| { |
| "utc": "2026-09-30T01:03:13.145569+00:00", |
| "selected": "unchanged base recipe: lr1e-4 wd.1 dropout.1 width128 phi6 budget128000", |
| "reason": "Two-layer development achieves sustained >90% held-out composition and 97.08% fixed endpoint; no retuning needed. Four-layer reference also succeeds. All development endpoints retained; no confirmation worlds read.", |
| "development": { |
| "dev-d2-original": { |
| "parameters": 416512, |
| "training_seconds": 125.8127673609415, |
| "evaluation_seconds": 1.1260491695720702, |
| "endpoint": { |
| "step": 128000, |
| "utc": "2026-09-30T01:00:16.690367+00:00", |
| "parameters": 416512, |
| "training_seconds": 125.8127673609415, |
| "capture_seconds": 0.6041618440067396, |
| "examples": 32768000, |
| "counts": { |
| "atomic": 4889904, |
| "composite": 27878096 |
| }, |
| "effective_input_tokens": 126182096, |
| "supervised_tokens": 65536000, |
| "estimated_training_flops": 318196678656000, |
| "last_batch_loss": 0.004912014119327068, |
| "atomic": { |
| "n": 1024, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 6.468004357884638e-06 |
| }, |
| "train_composite": { |
| "n": 5838, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 5.646218596666586e-06 |
| }, |
| "test_composite": { |
| "n": 754, |
| "answer_accuracy": 0.9708222811671088, |
| "accuracy": 0.9708222811671088, |
| "nll": 0.04894252493977547 |
| }, |
| "ood_composite": { |
| "n": 14, |
| "answer_accuracy": 0.0, |
| "accuracy": 0.0, |
| "nll": 6.5166850090026855 |
| }, |
| "two_calls": { |
| "n": 754, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 5.410871835920261e-06 |
| }, |
| "evaluation_seconds": 1.1260491695720702 |
| } |
| }, |
| "dev-d4-reference": { |
| "parameters": 813056, |
| "training_seconds": 182.0216330656549, |
| "evaluation_seconds": 1.648130600224249, |
| "endpoint": { |
| "step": 128000, |
| "utc": "2026-09-30T01:01:28.449357+00:00", |
| "parameters": 813056, |
| "training_seconds": 182.0216330656549, |
| "capture_seconds": 0.5995612218976021, |
| "examples": 32768000, |
| "counts": { |
| "atomic": 4889904, |
| "composite": 27878096 |
| }, |
| "effective_input_tokens": 126182096, |
| "supervised_tokens": 65536000, |
| "estimated_training_flops": 629044936704000, |
| "last_batch_loss": 0.0007603238918818533, |
| "atomic": { |
| "n": 1024, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 3.510803480821778e-06 |
| }, |
| "train_composite": { |
| "n": 5838, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 5.936707566434052e-06 |
| }, |
| "test_composite": { |
| "n": 754, |
| "answer_accuracy": 0.9893899204244032, |
| "accuracy": 0.9893899204244032, |
| "nll": 0.014771726913750172 |
| }, |
| "ood_composite": { |
| "n": 14, |
| "answer_accuracy": 0.07142857142857142, |
| "accuracy": 0.07142857142857142, |
| "nll": 4.617985248565674 |
| }, |
| "two_calls": { |
| "n": 754, |
| "answer_accuracy": 1.0, |
| "accuracy": 1.0, |
| "nll": 2.9723194074904313e-06 |
| }, |
| "evaluation_seconds": 1.648130600224249 |
| } |
| } |
| }, |
| "confirmation_worlds": [ |
| 143011, |
| 143012, |
| 143013 |
| ], |
| "initializations": [ |
| 14301, |
| 14302 |
| ], |
| "primary_runs": [ |
| "conf-w143011-s14301-d1", |
| "conf-w143011-s14301-d2", |
| "conf-w143011-s14302-d1", |
| "conf-w143011-s14302-d2", |
| "conf-w143012-s14301-d1", |
| "conf-w143012-s14301-d2", |
| "conf-w143012-s14302-d1", |
| "conf-w143012-s14302-d2", |
| "conf-w143013-s14301-d1", |
| "conf-w143013-s14301-d2", |
| "conf-w143013-s14302-d1", |
| "conf-w143013-s14302-d2" |
| ], |
| "conditional_wide_runs": [ |
| "conf-w143011-s14301-d1wide", |
| "conf-w143011-s14302-d1wide", |
| "conf-w143012-s14301-d1wide", |
| "conf-w143012-s14302-d1wide", |
| "conf-w143013-s14301-d1wide", |
| "conf-w143013-s14302-d1wide" |
| ], |
| "parameter_control": { |
| "width": 180, |
| "parameters": 419220, |
| "two_layer_parameters": 416512, |
| "relative_difference": 0.0065016133988937685, |
| "selection": "nearest parameter count over widths multiple of 4; no accuracy used" |
| }, |
| "boundary": "Shared training recipe, not per-depth optimal hyperparameters; no linear scaling law from two depths." |
| } |
|
|
