Boundaries
8 Settings / 8
Dashboard

Settings

Configuration and calibration

Allocation Parameters
Parameter Value
beta_u 1.000
beta_n 1.000
beta_w 1.505
beta_pi 1.000
beta_e 0.800
gamma_un 0.200
gamma_uw 0.300
alpha_ai_precision 2.000
alpha_hybrid_uncertainty 0.800
alpha_human_conflict 1.510
Safety Thresholds
Threshold Value
u_critical 0.90
n_critical 0.90
w_critical 0.85
e_critical 0.90
w_prohibit_ai 0.70
n_prohibit_ai 0.80
u_prohibit_ai 0.80
governance_tier_high 0.70
governance_tier_low 0.30
min_confidence 0.30
surprise_escalation_threshold 0.70
surprise_tier_boost 1.00
surprise_threshold_adaptive 1.00
surprise_threshold_percentile 0.90
surprise_threshold_min 0.50
surprise_threshold_max 0.95
surprise_history_size 100.00
Parameter Calibration

Calibrate allocation parameters based on recorded outcomes. This uses precision-weighted gradient descent to optimize parameters.

Higher = faster learning, may overshoot
Learning Statistics

Loading...

Surprise-Triggered Auditing

When prediction error exceeds the threshold (default: 0.7), the system automatically:

  • Escalates governance tier for more thorough documentation
  • Adds SURPRISE_ESCALATION flag to the decision
  • Increases scrutiny for unexpected situations

Theoretical basis: In Active Inference, high "surprise" indicates model inadequacy, warranting increased human oversight.

Learning from Overrides (κ Multiplier)

When humans override AI recommendations, the system learns with boosted precision:

  • Override precision: ωoverride = κ · ωbase, where κ = 1.5 (default)
  • Escalations (AI→Human): Higher weight on safety corrections
  • Delegations (Human→AI): System learns efficiency patterns

Effect: Unified score weights (wi) are adjusted via gradient descent: Δwi = η · κ · ωbase · εr · ∂r/∂wi

How System Learning Works
What the Metrics Mean
Total Updates Number of batch gradient descent updates performed. Each update adjusts parameters based on ~50-100 recorded outcomes.
Pending Outcomes Outcomes waiting to be processed. A batch update triggers when this reaches the batch size threshold.
Mean Recent Error Average prediction error over recent decisions. Lower is better. Target: < 0.15
Std Recent Error Variability in prediction errors. Lower values indicate more consistent performance.
Cumulative Error Total accumulated error across all decisions. Used internally for gradient computation.
Override Learnings Count of parameter updates from human overrides. When humans correct AI recommendations, the system learns from these corrections.
Escalation Corrections Overrides where AI was corrected to Human mode. These use κ-boosted precision (κ=1.5) for safety learning.
Surprise Escalations Times when high prediction error automatically escalated governance tier for more thorough documentation.
How to Observe Improvement
  1. Run simulations with "Use Real API" enabled on the Simulation page to generate decisions
  2. Record outcomes - The system learns when actual outcomes are recorded via the API
  3. Watch Mean Error decrease - After multiple batch updates, the mean error should trend downward
  4. Click "Run Calibration" to trigger immediate parameter adjustment
Improvement Indicators
Optimized Mean Error < 0.15 — System is well-calibrated
Learning Mean Error 0.15 - 0.30 — System is adapting
Needs Work Mean Error > 0.30 — Run calibration
Unified Score Function & Learning
Unified score function (sigmoid of weighted features):
r(x;θ) = σ(Σi wi·φi(xi) + Σi<j wij·φij(xi,xj) + bias)
Mode thresholding:
T(r) = AI if r < b₁, Hybrid if b₁ ≤ r < b₂, Human if r ≥ b₂
Gradient learning rule:
Δwi = η · ωobs · εr · σ'(z) · φi(xi)
Where: x = (U, N, W, π̃, ε̃) ∈ [0,1]⁵ dimensional vector, η = learning rate (default 0.01),
ωobs = observation precision, εr = y - r̂ (score error), σ' = sigmoid derivative,
(b₁, b₂) = (0.3, 0.7) default thresholds
Error Trend Over Time

No error history available yet. Run simulations with outcomes to see the trend.

Current Status: -
Trend: -
General Settings
Debug Mode OFF
Log Level info
Demo UI ENABLED
Storage Backend memory
Allocation Temperature 1.0
Auto-load Plugins ON
System Self-Test

Run the automated test suite to verify system functionality. This executes all pytest tests and displays the results.

Configuration via Environment

MCAT can be configured via environment variables:

# General
MCAT_DEBUG=true
MCAT_LOG_LEVEL=DEBUG
MCAT_ENABLE_DEMO_UI=true

# Unified Score weights (sign-constrained)
MCAT_UNIFIED_W_U=1.0      # Uncertainty (≥0)
MCAT_UNIFIED_W_N=1.0      # Novelty (≥0)
MCAT_UNIFIED_W_W=1.5      # Value conflict (≥0)
MCAT_UNIFIED_W_PI=-1.0    # Precision (≤0)
MCAT_UNIFIED_W_E=0.8      # Prediction error (≥0)
MCAT_UNIFIED_B1=0.3       # AI/Hybrid threshold
MCAT_UNIFIED_B2=0.7       # Hybrid/Human threshold

# Learning parameters
MCAT_LEARNING_RATE=0.01   # η - gradient learning rate
MCAT_LEARNING_KAPPA=1.5   # κ - override precision multiplier

# Thresholds
MCAT_THRESHOLD_U_CRITICAL=0.9
MCAT_THRESHOLD_W_CRITICAL=0.85

# Storage
MCAT_STORAGE_BACKEND=memory
MCAT_DATABASE_URL=sqlite+aiosqlite:///mcat.db