-15211:1970c2158b1ffa416d159d03c3370b9a462aee35:13017:037226db412a041b", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350474, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:39.199778+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2021-26788:de5336016edbe1e90327d0ed1cba5c4e49114366:17432:fb81770e84160d1a", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350500, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:40.764908+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8952:82939d7999dfc1f1998c4b1c12e2f19edbdff272:5080:9b541ee45bd7bb4c", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350509, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:41.419459+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8952:82939d7999dfc1f1998c4b1c12e2f19edbdff272:5080:9b541ee45bd7bb4c", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350512, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:41.629227+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2017-12595:ad527a64f93dca12f6aabab2ca99ae5eb352ab4b:199167:9645c6899f00375e", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350515, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:41.938382+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2016-10190:2a05c8f813de6f2278827734bf8102291e7484aa:6433:710d258214d57591", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 332, "generation_seed": 335350529, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:00:43.960747+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2018-10017:492022c7297ede682161d9c0ec2de15526424e76:10607:9583c00491aa1a3a", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360055, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:01.390588+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:e0485802ec72996c20026da320650d8362f555bd:204016:ea6e76bc9c8e753b", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360062, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:03.683028+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:e0485802ec72996c20026da320650d8362f555bd:204016:ea6e76bc9c8e753b", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360064, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:03.884926+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-15211:1970c2158b1ffa416d159d03c3370b9a462aee35:12949:24f57382cc2a3f3d", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360068, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:04.174895+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-20761:35ab4475a7df9b2a4bcab235e379c0c3ec543658:11604:2b9624e5a747d2b1", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360090, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:06.528392+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-6111:3c8e4852477d5b1e2da877808c998dc57db9460f:23252:5423f7a1b24a1409", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360101, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:08.183332+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-6111:3c8e4852477d5b1e2da877808c998dc57db9460f:23255:36f224a65b53c9f7", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360117, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:08.899484+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2015-1302:fff450abc4e2fb330ba700547a8e6a7b0fb90a6e:6749:66b38482f34a449c", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360127, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:09.465342+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2012-1571:1859fdb4e67c49c463c4e0078054335cd46ba295:6724:fd770e0b9f9e1671", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360140, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:10.833433+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2019-19481:b75c002cfb1fd61cd20ec938ff4937d7b1a94278:12121:1e98e06b47073acd", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360165, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:12.158499+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2016-5222:e4ebe078840e65d673722e94f8251b334030b5e8:7256:3c4b17afdfe4b7a6", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360174, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:12.870530+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2017-12994:ffde45acf3348f8353fb4064a1b21683ee6b5ddf:2599:96e96c9721216f61", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360179, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:13.844334+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2020-11089:6b485b146a1b9d6ce72dfd7b5f36456c166e7a16:7320:c7a35eef57da06c3", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360195, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:18.924899+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:fbe2837bc80