HITL Escalation Report

HITL Escalation Report — Human-in-the-Loop Verification Protocol

The HITL Escalation Report documents every instance where the Digital Refinery's AI classification system requires human verification. This occurs when the primary Gemma 4 E2B model produces a confidence score below the 85% threshold — the AMBER band where automated classification alone is insufficient for regulatory defensibility.

When triggered, the escalation protocol engages the secondary E4B Visual Fallacy guard model (4.5B parameters) for dual-model consensus. If both models disagree or neither achieves sufficient confidence, the load is flagged for human operator review with full visual evidence and preliminary AI analysis.

The report captures: primary model (E2B) classification and confidence score, secondary model (E4B) classification and confidence score, consensus status (agree/disagree/abstain), visual evidence with AI-highlighted regions of concern, ambient audio analysis summary, operator decision and reasoning (when human review is completed), and total escalation-to-resolution time.

This escalation mechanism is critical for addressing the Visual Fallacy failure mode — where weathered CCA-treated wood appears visually identical to naturally aged clean wood. The RGB blind spot in standard camera-based classification makes this one of the highest-risk scenarios in Ontario construction waste processing, where misclassification can result in contaminated material entering the clean pyrogenic biocarbon feedstock stream.

Each HITL Escalation Report becomes part of the SR&ED (Scientific Research and Experimental Development) evidence chain, documenting the technical uncertainty inherent in edge AI classification and the systematic approach to resolution. This supports 905WOOD.COM SALES' annual SR&ED claim under the 23% combined federal-provincial tax credit program.