Эксперимент: Critical-depth sparse attention (#700)

{ "worked": false, "confidence": 8, "verdict": "Built a finite-grid dyadic strong-halo implementation with horizontal/vertical ancestor depths, complementary critical height H_s, exponential packing metrics, antichain routing, and deterministic parameter sweeps. Two algebraic predictions were confirmed: H_s is exactly symmetric under s↔1−s in this symmetric setup, and mean H scales exactly linearly with Lambda; however, the directional/critical-exponent sweep was completely flat because the full active mask makes the halo trivial, so it does not validate the claimed nontrivial concentration-control mechanism. The router demo reduced p99/max overlap, but only by pruning 128 candidates to 70, making it an unequal-compute comparison.", "metrics": { "baseline": "128 unconstrained candidates; mean H=0.3922, p99=1.2, max=1.6, exp_mean=1.1056", "idea": "70 greedy-antichain candidates; mean H=0.1734, p99=0.6, max=0.6, exp_mean=1.0450; unequal candidate count prevents attributing this solely to critical-depth weighting" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 critical_depth_experiment.py", "files": [ "critical_depth_experiment.py", "results.json", "run_output.txt" ], "limitations": "No neural attention training or validation accuracy was tested; no equal-selected-token/FLOP comparison was performed; the active mask was fully populated, making the dyadic halo trivial and causing the directional s sweep to be uninformative; no GPU benchmark or memory measurement was run; the exponential theorem was not tested across a meaningful family of sparse masks." }