f155c001781041a54c58b5524fc14:197061:f81d4933fd3a3a85", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360211, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:19.694241+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:fbe2837bc80f155c001781041a54c58b5524fc14:197061:f81d4933fd3a3a85", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360214, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:20.073112+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2016-5164:93bc623489bdcfc7e9127614fcfb3258edf3f0f9:22332:17ec301edab36a45", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360261, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:25.060777+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2013-1774:1ee0a224bc9aad1de496c795f96bc6ba2c394811:1075:33c8d63c48952d34", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360268, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:25.577365+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2013-1774:1ee0a224bc9aad1de496c795f96bc6ba2c394811:1075:33c8d63c48952d34", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360270, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:25.695285+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2011-2918:a8b0ca17b80e92faab46ee7179ba9e99ccb61233:4012:54f35dcfd9f7c7fe", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360315, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:29.146276+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2011-2918:a8b0ca17b80e92faab46ee7179ba9e99ccb61233:4012:54f35dcfd9f7c7fe", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360317, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:29.242164+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2019-15694:0943c006c7d900dfc0281639e992791d6c567438:14808:ea9ddc25ae2ce9f3", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360372, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:35.794416+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2013-1767:5f00110f7273f9ff04ac69a5f85bb535a4fd0987:9246:ea9637ec0ae9cff9", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360380, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:36.229883+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2013-1767:5f00110f7273f9ff04ac69a5f85bb535a4fd0987:9246:ea9637ec0ae9cff9", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360382, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:36.340120+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2016-5149:53262b5285efbbfdbbcc17fbf5200890f1ee295c:22311:16cf46b435176a80", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360386, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:36.564240+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8958:b8f17d08b7418204bf8a05a5c24e87b2fc395b75:19861:d29a699bcd9dad11", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360419, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:38.462206+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8958:b8f17d08b7418204bf8a05a5c24e87b2fc395b75:19861:d29a699bcd9dad11", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360421, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:38.933022+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2015-8958:b8f17d08b7418204bf8a05a5c24e87b2fc395b75:19861:d29a699bcd9dad11", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360423, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:39.412542+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:605ec91e5a7330d61be313637e495fa02a6dc264:212145:1d004c4256e9f06b", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360434, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:40.571044+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2018-6057:c0c8978849ac57e4ecd613ddc8ff7852a2054734:23406:93d672fd873d6b30", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360447, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:41.298938+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:NOCVE:104c307147ad379617472dd91a5bcb368d72bd6d:204406:43ee1d1d1ee75e63", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360466, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:42.545602+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2014-6262:64ed5314af1255ab6dded45f70b39cdeab5ae2ec:8091:5524e77de51d4f7c", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360481, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:43.187592+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2019-13391:f6ffc702c6eecd963587273a429dcd608c648984:20088:ad10074f2459fbb3", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360484, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:43.596632+00:00", "raw": "1", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "secvuleval:CVE-2019-13391:f6ffc702c6eecd963587273a429dcd608c648984:20088:ad10074f2459fbb3", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360485, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:43.775642+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}
