A small feed-forward network makes approve/deny decisions (blue dots flowing left to right, turning green or red on the output layer). The dark box in the middle hides the hidden layers — how much of it you can see depends on the transparency level. Four panel controls mirror real AI-transparency levers: how much of the pipeline is disclosed, whether individual decisions are explained, and whether every decision is logged for audit.
Trust score = 0.40 × Transparency + 0.35 × Explainability + 0.25 × Auditability
Transparency: black box = 0, limited = 0.35, selective = 0.70, full = 1.00
Explainability: on = 1.00, off = 0.15
Auditability: audit trail on = 1.00, off = 0.00
Documentation completeness = (revealed components) / 4 × 100%
- Black box — only inputs and the final decision are visible; the hidden layers stay sealed inside the dark shroud, matching a proprietary model with no disclosure.
- Limited — hidden nodes become faintly visible and a Model Card (purpose, architecture, limitations) is published, but connection weights stay hidden — typical of most production APIs.
- Selective — connection weights become visible and a Datasheet (training-data provenance, known bias) is added, as if disclosed to a regulator or auditor rather than the public.
- Full — every weight is visible and a public Transparency Report is published, matching open-source or fully-disclosed research models.
- Explainability — colors each connection by its learned weight (blue = positive, amber = negative) and pulses them, approximating a feature-importance / SHAP-style explanation of each decision.
- Audit trail — logs every decision with a sequential ID and outcome; without it, decisions vanish with no record, exactly the compliance gap the EU AI Act and similar rules target for high-risk systems.