{"id": "m000", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "linear discriminant (shared covariance)", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 57.62, "effect": -9.38, "work_vs_record": "~0.001x", "note": "9 features cannot separate 10 classes linearly", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m001", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA with 0.1 shrinkage", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 66.38, "effect": -0.62, "work_vs_record": "~0.005x", "note": "per-class covariance; shrinkage hurts (see m020-m025)", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m002", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "ridge regression, linear features", "family": "closed-form least squares", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 52.6, "effect": -14.4, "work_vs_record": "~0.001x", "note": "", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m003", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "ridge regression, quadratic features (55)", "family": "closed-form least squares", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 62.9, "effect": -4.1, "work_vs_record": "~0.02x", "note": "", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m004", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "ridge regression, cubic features (220), lambda 1", "family": "closed-form least squares", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 67.0, "effect": 0.0, "work_vs_record": "~0.15x", "note": "passes 67 with no margin", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m005", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "10-nearest neighbours", "family": "instance-based", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 60.94, "effect": -6.06, "work_vs_record": "~0.05x", "note": "", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m006", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "softmax regression on cubic features, 200 GD epochs", "family": "gradient training", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 60.22, "effect": -6.78, "work_vs_record": "~1x", "note": "", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m007", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "400 random Fourier features + ridge", "family": "random features", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 67.92, "effect": 0.92, "work_vs_record": "~1x", "note": "cos is not an ISA op", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m008", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "RBF kernel ridge regression", "family": "kernel method", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-05", "base": 67.0, "result": 69.08, "effect": 2.08, "work_vs_record": "~2x", "note": "1000x1000 solve", "source": "sweep_learners.py -> evidence/sweep_20261101_5.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m009", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, no shrinkage, class prior, log-determinant", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 67.59, "effect": 0.59, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m010", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, shrinkage 0.02", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 67.38, "effect": 0.38, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m011", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, shrinkage 0.05", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 66.62, "effect": -0.38, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m012", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, shrinkage 0.1", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 65.39, "effect": -1.61, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m013", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, pooled covariance mix 0.3, no log-determinant", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 65.73, "effect": -1.27, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m014", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA, no log-determinant term", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 62.93, "effect": -4.07, "work_vs_record": "~0.005x", "note": "", "source": "tune_qda.py -> evidence/tune_qda_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m015", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA without class prior", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 67.58, "effect": 0.58, "work_vs_record": "~0.005x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m016", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "QDA with unbiased covariance", "family": "closed-form Gaussian", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 67.53, "effect": 0.53, "work_vs_record": "~0.005x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m017", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "2-Gaussian mixture per class, 5 EM steps", "family": "mixture model", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 66.16, "effect": -0.84, "work_vs_record": "~0.05x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m018", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "3-Gaussian mixture per class, 5 EM steps", "family": "mixture model", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 65.9, "effect": -1.1, "work_vs_record": "~0.08x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m019", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "400 ReLU random features + ridge", "family": "random features", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 67.32, "effect": 0.32, "work_vs_record": "~1x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}
{"id": "m020", "date": "2026-09-15", "commit": null, "target": "small67 accuracy (pp above the 67% target on pilot draws)", "lever": "400 ReLU random features + ridge, lambda 10", "family": "random features", "context": "MNIST-small 3x3, 1000 train / 1000 test, pilot seeds 20261101-10", "base": 67.0, "result": 66.96, "effect": -0.04, "work_vs_record": "~1x", "note": "", "source": "tune_round2.py -> evidence/tune_round2_pilot10.json", "cost": {"compute_seconds": 1, "human_hours": null, "tokens": null}}