+{"prompt_id": "primevul:CVE-2010-0011:1958b52d41cba96956dc1995660de49525ed1047:568:c7a4eaccd72cbcb1", "dataset_snapshot_hash": "da8b90e775e7328b20da8464beac1b9c36b3dfd59f477e75ba7896923a9f92ce", "adapter_step": 333, "generation_seed": 336360535, "tool_versions": {"joern": {"used_per_rollout": false, "installed_version": "4.0.645", "jar_sha256": "12dd1643761ce3cb4bc7a51e973f7fafa449a4a40af6750a1c56d426a297f240"}, "tree_sitter_grammar_hashes": {"tree_sitter_c": "800dac4dea22edf6e342326c7bd11d3c0ada8b64af2259110bafa839429466c9", "tree_sitter_cpp": "99d80b33712cb43559cd5d35d59c57f0c323bae496c863e5e3a2a8746824fe5c"}, "packages": {"torch": "2.8.0", "transformers": "4.56.2", "trl": "0.21.0", "vllm": "0.10.2", "peft": "0.21.2", "tree-sitter": "0.26.0", "tree-sitter-c": "0.24.2", "tree-sitter-cpp": "0.23.4"}}, "wall_timestamp": "2026-10-09T14:07:47.349907+00:00", "raw": "0", "model_revision": "c03e6d358207e414f1eca0bb1891e29f1db0e242"}

results/phase6/scale_review21_step324_lr5e6/progress.json

@@ -1,7 +1,9 @@
 {
-  "phase": "scoring_rollouts",
-  "completed_steps": 331,
-  "elapsed_gpu_wall_hours": 1.0142274908224742,
-  "updated_at": "2026-10-09T13:52:01.703288+00:00",
-  "adapter_step": 331
+  "phase": "optimizer_step_completed",
+  "completed_steps": 334,
+  "elapsed_gpu_wall_hours": 1.3447383081250721,
+  "updated_at": "2026-10-09T14:11:51.542222+00:00",
+  "entropy_nats": 0.5097311908977622,
+  "parse_rate": 0.98828125,
+  "zero_variance_fraction": 0.0
 }

results/phase6/scale_review21_step324_lr5e6/scale.log

@@ -94,3 +94,15 @@ PHASE optimizer_step_completed {'entropy_nats': 0.5321434676768195, 'parse_rate'
 INFO 10-09 13:51:41 [block_pool.py:292] Successfully reset prefix cache
 WARNING 10-09 13:51:41 [tokenizer.py:286] No tokenizer found in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000331, using base model tokenizer instead. (Exception: Unrecognized model in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000331. Should have a `model_type` key in its config.json, or contain one of the following strings in its name: aimv2, aimv2_vision_model, albert, align, altclip, apertus, arcee, aria, aria_text, audio-spectrogram-transformer, autoformer, aya_vision, bamba, bark, bart, beit, bert, bert-generation, big_bird, bigbird_pegasus, biogpt, bit, bitnet, blenderbot, blenderbot-small, blip, blip-2, blip_2_qformer, bloom, bridgetower, bros, camembert, canine, chameleon, chinese_clip, chinese_clip_vision_model, clap, clip, clip_text_model, clip_vision_model, clipseg, clvp, code_llama, codegen, cohere, cohere2, cohere2_vision, colpali, colqwen2, conditional_detr, convbert, convnext, convnextv2, cpmant, csm, ctrl, cvt, d_fine, dab-detr, dac, data2vec-audio, data2vec-text, data2vec-vision, dbrx, deberta, deberta-v2, decision_transformer, deepseek_v2, deepseek_v3, deepseek_vl, deepseek_vl_hybrid, deformable_detr, deit, depth_anything, depth_pro, deta, detr, dia, diffllama, dinat, dinov2, dinov2_with_registers, dinov3_convnext, dinov3_vit, distilbert, doge, donut-swin, dots1, dpr, dpt, efficientformer, efficientloftr, efficientnet, electra, emu3, encodec, encoder-decoder, eomt, ernie, ernie4_5, ernie4_5_moe, ernie_m, esm, evolla, exaone4, falcon, falcon_h1, falcon_mamba, fastspeech2_conformer, fastspeech2_conformer_with_hifigan, flaubert, flava, florence2, fnet, focalnet, fsmt, funnel, fuyu, gemma, gemma2, gemma3, gemma3_text, gemma3n, gemma3n_audio, gemma3n_text, gemma3n_vision, git, glm, glm4, glm4_moe, glm4v, glm4v_moe, glm4v_moe_text, glm4v_text, glpn, got_ocr2, gpt-sw3, gpt2, gpt_bigcode, gpt_neo, gpt_neox, gpt_neox_japanese, gpt_oss, gptj, gptsan-japanese, granite, granite_speech, granitemoe, granitemoehybrid, granitemoeshared, granitevision, graphormer, grounding-dino, groupvit, helium, hgnet_v2, hiera, hubert, hunyuan_v1_dense, hunyuan_v1_moe, ibert, idefics, idefics2, idefics3, idefics3_vision, ijepa, imagegpt, informer, instructblip, instructblipvideo, internvl, internvl_vision, jamba, janus, jetmoe, jukebox, kosmos-2, kosmos-2.5, kyutai_speech_to_text, layoutlm, layoutlmv2, layoutlmv3, led, levit, lfm2, lightglue, lilt, llama, llama4, llama4_text, llava, llava_next, llava_next_video, llava_onevision, longformer, longt5, luke, lxmert, m2m_100, mamba, mamba2, marian, markuplm, mask2former, maskformer, maskformer-swin, mbart, mctct, mega, megatron-bert, metaclip_2, mgp-str, mimi, minimax, mistral, mistral3, mixtral, mlcd, mllama, mm-grounding-dino, mobilebert, mobilenet_v1, mobilenet_v2, mobilevit, mobilevitv2, modernbert, modernbert-decoder, moonshine, moshi, mpnet, mpt, mra, mt5, musicgen, musicgen_melody, mvp, nat, nemotron, nezha, nllb-moe, nougat, nystromformer, olmo, olmo2, olmoe, omdet-turbo, oneformer, open-llama, openai-gpt, opt, ovis2, owlv2, owlvit, paligemma, patchtsmixer, patchtst, pegasus, pegasus_x, perceiver, perception_encoder, perception_lm, persimmon, phi, phi3, phi4_multimodal, phimoe, pix2struct, pixtral, plbart, poolformer, pop2piano, prompt_depth_anything, prophetnet, pvt, pvt_v2, qdqbert, qwen2, qwen2_5_omni, qwen2_5_vl, qwen2_5_vl_text, qwen2_audio, qwen2_audio_encoder, qwen2_moe, qwen2_vl, qwen2_vl_text, qwen3, qwen3_moe, rag, realm, recurrent_gemma, reformer, regnet, rembert, resnet, retribert, roberta, roberta-prelayernorm, roc_bert, roformer, rt_detr, rt_detr_resnet, rt_detr_v2, rwkv, sam, sam2, sam2_hiera_det_model, sam2_video, sam2_vision_model, sam_hq, sam_hq_vision_model, sam_vision_model, seamless_m4t, seamless_m4t_v2, seed_oss, segformer, seggpt, sew, sew-d, shieldgemma2, siglip, siglip2, siglip_vision_model, smollm3, smolvlm, smolvlm_vision, speech-encoder-decoder, speech_to_text, speech_to_text_2, speecht5, splinter, squeezebert, stablelm, starcoder2, superglue, superpoint, swiftformer, swin, swin2sr, swinv2, switch_transformers, t5, t5gemma, table-transformer, tapas, textnet, time_series_transformer, timesfm, timesformer, timm_backbone, timm_wrapper, trajectory_transformer, transfo-xl, trocr, tvlt, tvp, udop, umt5, unispeech, unispeech-sat, univnet, upernet, van, video_llava, videomae, vilt, vipllava, vision-encoder-decoder, vision-text-dual-encoder, visual_bert, vit, vit_hybrid, vit_mae, vit_msn, vitdet, vitmatte, vitpose, vitpose_backbone, vits, vivit, vjepa2, voxtral, voxtral_encoder, wav2vec2, wav2vec2-bert, wav2vec2-conformer, wavlm, whisper, xclip, xcodec, xglm, xlm, xlm-prophetnet, xlm-roberta, xlm-roberta-xl, xlnet, xlstm, xmod, yolos, yoso, zamba, zamba2, zoedepth, siglip_visual_tokenizer, aimv2_visual_tokenizer)
 PHASE scoring_rollouts {'adapter_step': 331}
+PHASE optimizer_step_completed {'entropy_nats': 0.5333679221763303, 'parse_rate': 0.9921875, 'zero_variance_fraction': 0.0}
+{'loss': -0.0, 'grad_norm': 0.487217515707016, 'learning_rate': 4.859150070105916e-06, 'num_tokens': 280473769.0, 'completions/mean_length': 136.076171875, 'completions/min_length': 43.0, 'completions/max_length': 646.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.076171875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 646.0, 'rewards/reward/mean': 0.22134509682655334, 'rewards/reward/std': 4.170248985290527, 'reward': 0.22134508192539215, 'reward_std': 3.6744396686553955, 'frac_reward_zero_std': 0.0, 'entropy': 0.5339720036572544, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'epoch': 1.2868217054263567}
+INFO 10-09 13:59:27 [block_pool.py:292] Successfully reset prefix cache
+WARNING 10-09 13:59:27 [tokenizer.py:286] No tokenizer found in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000332, using base model tokenizer instead. (Exception: Unrecognized model in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000332. Should have a `model_type` key in its config.json, or contain one of the following strings in its name: aimv2, aimv2_vision_model, albert, align, altclip, apertus, arcee, aria, aria_text, audio-spectrogram-transformer, autoformer, aya_vision, bamba, bark, bart, beit, bert, bert-generation, big_bird, bigbird_pegasus, biogpt, bit, bitnet, blenderbot, blenderbot-small, blip, blip-2, blip_2_qformer, bloom, bridgetower, bros, camembert, canine, chameleon, chinese_clip, chinese_clip_vision_model, clap, clip, clip_text_model, clip_vision_model, clipseg, clvp, code_llama, codegen, cohere, cohere2, cohere2_vision, colpali, colqwen2, conditional_detr, convbert, convnext, convnextv2, cpmant, csm, ctrl, cvt, d_fine, dab-detr, dac, data2vec-audio, data2vec-text, data2vec-vision, dbrx, deberta, deberta-v2, decision_transformer, deepseek_v2, deepseek_v3, deepseek_vl, deepseek_vl_hybrid, deformable_detr, deit, depth_anything, depth_pro, deta, detr, dia, diffllama, dinat, dinov2, dinov2_with_registers, dinov3_convnext, dinov3_vit, distilbert, doge, donut-swin, dots1, dpr, dpt, efficientformer, efficientloftr, efficientnet, electra, emu3, encodec, encoder-decoder, eomt, ernie, ernie4_5, ernie4_5_moe, ernie_m, esm, evolla, exaone4, falcon, falcon_h1, falcon_mamba, fastspeech2_conformer, fastspeech2_conformer_with_hifigan, flaubert, flava, florence2, fnet, focalnet, fsmt, funnel, fuyu, gemma, gemma2, gemma3, gemma3_text, gemma3n, gemma3n_audio, gemma3n_text, gemma3n_vision, git, glm, glm4, glm4_moe, glm4v, glm4v_moe, glm4v_moe_text, glm4v_text, glpn, got_ocr2, gpt-sw3, gpt2, gpt_bigcode, gpt_neo, gpt_neox, gpt_neox_japanese, gpt_oss, gptj, gptsan-japanese, granite, granite_speech, granitemoe, granitemoehybrid, granitemoeshared, granitevision, graphormer, grounding-dino, groupvit, helium, hgnet_v2, hiera, hubert, hunyuan_v1_dense, hunyuan_v1_moe, ibert, idefics, idefics2, idefics3, idefics3_vision, ijepa, imagegpt, informer, instructblip, instructblipvideo, internvl, internvl_vision, jamba, janus, jetmoe, jukebox, kosmos-2, kosmos-2.5, kyutai_speech_to_text, layoutlm, layoutlmv2, layoutlmv3, led, levit, lfm2, lightglue, lilt, llama, llama4, llama4_text, llava, llava_next, llava_next_video, llava_onevision, longformer, longt5, luke, lxmert, m2m_100, mamba, mamba2, marian, markuplm, mask2former, maskformer, maskformer-swin, mbart, mctct, mega, megatron-bert, metaclip_2, mgp-str, mimi, minimax, mistral, mistral3, mixtral, mlcd, mllama, mm-grounding-dino, mobilebert, mobilenet_v1, mobilenet_v2, mobilevit, mobilevitv2, modernbert, modernbert-decoder, moonshine, moshi, mpnet, mpt, mra, mt5, musicgen, musicgen_melody, mvp, nat, nemotron, nezha, nllb-moe, nougat, nystromformer, olmo, olmo2, olmoe, omdet-turbo, oneformer, open-llama, openai-gpt, opt, ovis2, owlv2, owlvit, paligemma, patchtsmixer, patchtst, pegasus, pegasus_x, perceiver, perception_encoder, perception_lm, persimmon, phi, phi3, phi4_multimodal, phimoe, pix2struct, pixtral, plbart, poolformer, pop2piano, prompt_depth_anything, prophetnet, pvt, pvt_v2, qdqbert, qwen2, qwen2_5_omni, qwen2_5_vl, qwen2_5_vl_text, qwen2_audio, qwen2_audio_encoder, qwen2_moe, qwen2_vl, qwen2_vl_text, qwen3, qwen3_moe, rag, realm, recurrent_gemma, reformer, regnet, rembert, resnet, retribert, roberta, roberta-prelayernorm, roc_bert, roformer, rt_detr, rt_detr_resnet, rt_detr_v2, rwkv, sam, sam2, sam2_hiera_det_model, sam2_video, sam2_vision_model, sam_hq, sam_hq_vision_model, sam_vision_model, seamless_m4t, seamless_m4t_v2, seed_oss, segformer, seggpt, sew, sew-d, shieldgemma2, siglip, siglip2, siglip_vision_model, smollm3, smolvlm, smolvlm_vision, speech-encoder-decoder, speech_to_text, speech_to_text_2, speecht5, splinter, squeezebert, stablelm, starcoder2, superglue, superpoint, swiftformer, swin, swin2sr, swinv2, switch_transformers, t5, t5gemma, table-transformer, tapas, textnet, time_series_transformer, timesfm, timesformer, timm_backbone, timm_wrapper, trajectory_transformer, transfo-xl, trocr, tvlt, tvp, udop, umt5, unispeech, unispeech-sat, univnet, upernet, van, video_llava, videomae, vilt, vipllava, vision-encoder-decoder, vision-text-dual-encoder, visual_bert, vit, vit_hybrid, vit_mae, vit_msn, vitdet, vitmatte, vitpose, vitpose_backbone, vits, vivit, vjepa2, voxtral, voxtral_encoder, wav2vec2, wav2vec2-bert, wav2vec2-conformer, wavlm, whisper, xclip, xcodec, xglm, xlm, xlm-prophetnet, xlm-roberta, xlm-roberta-xl, xlnet, xlstm, xmod, yolos, yoso, zamba, zamba2, zoedepth, siglip_visual_tokenizer, aimv2_visual_tokenizer)
+PHASE scoring_rollouts {'adapter_step': 332}
+PHASE optimizer_step_completed {'entropy_nats': 0.5398191982927094, 'parse_rate': 0.98828125, 'zero_variance_fraction': 0.0}
+{'loss': 0.0, 'grad_norm': 0.4158616065979004, 'learning_rate': 4.858279533144357e-06, 'num_tokens': 281444973.0, 'completions/mean_length': 137.5703125, 'completions/min_length': 45.0, 'completions/max_length': 754.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.5703125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 754.0, 'rewards/reward/mean': -0.5707991719245911, 'rewards/reward/std': 4.862054347991943, 'reward': -0.5707991719245911, 'reward_std': 4.112673759460449, 'frac_reward_zero_std': 0.0, 'entropy': 0.5419605123461224, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'epoch': 1.2906976744186047}
+INFO 10-09 14:06:27 [block_pool.py:292] Successfully reset prefix cache
+WARNING 10-09 14:06:27 [tokenizer.py:286] No tokenizer found in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000333, using base model tokenizer instead. (Exception: Unrecognized model in /home/ubuntu/ib_connect_research/project/tracegrpo/results/phase6/scale_review21_step324_lr5e6/live_adapters/adapter-step-000333. Should have a `model_type` key in its config.json, or contain one of the following strings in its name: aimv2, aimv2_vision_model, albert, align, altclip, apertus, arcee, aria, aria_text, audio-spectrogram-transformer, autoformer, aya_vision, bamba, bark, bart, beit, bert, bert-generation, big_bird, bigbird_pegasus, biogpt, bit, bitnet, blenderbot, blenderbot-small, blip, blip-2, blip_2_qformer, bloom, bridgetower, bros, camembert, canine, chameleon, chinese_clip, chinese_clip_vision_model, clap, clip, clip_text_model, clip_vision_model, clipseg, clvp, code_llama, codegen, cohere, cohere2, cohere2_vision, colpali, colqwen2, conditional_detr, convbert, convnext, convnextv2, cpmant, csm, ctrl, cvt, d_fine, dab-detr, dac, data2vec-audio, data2vec-text, data2vec-vision, dbrx, deberta, deberta-v2, decision_transformer, deepseek_v2, deepseek_v3, deepseek_vl, deepseek_vl_hybrid, deformable_detr, deit, depth_anything, depth_pro, deta, detr, dia, diffllama, dinat, dinov2, dinov2_with_registers, dinov3_convnext, dinov3_vit, distilbert, doge, donut-swin, dots1, dpr, dpt, efficientformer, efficientloftr, efficientnet, electra, emu3, encodec, encoder-decoder, eomt, ernie, ernie4_5, ernie4_5_moe, ernie_m, esm, evolla, exaone4, falcon, falcon_h1, falcon_mamba, fastspeech2_conformer, fastspeech2_conformer_with_hifigan, flaubert, flava, florence2, fnet, focalnet, fsmt, funnel, fuyu, gemma, gemma2, gemma3, gemma3_text, gemma3n, gemma3n_audio, gemma3n_text, gemma3n_vision, git, glm, glm4, glm4_moe, glm4v, glm4v_moe, glm4v_moe_text, glm4v_text, glpn, got_ocr2, gpt-sw3, gpt2, gpt_bigcode, gpt_neo, gpt_neox, gpt_neox_japanese, gpt_oss, gptj, gptsan-japanese, granite, granite_speech, granitemoe, granitemoehybrid, granitemoeshared, granitevision, graphormer, grounding-dino, groupvit, helium, hgnet_v2, hiera, hubert, hunyuan_v1_dense, hunyuan_v1_moe, ibert, idefics, idefics2, idefics3, idefics3_vision, ijepa, imagegpt, informer, instructblip, instructblipvideo, internvl, internvl_vision, jamba, janus, jetmoe, jukebox, kosmos-2, kosmos-2.5, kyutai_speech_to_text, layoutlm, layoutlmv2, layoutlmv3, led, levit, lfm2, lightglue, lilt, llama, llama4, llama4_text, llava, llava_next, llava_next_video, llava_onevision, longformer, longt5, luke, lxmert, m2m_100, mamba, mamba2, marian, markuplm, mask2former, maskformer, maskformer-swin, mbart, mctct, mega, megatron-bert, metaclip_2, mgp-str, mimi, minimax, mistral, mistral3, mixtral, mlcd, mllama, mm-grounding-dino, mobilebert, mobilenet_v1, mobilenet_v2, mobilevit, mobilevitv2, modernbert, modernbert-decoder, moonshine, moshi, mpnet, mpt, mra, mt5, musicgen, musicgen_melody, mvp, nat, nemotron, nezha, nllb-moe, nougat, nystromformer, olmo, olmo2, olmoe, omdet-turbo, oneformer, open-llama, openai-gpt, opt, ovis2, owlv2, owlvit, paligemma, patchtsmixer, patchtst, pegasus, pegasus_x, perceiver, perception_encoder, perception_lm, persimmon, phi, phi3, phi4_multimodal, phimoe, pix2struct, pixtral, plbart, poolformer, pop2piano, prompt_depth_anything, prophetnet, pvt, pvt_v2, qdqbert, qwen2, qwen2_5_omni, qwen2_5_vl, qwen2_5_vl_text, qwen2_audio, qwen2_audio_encoder, qwen2_moe, qwen2_vl, qwen2_vl_text, qwen3, qwen3_moe, rag, realm, recurrent_gemma, reformer, regnet, rembert, resnet, retribert, roberta, roberta-prelayernorm, roc_bert, roformer, rt_detr, rt_detr_resnet, rt_detr_v2, rwkv, sam, sam2, sam2_hiera_det_model, sam2_video, sam2_vision_model, sam_hq, sam_hq_vision_model, sam_vision_model, seamless_m4t, seamless_m4t_v2, seed_oss, segformer, seggpt, sew, sew-d, shieldgemma2, siglip, siglip2, siglip_vision_model, smollm3, smolvlm, smolvlm_vision, speech-encoder-decoder, speech_to_text, speech_to_text_2, speecht5, splinter, squeezebert, stablelm, starcoder2, superglue, superpoint, swiftformer, swin, swin2sr, swinv2, switch_transformers, t5, t5gemma, table-transformer, tapas, textnet, time_series_transformer, timesfm, timesformer, timm_backbone, timm_wrapper, trajectory_transformer, transfo-xl, trocr, tvlt, tvp, udop, umt5, unispeech, unispeech-sat, univnet, upernet, van, video_llava, videomae, vilt, vipllava, vision-encoder-decoder, vision-text-dual-encoder, visual_bert, vit, vit_hybrid, vit_mae, vit_msn, vitdet, vitmatte, vitpose, vitpose_backbone, vits, vivit, vjepa2, voxtral, voxtral_encoder, wav2vec2, wav2vec2-bert, wav2vec2-conformer, wavlm, whisper, xclip, xcodec, xglm, xlm, xlm-prophetnet, xlm-roberta, xlm-roberta-xl, xlnet, xlstm, xmod, yolos, yoso, zamba, zamba2, zoedepth, siglip_visual_tokenizer, aimv2_visual_tokenizer)
+PHASE scoring_rollouts {'adapter_step': 333}
+PHASE optimizer_step_completed {'entropy_nats': 0.5097311908977622, 'parse_rate': 0.98828125, 'zero_variance_fraction': 0.0}
+{'loss': -0.0, 'grad_norm': 0.4894426167011261, 'learning_rate': 4.857406392712601e-06, 'num_tokens': 282422131.0, 'completions/mean_length': 142.55859375, 'completions/min_length': 39.0, 'completions/max_length': 1073.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.55859375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 1073.0, 'rewards/reward/mean': -0.2507460117340088, 'rewards/reward/std': 4.825312614440918, 'reward': -0.2507460117340088, 'reward_std': 4.070449352264404, 'frac_reward_zero_std': 0.0, 'entropy': 0.5206792474418762, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'epoch': 1.2945736434108528}

results/phase6/scale_review21_step324_lr5e6/step_metrics.jsonl

@@ -5,3 +5,6 @@
