Эксперимент: Probe-Then-Partitioned Multi-Task Trunk (#497)

{ "worked": true, "confidence": 8, "verdict": "Built a NumPy mechanism verification and a PyTorch probe-then-partitioned multi-task trunk MVP. The core predictions manifested: cosine-distance identities were exact, the small-noise embedding law E[d]≈7σ² matched within 2.3% for σ≤0.06, and the analytic partition gain matched the angle sweep to numerical precision. In the learned toy regression, probe partitioning reduced mean MSE from 0.340 to 0.032 and worst-task MSE from 0.992 to 0.078, but the discovered partition used singleton trunks and was not parameter/FLOP matched.", "metrics": { "baseline": "Shared trunk: mean MSE 0.3396, worst-task MSE 0.9924.", "idea": "Probe partition: mean MSE 0.0320, worst-task MSE 0.0779; discovered groups [0,0,1,2,0,0]. Oracle partition: mean MSE 0.0353, worst-task MSE 0.0774." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py && /home/maxwelhelp/main/bin/python3 torch_mini.py", "files": [ "experiment.py", "torch_mini.py", "results.json", "torch_results.json" ], "limitations": "This is a tiny synthetic regression experiment, not NYUv2 or Multi-MNIST. HDBSCAN itself was not used; a lightweight MST-cut approximation was implemented. The learned comparison did not equalize parameter count, inference FLOPs, or probe-plus-training cost, and gradient-conflict rate, validation generalization, label-permutation robustness, and half-budget probing were not tested." }