Stage 1: Qualification (DeBERTa)
Waiting...
Why separate qualification?
DeBERTa reads language patterns and outputs confidence scores (0.0–1.0). No judgment yet — just signal extraction. This is auditable and reproducible.
Stage 2: Rules (Deterministic)
Waiting...
Why deterministic rules?
Code handles judgment. Thresholds are explicit: >0.8 = SOLID, 0.5–0.8 = EXAMINE, <0.5 = ATTENTION. Same input always produces same output. No AI hallucination of criteria.
Stage 3: Language (Haiku)
Waiting...
Why separate language generation?
Haiku narrates decisions already made. It cannot change the judgment — only explain it. The explanation is constrained by the rules, not invented by AI.