 {"optimizer_step": 329, "adapter_step": 328, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.990234375, "verifier_failures": 6, "step_failure_rate": 0.01171875, "label_accuracy": 0.5078125, "reward_component_means": {"R1": 0.02734375, "R2": 0.2653845822485795, "R3": 0.022459959959959957, "R4": 0.44140625, "R5": -0.057031250000000006}, "completion_tokens": {"mean": 134.693359375, "p50": 115.5, "p95": 249.0, "max": 1247}, "failure_causes": {"parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 2, column 80": 1, "parser_failure:contract_v2: ValueError: invalid v2 trace step": 1, "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 1, "parser_failure:contract_v2: ParseError: mismatched tag: line 8, column 31": 1, "ValueError:judge_invalid_score:<analysis>\n<explanation>The code": 1, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 2, column 117": 1}, "generated_at": "2026-10-09T13:28:21.901387+00:00", "aggregate_parse_rate": 0.9866309684236771, "verifier_window_failure_rate": 0.01064453125, "verifier_window_failures": 109, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 168449, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.01064453125, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.860883329777125e-06, "entropy_nats": 0.5372266856046501, "entropy_completion_tokens": 68963, "completed_at": "2026-10-09T13:32:56.445171+00:00"}
 {"optimizer_step": 330, "adapter_step": 329, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.9921875, "verifier_failures": 4, "step_failure_rate": 0.0078125, "label_accuracy": 0.505859375, "reward_component_means": {"R1": 0.01953125, "R2": 0.28547945717370193, "R3": 0.022238658777120314, "R4": 0.439453125, "R5": -0.04453125}, "completion_tokens": {"mean": 141.78125, "p50": 125.5, "p95": 286.34999999999997, "max": 1397}, "failure_causes": {"parser_failure:contract_v2: ValueError: invalid v2 trace step": 1, "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 1, "parser_failure:contract_v2: ParseError: undefined entity: line 6, column 87": 1, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 7, column 114": 1}, "generated_at": "2026-10-09T13:35:45.412410+00:00", "aggregate_parse_rate": 0.9866478062984949, "verifier_window_failure_rate": 0.01064453125, "verifier_window_failures": 109, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 168961, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.01064453125, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.860018002636229e-06, "entropy_nats": 0.5038575398743297, "entropy_completion_tokens": 72592, "completed_at": "2026-10-09T13:39:43.761916+00:00"}
 {"optimizer_step": 331, "adapter_step": 330, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.982421875, "verifier_failures": 9, "step_failure_rate": 0.017578125, "label_accuracy": 0.51171875, "reward_component_means": {"R1": 0.041015625, "R2": 0.26127895963701636, "R3": 0.036118659420289856, "R4": 0.451171875, "R5": -0.0748046875}, "completion_tokens": {"mean": 141.419921875, "p50": 129.5, "p95": 296.45, "max": 595}, "failure_causes": {"parser_failure:contract_v2: ValueError: v2 requires exactly one <analysis> block without surrounding text": 1, "parser_failure:contract_v2: ParseError: mismatched tag: line 2, column 234": 1, "parser_failure:contract_v2: ParseError: mismatched tag: line 10, column 2": 1, "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 1, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 5, column 40": 1, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 5, column 46": 1, "parser_failure:contract_v2: ParseError: mismatched tag: line 9, column 2": 1, "parser_failure:contract_v2: ValueError: invalid v2 trace step": 2}, "generated_at": "2026-10-09T13:44:35.383082+00:00", "aggregate_parse_rate": 0.9866350392097857, "verifier_window_failure_rate": 0.011328125, "verifier_window_failures": 116, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 169473, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.011328125, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.859150070105916e-06, "entropy_nats": 0.5321434676768195, "entropy_completion_tokens": 72407, "completed_at": "2026-10-09T13:50:11.607610+00:00"}
+{"optimizer_step": 332, "adapter_step": 331, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.9921875, "verifier_failures": 4, "step_failure_rate": 0.0078125, "label_accuracy": 0.533203125, "reward_component_means": {"R1": 0.07421875, "R2": 0.24101022533402044, "R3": 0.029607451763140388, "R4": 0.48046875, "R5": -0.042578125}, "completion_tokens": {"mean": 136.076171875, "p50": 127.0, "p95": 259.9, "max": 646}, "failure_causes": {"parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 2, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 11, column 43": 1, "parser_failure:contract_v2: ParseError: mismatched tag: line 2, column 34": 1}, "generated_at": "2026-10-09T13:52:45.304479+00:00", "aggregate_parse_rate": 0.9866517633908874, "verifier_window_failure_rate": 0.0111328125, "verifier_window_failures": 114, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 169985, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.0111328125, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.858279533144357e-06, "entropy_nats": 0.5333679221763303, "entropy_completion_tokens": 69671, "completed_at": "2026-10-09T13:56:32.691748+00:00"}
