Distinguish useful assumptions from harmful bias
Does anchoring a moving network prove that stronger priors make reasoning better? No. An inductive bias restricts what a learner can infer; harmful bias concerns unfair outcomes in context. This page trains no AI, uses no participant data and checks no prediction against a ground truth. Its physics metaphor can illustrate a chosen dependency, but cannot validate the blank-slate claim.
Worked formula and the special baseline
For weight 0.40, prior strength 0.65 and depth 4: divergence = 82×0.40 + 18×(1−0.65) + 1.5×4 − 10 = 35.1, displayed 35.1%. Amplification = 1 + 3.8×0.40×(1.1−0.8×0.65) = 1.8816, displayed 1.9x. Percent and multiplier labels have no empirical denominator.
The default (0.85, 0.20, 4) bypasses those equations and forces 78.4%, 4.2x, High and Premise Collapse. The general equations would give 80.1% and 4.0362x. This discontinuity is a fixture, not an experimental observation. Divergence is clipped to 4.2–99.4; categories use thresholds 35, 65 and 75.
Random motion and replay limits
Math.random varies initial node positions, radii and anomaly forces without a saved seed. Replay can change motion while formula metrics remain fixed. The chart combines 0.8×score, summed node speed and horizontal premise displacement, then clips to 0–100; these mixed units form an authored motion proxy. The displayed 120-step counter is capped while animation continues. JSON exports parameters and formula results, not the random state or full trajectory, so it cannot reproduce a run.
Primary research and better questions
NIST distinguishes computational/statistical, human and systemic sources of harmful AI bias. Feinman and Lake study how neural networks acquire a shape preference using controlled synthetic training data. Neither source validates this page’s coefficients or a universal claim about human development. For a real test, define a task, compare assumptions on held-out data, report uncertainty and assess which people experience errors.