| { |
| "study": "Path A: curriculum head-to-head on a text policy (Qwen3-1.7B-4bit, MLX-LoRA, M5)", |
| "protocol": "same model, same oracle (plumb_forge score_predictions), same held-out seed-777 benchmark (n=1000) \u2014 only the training curriculum varies", |
| "date": "2026-08-22", |
| "hub": { |
| "dataset": "https://huggingface.co/datasets/caiotheodoro/plumb", |
| "handseeded": "https://huggingface.co/caiotheodoro/plumb-handseeded", |
| "ornith": "https://huggingface.co/caiotheodoro/plumb-ornith", |
| "blended": "https://huggingface.co/caiotheodoro/plumb-blended", |
| "collection": "https://huggingface.co/collections/caiotheodoro/plumb-6a89f2ea18c97d11b0238529" |
| }, |
| "cis_95_bootstrap_10k_seed11": { |
| "handseeded": { |
| "sw-recall": [ |
| 0.29, |
| 0.347 |
| ], |
| "precision": [ |
| 0.279, |
| 0.337 |
| ], |
| "exact": [ |
| 0.154, |
| 0.203 |
| ] |
| }, |
| "ornith": { |
| "sw-recall": [ |
| 0.214, |
| 0.268 |
| ], |
| "precision": [ |
| 0.098, |
| 0.124 |
| ], |
| "exact": [ |
| 0.066, |
| 0.103 |
| ] |
| }, |
| "blended": { |
| "sw-recall": [ |
| 0.306, |
| 0.363 |
| ], |
| "precision": [ |
| 0.342, |
| 0.406 |
| ], |
| "exact": [ |
| 0.201, |
| 0.254 |
| ] |
| }, |
| "pow-ornith": { |
| "sw-recall": [ |
| 0.139, |
| 0.186 |
| ], |
| "precision": [ |
| 0.102, |
| 0.135 |
| ], |
| "exact": [ |
| 0.052, |
| 0.085 |
| ] |
| }, |
| "pow-handseeded": { |
| "sw-recall": [ |
| 0.154, |
| 0.203 |
| ], |
| "precision": [ |
| 0.152, |
| 0.201 |
| ], |
| "exact": [ |
| 0.071, |
| 0.108 |
| ] |
| }, |
| "pow-blended": { |
| "sw-recall": [ |
| 0.238, |
| 0.295 |
| ], |
| "precision": [ |
| 0.204, |
| 0.255 |
| ], |
| "exact": [ |
| 0.135, |
| 0.184 |
| ] |
| } |
| }, |
| "rows_full_size": [ |
| { |
| "run": "base", |
| "n": 1000, |
| "train_tasks": 0, |
| "parse_rate": 0.127, |
| "severity_weighted_recall": 0.02, |
| "high_recall": 0.034, |
| "precision": 0.118, |
| "corrected_figure_exact_match": 0.0 |
| }, |
| { |
| "run": "handseeded", |
| "n": 1000, |
| "train_tasks": 223, |
| "parse_rate": 1.0, |
| "severity_weighted_recall": 0.318, |
| "high_recall": 0.336, |
| "precision": 0.308, |
| "corrected_figure_exact_match": 0.178 |
| }, |
| { |
| "run": "ornith", |
| "n": 1000, |
| "train_tasks": 58, |
| "parse_rate": 0.997, |
| "severity_weighted_recall": 0.241, |
| "high_recall": 0.287, |
| "precision": 0.111, |
| "corrected_figure_exact_match": 0.084 |
| }, |
| { |
| "run": "blended", |
| "n": 1000, |
| "train_tasks": 281, |
| "parse_rate": 1.0, |
| "severity_weighted_recall": 0.334, |
| "high_recall": 0.365, |
| "precision": 0.374, |
| "corrected_figure_exact_match": 0.228 |
| } |
| ], |
| "rows_equal_n_58_random": [ |
| { |
| "run": "equal-handseeded", |
| "n": 1000, |
| "train_tasks": 49, |
| "parse_rate": 0.998, |
| "severity_weighted_recall": 0.106, |
| "high_recall": 0.036, |
| "precision": 0.126, |
| "corrected_figure_exact_match": 0.023 |
| }, |
| { |
| "run": "equal-ornith", |
| "n": 1000, |
| "train_tasks": 49, |
| "parse_rate": 0.999, |
| "severity_weighted_recall": 0.124, |
| "high_recall": 0.083, |
| "precision": 0.111, |
| "corrected_figure_exact_match": 0.038 |
| }, |
| { |
| "run": "equal-blended", |
| "n": 1000, |
| "train_tasks": 49, |
| "parse_rate": 0.997, |
| "severity_weighted_recall": 0.062, |
| "high_recall": 0.013, |
| "precision": 0.1, |
| "corrected_figure_exact_match": 0.02 |
| } |
| ], |
| "rows_equal_n_58_stratified": [ |
| { |
| "run": "strat-handseeded", |
| "n": 1000, |
| "train_tasks": 49, |
| "parse_rate": 0.999, |
| "severity_weighted_recall": 0.053, |
| "high_recall": 0.081, |
| "precision": 0.132, |
| "corrected_figure_exact_match": 0.011 |
| }, |
| { |
| "run": "strat-blended", |
| "n": 1000, |
| "train_tasks": 49, |
| "parse_rate": 0.968, |
| "severity_weighted_recall": 0.169, |
| "high_recall": 0.094, |
| "precision": 0.142, |
| "corrected_figure_exact_match": 0.095 |
| } |
| ], |
| "rows_exploratory_pow184": [ |
| { |
| "run": "pow-ornith", |
| "n": 1000, |
| "train_tasks": 184, |
| "n_train_split": 156, |
| "parse_rate": 1.0, |
| "severity_weighted_recall": 0.163, |
| "high_recall": 0.164, |
| "precision": 0.118, |