+{"optimizer_step": 333, "adapter_step": 332, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.98828125, "verifier_failures": 6, "step_failure_rate": 0.01171875, "label_accuracy": 0.45703125, "reward_component_means": {"R1": -0.07421875, "R2": 0.21056271525827627, "R3": 0.020094086021505374, "R4": 0.423828125, "R5": -0.062890625}, "completion_tokens": {"mean": 137.5703125, "p50": 127.0, "p95": 275.9, "max": 754}, "failure_causes": {"parser_failure:contract_v2: ParseError: mismatched tag: line 16, column 2": 1, "parser_failure:contract_v2: ValueError: 'CWE-overflow' does not match '^(CWE-[0-9]+|NONE)$'": 1, "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 2, "parser_failure:contract_v2: ParseError: mismatched tag: line 20, column 2": 1, "parser_failure:contract_v2: ValueError: 'CWE-59)' does not match '^(CWE-[0-9]+|NONE)$'": 1}, "generated_at": "2026-10-09T14:00:45.381089+00:00", "aggregate_parse_rate": 0.9866566567153674, "verifier_window_failure_rate": 0.0115234375, "verifier_window_failures": 118, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 170497, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.0115234375, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.857406392712601e-06, "entropy_nats": 0.5398191982927094, "entropy_completion_tokens": 70436, "completed_at": "2026-10-09T14:04:44.345485+00:00"}
+{"optimizer_step": 334, "adapter_step": 333, "n_rollouts": 512, "groups": 64, "zero_groups": 0, "zero_variance_fraction": 0.0, "parse_rate": 0.98828125, "verifier_failures": 6, "step_failure_rate": 0.01171875, "label_accuracy": 0.486328125, "reward_component_means": {"R1": -0.015625, "R2": 0.2463194344420451, "R3": 0.024276377217553685, "R4": 0.47265625, "R5": -0.058007812500000006}, "completion_tokens": {"mean": 142.55859375, "p50": 125.0, "p95": 288.34999999999997, "max": 1073}, "failure_causes": {"parser_failure:contract_v2: ValueError: 'CWE-125,CWE-119' does not match '^(CWE-[0-9]+|NONE)$'": 1, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 2, column 59": 1, "parser_failure:contract_v2: ValueError: '' should be non-empty": 2, "parser_failure:contract_v2: ParseError: not well-formed (invalid token): line 5, column 114": 1, "parser_failure:contract_v2: ValueError: duplicate or overlapping statement ranges": 1}, "generated_at": "2026-10-09T14:07:48.246021+00:00", "aggregate_parse_rate": 0.9866615207386745, "verifier_window_failure_rate": 0.01142578125, "verifier_window_failures": 117, "verifier_window_rollouts": 10240, "canary_in_failure_window": false, "canaries_in_failure_window": 0, "all_generated_rollouts_including_canary": 171009, "dgrpo6b_failure_decision": {"source": "D-GRPO-6b (2026-10-07 delegated prospective ruling)", "rolling_failure_rate": 0.01142578125, "rolling_window_steps": 20, "nonparser_rolling_rate": 9.765625e-05, "completed_block_wilson_lowers": [0.018117845054430304, 0.016752423051967957, 0.013668988573899227, 0.014120917591715385, 0.013307864394505934, 0.011239402594480075, 0.009631719614628635, 0.008920953010197434, 0.010970713591566107, 0.012047080241744917, 0.009631719614628635, 0.009631719614628635, 0.008743666775120913, 0.007244158561402301, 0.009542737334915601, 0.009098405671651343], "warning": false, "stop": false, "stop_reason": null}, "learning_rate_change_step": 325, "base_learning_rate": 5e-06, "learning_rate_source": "D-GRPO-15 / Review #21", "effective_learning_rate": 4.856530649774567e-06, "entropy_nats": 0.5097311908977622, "entropy_completion_tokens": 72990, "completed_at": "2026-10-09T14:11:51.503238+00:00"}

scripts/stop_scale_at_step400.py

@@ -33,7 +33,7 @@ def main():
             from src.rl.full_state_checkpoint import read_extra_state
             manifest=json.loads((CHECKPOINT/'full_state_manifest.json').read_text())
             if manifest['optimizer_step']!=400:
-                time.sleep(.05);continue
+                time.sleep(.5);continue
             for name,expected in manifest['sha256'].items():
                 if sha(CHECKPOINT/name)!=expected:raise RuntimeError('checkpoint hash mismatch: '+name)
             extra=read_extra_state(CHECKPOINT)
@@ -70,7 +70,7 @@ def main():
             LOG.write_text(json.dumps(dict(status='run_stopped_before_400',
                 observed=status,checked_at=utc()),indent=2)+'\n')
             return
-        time.sleep(.05)
+        time.sleep(.5)
 
 if __name__=='__main__':
     try:main()