Эксперимент: Schur-Riesz Greedy Adapter Expansion (#860)

{ "worked": true, "confidence": 9, "verdict": "Built a NumPy MVP of Schur-Riesz greedy adapter expansion with Y-weighted projection, residual candidate responses, gain ranking, singular-value filtering, and fixed-order comparison. The math checks manifested quantitatively: projection orthogonality was 2.85e-14, idempotence error was 1.53e-15, residual gain was approximately zero at zero novel amplitude and followed exact quadratic scaling (ratios 0.0625, 0.25, 1, 4 for amplitudes 0.25, 0.5, 1, 2), while predicted and observed least-squares loss drops matched to 1.83e-16 relative error. In the controlled toy regression, greedy selected all three useful blocks and reached loss 9.51 versus 483.09 for fixed-order expansion, a 98.0% lower final loss at the same three added scalar blocks; this is a clear mechanism signal, not evidence of production-scale neural-adapter gains.", "metrics": { "baseline": "Fixed-order expansion selected blocks [0,1,2], all labeled redundant; loss 501.51 -> 483.09 after 3 additions.", "idea": "Projected-gain greedy selected blocks [7,5,11], all labeled useful; loss 501.51 -> 9.51 after 3 additions; final loss ratio idea/baseline = 0.01968. Math: orthogonality 2.85e-14, idempotence 1.53e-15, gain amplitude ratios exactly [0.0625,0.25,1,4], loss-drop relative prediction error 1.83e-16." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 schur_riesz_mvp.py > results.json", "files": [ "schur_riesz_mvp.py", "results.json" ], "limitations": "This is a deterministic synthetic weighted least-squares experiment, not a Transformer, LoRA, MoE, or held-out neural fine-tuning benchmark. It does not measure GPU runtime, FLOPs, parameter-memory savings, train-after-selection acceptance thresholds, minibatch Jacobian operators, or robustness across datasets and random seeds." }