Same governing rule as the 2D reconciliation-workflow simulator, rebuilt in real 3D space instead of a flat card layout: a physician score p and an AI score a are drawn independently for every case — p uniformly from [0.28, 0.88], and a = clamp(p + noise, 0.02, 0.98) with noise drawn from [-n, n]. The gap is Δ = |p − a|. If Δ exceeds the reconciliation threshold T the case is flagged as a discrepancy and routed through the reconciliation pad; otherwise it counts as agreement.
- Case complexity — sets the noise half-width n = lerp(0.03, 0.55, (v−1)/9), exactly as in the 2D model. Higher complexity means the AI opinion can drift further from the physician's.
- Reconciliation threshold — sets T = lerp(0.05, 0.42, (v−1)/9). Strict (low) settings flag more cases as discrepancies; lenient (high) settings flag fewer.
- Case cycle speed — a 3D-only control: how fast the arena advances through physician → AI → compare → classify → outcome for each case.
In the arena, the blue column's height is the physician score and the purple column's height is the AI score, both rising to true scale (0–1 mapped to 0–2.4 world units). The bridge beam only appears once both scores have risen, turns green and solid for agreement or amber and segmented for a discrepancy, and a pink ring in front of the columns lights up while a flagged case is being reconciled. A row of tally blocks behind the arena keeps the last 24 verdicts visible at a glance.