| "corrected_figure_exact_match": 0.068, |
| "val_loss": 0.4697 |
| }, |
| { |
| "run": "pow-handseeded", |
| "n": 1000, |
| "train_tasks": 184, |
| "n_train_split": 156, |
| "parse_rate": 0.997, |
| "severity_weighted_recall": 0.178, |
| "high_recall": 0.206, |
| "precision": 0.176, |
| "corrected_figure_exact_match": 0.089, |
| "val_loss": 0.4918 |
| }, |
| { |
| "run": "pow-blended", |
| "n": 1000, |
| "train_tasks": 184, |
| "n_train_split": 156, |
| "parse_rate": 0.999, |
| "severity_weighted_recall": 0.266, |
| "high_recall": 0.365, |
| "precision": 0.229, |
| "corrected_figure_exact_match": 0.159, |
| "val_loss": 0.4803 |
| } |
| ], |
| "do_not_publish_as_method": true, |
| "exploratory_caveats": [ |
| "pow-* and the reward-only / instruction-anchored grows fit DeploymentProfile on forge/data/benchmark.jsonl (seed-777) and told the proposer to match the held-out eval \u2014 eval-set leakage. Not a methods result.", |
| "Acceptance was r>0, so multiplicative M never filtered. count_match compared P_bench(count) to uniform, so the most common deployment class (single-injection) scored lowest.", |
| "D was estimated against the linear placeholder (frontier at difficulty ~0.8), not the 1.7B policy.", |
| "N had no floor: late naive-grow accepts had novelty 0.01-0.07. pow-ornith@184 (0.163 sw-recall) is WORSE than ornith@58 (0.241) \u2014 growing a self-proposed buffer without a novelty floor added near-duplicates, not information.", |
| "pow-blended / anchored2 / mockanchored were still in flight at harvest of pow-ornith and pow-handseeded; fold in when eval artifacts land, still as exploratory.", |
| "The clean protocol (train prior, mix constraint, novelty floor, 1.7B/error-profile D) is the methods experiment." |
| ], |
| "train_val_loss": { |
| "ornith": 0.5948, |
| "handseeded": 0.4861, |
| "blended": 0.4726, |
| "equal-handseeded": 0.6289, |
| "equal-ornith": 0.5919, |
| "equal-blended": 0.6303, |
| "strat-handseeded": 0.6065, |
| "strat-blended": 0.6058, |
| "pow-ornith": 0.4697, |
| "pow-handseeded": 0.4918 |
| }, |
| "interpretation": [ |
| "Full size (powered): Ornith-only underperforms hand-seeded with non-overlapping CIs (sw-recall 0.241 [0.214,0.268] vs 0.318 [0.290,0.347]). Blend beats both on point estimates; vs hand-seeded the recall CIs overlap and precision is the surviving interval (0.374 [0.342,0.406] vs 0.308 [0.279,0.337]).", |
| "Equal-N at 58 tasks (underpowered): all curricula collapse to 0.05-0.17 sw-recall \u2014 58 tasks cannot train this task well, and the ordering is noisy/unstable (random vs stratified flips hand-seeded). No conclusion about 'blended-wins-is-just-more-data' can be drawn from an underpowered small-N run.", |
| "The one stable small-N signal: the stratified blend@58 (0.169 sw-recall, 0.095 exact) is the best small-N curriculum \u2014 a coverage-preserving mix of matched + Ornith tasks beats either alone, reinforcing the full-size blend conclusion.", |
| "Stratifying hand-seeded@58 to force 20 PASS + all 9 types diluted per-type counts and hurt recall (0.053) \u2014 small-N composition is dominated by allocation tradeoffs, i.e., noise.", |
| "EXPLORATORY, do not publish as method: naive grow 58\u2192184 hurt (0.241 \u2192 0.163). Novelty too weak. Reward-only M did not steer under accept-any-positive. Clean equal-N with train-prior mix constraint is the follow-up." |
| ], |
| "artifacts": { |
| "per_run": "curriculum/data/eval/eval-{run}.json", |
| "head2head": "curriculum/data/eval/head2head.json", |
| "equalN": "curriculum/data/eval/equalN-head2head.json", |
| "adapters": "curriculum/adapters/text-{handseeded,ornith,blended,equal-*,strat-*,pow-*}" |
| }, |
| "rows_equal_n_clean": [ |
| { |
| "run": "eq-handseeded", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.88, |
| "severity_weighted_recall": 0.06, |
| "high_recall": 0.094, |
| "precision": 0.11, |
| "corrected_figure_exact_match": 0.0, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.05993606819392648, |
| "ci95": [ |
| 0.04429028815368197, |
| 0.07626893436088227 |
| ] |
| }, |
| "precision": { |
| "point": 0.10955710955710955, |
| "ci95": [ |
| 0.08168316831683169, |
