Эксперимент: Dual-Rail Ratio Arithmetic Layer (#1200)

{ "worked": false, "confidence": 9, "verdict": "Built the dual-rail tensor, direct-sum, signed multiplication, gauge normalization, exact-algebra checks, and matched deep-product benchmark. The algebraic identities were verified to roughly 1e-16 relative error, and normalized rails stayed finite at depths where scalar float32 products overflowed; however, the ratio denominator underflowed, producing infinite decoded values, and normalization did not reliably improve relative error. Thus the bounded-rail signal exists, but the promised usable numerical stability was not demonstrated.", "metrics": { "baseline": "Float32 scalar product overflowed by depth 950 for factor 1.1; at depth 500 relative error was 1.12e-6 and at depth 1000 output was non-finite.", "idea": "Normalized pair rails remained finite through depth 1000 for factor 1.1, but decoding became inf because the denominator underflowed; relative error was only inconsistently better, such as 1.44e-7 versus 3.71e-7 at depth 100, but 6.49e-7 versus 3.62e-7 at depth 200. Exact unsigned and signed algebra maximum relative errors were 3.7e-16." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 dual_rail_experiment.py && /home/maxwelhelp/main/bin/python3 mini_benchmark.py", "files": [ "dual_rail_experiment.py", "mini_benchmark.py" ], "limitations": "No trained multiplicative MLP or optimization experiment was run, and no FLOP- or parameter-matched neural comparison was performed. The test used sequential positive products only; signed cancellation, matrix operations, activation nonlinearities, and gradient stability remain untested. The implementation renormalizes after multiplication, so intermediate products can still overflow in general, and ratio decoding needs a log-domain or guarded representation to avoid denominator underflow." }