E-LAB-05 · EntropyLab · April 2026

ENTRO-EVO

Adaptive Entropy Weighting for Self-Calibrating Intelligence Systems.
Learning optimal control weights online via gradient descent — no manual calibration.

Ψ_norm · Normalized Entropy State · Real-time Learning

0.07

Ψ_norm (target = 0.1)

CONVERGING · Learning weights

GitHub Repository DOI: 10.5281/zenodo.19464489

01 · Core Innovation

Adaptive Entropy Weighting

Instead of manual calibration, ENTRO-EVO learns optimal control weights online through gradient descent.

AEW Update Rule

w(t+1) = clip(w(t) - η·∇L, w_min, w_max)

η(t) = η₀/(1+κ·t) · L(t) = (Ψ_norm - Ψ*)²

Dynamic Thresholding

θ(t+1) = θ_base + (1 - H(t))·Δθ_max

H(t) = stability history index · θ_base = 1.2 · Δθ_max = 0.6

Cross-Domain Transfer

w_init = λ·w_reg(e*) + (1-λ)·w_default

λ = similarity(F(e), F(e*)) · 71.5% faster re-adaptation

02 · Validation Results

Three-Regime Performance

Regime Steady-State Error Budget Violations Final Weights [w₁,w₂,w₃]

Scraper (High Volatility) 0.365 0 [0.20, 0.10, 0.70]

LLM (Slow Saturation) 0.062 0 [0.80, 0.10, 0.10]

Phase Shift (Abrupt Transition) 0.367 0 [0.17, 0.10, 0.73]

Best PSI · v11.0 (Thermodynamic Controller)

Ψ = 0.075

FINAL WEIGHTS

[0.02, 0.02, 0.97]

TEMPERATURE

0.72

PHASE

stabilize

ENTROPY

0.17

Best Balance · v7.0 (Softmax Normalization)

Ψ = 0.697

FINAL WEIGHTS

[0.33, 0.33, 0.34]

PERFECT BALANCE

No collapse

ENTROPY

1.10

Best Compromise · v5.1 (Dual-Objective)

Ψ = 0.246

FINAL WEIGHTS

[0.38, 0.22, 0.40]

K (GAIN)

0.73

BUDGET VIOLATIONS

# pip install entro-evo
from entro_evo import AdaptiveEntropyWeighting

aew = AdaptiveEntropyWeighting(eta_0=0.01, target_psi=0.339)
weights = aew.step(psi_norm=0.5, d_psi=0.1, d2_psi=0.05, u_t=0.3)

# → Output
weights = [0.48, 0.31, 0.21] · Converging to target

"The marks of a mature control system are not its performance under known conditions —
it is its grace under unknown ones. ENTRO-EVO gives entropy-regulated intelligence
the grace to learn the physics of its own environment."

— Samir Baladi · ENTRO-EVO · April 2026