SAGE: A Guardrailed LLM Tutoring System with Scaffolded Self-Regulated Learning for Adult Education

Abstract

We present SAGE (Scaffolded, Audited, Guided Education), a deployable guardrailed LLM tutoring system for adult learners. Left unrestricted, chat assistants often hand over the answer, which can improve in-the-moment performance but may erode the self-regulation that supports durable learning. SAGE wraps a commercial LLM in interface-level guardrails that require learners to engage before the model answers: a lightweight intent classifier routes each turn to one of three self-regulated-learning scaffolds (write-your-approach-first, paraphrase-before-proceeding, attempt-before-reveal), so a request to ''just give me the answer'' is held behind a guided step rather than answered outright. As a secondary safeguard, an online counterfactual fairness audit checks whether the guardrailed tutor responds differently across age band, language dominance, prior education, and digital fluency, while an agentic rectifier regenerates any flagged turn; a pseudonymous interaction log supports per-cohort analytics. The demonstration runs the same business-education prompt through an unrestricted LLM and through SAGE side by side, allowing an attendee to watch the scaffold gate hold the answer, then exercise the counterfactual swap, inspect the per-attribute audit panel, and trigger the rectifier. The system is open-source and available at https://github.com/kwanhui/public-llmLearnFair-CIKM2026 to allow other groups to deploy and audit their own scaffolds.