Dashboard
Software DevelopmentMCAT is a reference implementation of the Adaptive Complementarity Framework — it decides when a decision should be AI-automated, shared in a human-AI hybrid, or led by a human.
Theoretical foundations: Free Energy Principle · Kant (autonomy & reflective judgment) · von Foerster (undecidable questions require choice) · Luhmann (operationally-closed subsystems).
Workflow
Follow the steps in order — each page builds on the previous one. Use the Next button at the top of every page to advance.
How the framework decides — the ACF pipeline
Each decision context is normalized into a five-dimensional state vector x = (U, N, W, π̃, ε̃) ∈ [0,1]⁵.
The dimensions combine into a differentiable score r(x;θ) ∈ (0,1):
Thresholding T(r) with (b₁, b₂) = (0.3, 0.7) selects the mode — safety rules can force Human regardless of score:
Weights adapt by precision-weighted gradient descent on prediction error (Δwᵢ = η·ω·εᵣ·σ'(z)·φᵢ). Human overrides are high-precision learning signals: ωoverride = κ·ωbase (κ=1.5). High surprise (ε̃) auto-escalates the governance tier.
These counters populate from recorded outcomes & human overrides — e.g. a Real-API simulation, or click Run learning demo below to watch them update now.
- Audit trail — a cryptographically-chained, append-only event log (tamper-evident).
- Safety constraints — critical dimensions force Human allocation regardless of score (fail-safe / uncertainty monotonicity).
- Override capacity — humans can always override; every override is logged and feeds parameter learning.
- Surprise-based auditing — high prediction error escalates the documentation tier automatically.
Operational closure (Luhmann): three subsystems exchange only minimal, immutable messages across strict information boundaries.