| 0.1388235294117647 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.0, |
| "ci95": [ |
| 0.0, |
| 0.0 |
| ] |
| } |
| } |
| }, |
| { |
| "run": "eq-naive", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.98, |
| "severity_weighted_recall": 0.135, |
| "high_recall": 0.101, |
| "precision": 0.079, |
| "corrected_figure_exact_match": 0.025, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.13465636654235483, |
| "ci95": [ |
| 0.1136426401546534, |
| 0.15606196389706822 |
| ] |
| }, |
| "precision": { |
| "point": 0.07853107344632769, |
| "ci95": [ |
| 0.06659142212189616, |
| 0.09080590238365494 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.02450719232818327, |
| "ci95": [ |
| 0.014787430683918669, |
| 0.03514293207448203 |
| ] |
| } |
| } |
| }, |
| { |
| "run": "eq-mix", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.843, |
| "severity_weighted_recall": 0.075, |
| "high_recall": 0.081, |
| "precision": 0.082, |
| "corrected_figure_exact_match": 0.011, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.07511987213638785, |
| "ci95": [ |
| 0.05898989898989899, |
| 0.09253139458030403 |
| ] |
| }, |
| "precision": { |
| "point": 0.08167330677290836, |
| "ci95": [ |
| 0.06517311608961303, |
| 0.09959758551307847 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.010522109749600427, |
| "ci95": [ |
| 0.005422153369481023, |
| 0.016440217391304347 |
| ] |
| } |
| } |
| }, |
| { |
| "run": "eq-error", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.99, |
| "severity_weighted_recall": 0.031, |
| "high_recall": 0.004, |
| "precision": 0.116, |
| "corrected_figure_exact_match": 0.002, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.03143313798614811, |
| "ci95": [ |
| 0.021973735535040954, |
| 0.04191197084384637 |
| ] |
| }, |
| "precision": { |
| "point": 0.11572700296735905, |
| "ci95": [ |
| 0.08282208588957055, |
| 0.15202702702702703 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.0019978689397975495, |
| "ci95": [ |
| 0.0, |
| 0.0047694753577106515 |
| ] |
| } |
| } |
| }, |
| { |
| "run": "eq-mockanchored", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.987, |
| "severity_weighted_recall": 0.125, |
| "high_recall": 0.166, |
| "precision": 0.108, |
| "corrected_figure_exact_match": 0.033, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.12493340436867342, |
| "ci95": [ |
| 0.10391016100662968, |
| 0.14705111402359108 |
| ] |
| }, |
| "precision": { |
| "point": 0.10797174571140263, |
| "ci95": [ |
| 0.08906882591093117, |
| 0.12763819095477386 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.03329781566329249, |
| "ci95": [ |
| 0.021077591885127124, |
| 0.0467350382967675 |
| ] |
| } |
| } |
| }, |
| { |
| "run": "eq-blend", |
| "n": 1000, |
| "n_train_split": 15, |
| "parse_rate": 0.98, |
| "severity_weighted_recall": 0.055, |
| "high_recall": 0.076, |
| "precision": 0.101, |
| "corrected_figure_exact_match": 0.001, |
| "ci95": { |
| "severity_weighted_recall": { |
| "point": 0.05487480021310603, |
| "ci95": [ |
| 0.039577121171049064, |
| 0.07064079296384197 |
| ] |
| }, |
| "precision": { |
| "point": 0.1010989010989011, |
| "ci95": [ |
| 0.07407407407407407, |
| 0.12873563218390804 |
| ] |
| }, |
| "corrected_figure_exact_match": { |
| "point": 0.0013319126265316996, |
| "ci95": [ |
| 0.0, |
| 0.00411635565312843 |
| ] |
| } |
| } |
| } |
| ], |
| "equal_n_clean": 18, |
| "equal_n_clean_sizes": { |
| "handseeded": 223, |
| "naive": 24, |
| "mix": 22, |
| "error": 18, |
| "mockanchored": 22 |
| }, |
| "clean_protocol_interpretation": [ |
| "equal N=18 (15 train / 3 valid); novelty floor 0.3 capped grow at 18-24.", |
| "no weak win: mix 0.075 [0.059, 0.093] below naive 0.135 [0.114, 0.156], no overlap.", |
| "mock-anchored 0.125 [0.104, 0.147] overlaps naive.", |
| "no strong win: error 0.031 [0.022, 0.042] below hand-seeded 0.060 [0.044, 0.076].", |
| "no super win: error below blend 0.055 [0.040, 0.071].", |
| "mix buffer: 16/22 accepts were controller deficit-fill; 9B proposer mode-collapsed to singles.", |
| "N=18 does not touch full-size bars (handseeded 0.318, blend 0.334)." |
| ] |
| } |
|
|
