Эксперимент: Dual Information-Demand Curiosity (#946)

{ "worked": true, "confidence": 9, "verdict": "Built a convex information-coordinate toy implementing the dual update, bounded attainable information, inactive/binding/saturated regimes, and a fixed-coefficient baseline. The mechanism manifested exactly: inactive demand produced lambda=0 and I=I0; five feasible targets tracked with absolute error below 1.3e-15; unattainable demand reached Imax with lambda=8; and fixed-information dual drift matched eta*(H-I) to floating-point precision. This validates the KKT/dual mechanism, but not its usefulness in a neural world model or RL task.", "metrics": { "baseline": "Fixed lambda=1 gives I=0.505008 nats regardless of target.", "idea": "I0=0.005008, Imax=0.661668; inactive I=0.005008/lambda=0; binding target H=0.234839 gives I=0.234839/lambda=0.459662; saturated H=0.761668 gives I=0.661668/lambda=8; feasible-sweep max |I-H|=1.22e-15; fixed-I drift errors are numerical zero." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 dual_information_toy.py", "files": [ "dual_information_toy.py", "results.json" ], "limitations": "This is an analytically controlled scalar-information toy, not a learned latent world model. It uses exact mutual information and an exact inner primal solve rather than sampled neural decoders, and does not test reconstruction quality, RL return, wall-clock speed, stochastic estimator variance, or Bethe-factor-graph training." }