From Clicks to Constructs: Event-Stream Process Mining, TreeSHAP Attribution, and Rasch Partial Credit Calibration of Behavioural Trace Indicators for Cross-Scale Science Reasoning

Coordinating molecular, cellular, and macroscopic explanations is a core requirement of secondary biology and chemistry, yet classroom assessments still infer a student's developmental level from written answers collected after an inquiry task ends. Digital process telemetry records the timestamped clicks, slider sweeps, pauses, and revisions a student produces inside a simulation, and it captures how learners move across scales in real time. Raw interaction counts, however, lack a shared unit of measurement with curriculum learning progressions. This study addresses that measurement gap through a two-part Monte Carlo validation. First, I specified a fixed-parameter benchmark of 86 cross-scale inquiry study clusters (312 effect rows, 48,920 learners) whose target values were anchored to the verified cross-scale inquiry literature (pooled Hedges' g = 0.642, adaptive next-step ZPD+1 prompts g = 0.784, unguarded language model access g = 0.412, Zero-Paste guardrails beta = 0.189), and I confirmed that a three-level correlated and hierarchical effects model recovers them. Second, using that benchmark alongside a simulated bilingual cohort of 1,240 secondary learners and 1,000 Monte Carlo replications, I calibrated six non-intrusive behavioural trace indicators against a 12-item ordered multiple-choice progression bank on a unified Rasch partial credit logit scale. Common-scale linking error was 0.118 logits (10.9% of the critical Level 2–3 transition interval). Trace-based classifications reproduced item-based progression levels with 81.45% accuracy and Gwet's first-order agreement coefficient of .768, exhibited strict measurement invariance across language tracks, gender, and prior achievement, and confirmed that genuine self-regulatory behaviours rather than interface guardrails drive diagnostic classification. Within the limits of a fixed-parameter simulation, behavioural traces can thus provide calibrated, equitable coordinates for real-time formative support in secondary science classrooms. Because every input is generated, all estimates reported here are computational recovery and calibration results rather than field-trial effect estimates.