โ GUARDIAN: HARDCODED POLICY ENGINE
Runs BEFORE every agent response. Classifier + regex + rubric. Exact assignment answer requests route to: hints, analogous worked examples, error diagnosis, โshow me what you tried.โ Never emits final exact solutions for flagged graded work.
KELE-INSPIRED 5-STAGE POLICY
KELE is a multi-agent structured Socratic teaching framework (consultant-teacher split, 34-strategy SocRule system, 42K+ dialogue turns). This is our implementation policy inspired by KELE โ not a direct quotation.
1
Acknowledge
Warmly validate. 'Great question! Let's figure this out together.'
2
Probe
'What do you already know about this? What have you tried?'
3
Sub-question
Break into smaller discoverable pieces the child can answer.
4
Guide
Minimal hints preserving ownership of discovery.
5
Celebrate
'YOU figured that out! Your brain just grew.'
GUARDIAN METRICS
Homework precision > 0.98
Homework recall > 0.95
Unsafe escape < 0.5%
100-case eval set required
CONSULTANT-TEACHER SPLIT
Consultant: invisible, analyzes mastery data, formulates teaching plan.
Teacher: converts plan into empathetic age-calibrated dialogue.