Эксперимент: Coxeter Folding Reversible Recurrence (#1407)

{ "worked": true, "confidence": 7, "verdict": "Built a complex-polygon Coxeter-style fold recurrence with fixed alternating schedule, exact reverse scheduling, denominator diagnostics, and a GRU memory comparison. Single-fold involution and one-cycle reverse reconstruction were at floating-point roundoff in float64; repeated float32 schedules accumulated error, and the singularity sweep showed the expected blow-up as the denominator approached zero. The reversible effect is real, but constant-memory training was not implemented because the MVP uses ordinary autograd; the memory result is therefore a native-unroll comparison, not proof of constant-memory backpropagation.", "metrics": { "baseline": "CPU GRU, batch=32, sequence length=32: 692224 native-autograd saved bytes; 291 saved tensors.", "idea": "CPU fold chain, 32 folds: 86016 native-autograd saved bytes; 418 saved tensors; reverse reconstruction relative error 2.248847158625722e-05. Float64 reverse errors were 4.50e-16 at 8 steps, 7.01e-15 at 40, and 2.27e-11 at 160; float32 errors were 3.32e-07, 1.28e-05, and 3.11e-02 respectively. Near-singular folded magnitude grew from 2.03 at denominator 0.209 to 2.16e8 at denominator 2.09e-9." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py && /home/maxwelhelp/main/bin/python3 benchmark.py", "files": [ "experiment.py", "benchmark.py", "results.json", "benchmark_results.json" ], "limitations": "No learned schedule, input injection, affine gauge normalization, sequential-MNIST/task loss, custom reversible backward, or true constant-memory training was tested. CUDA encountered CUDNN_STATUS_INTERNAL_ERROR_HOST_ALLOCATION_FAILED, so the benchmark used its CPU fallback." }