Post by Greg Pilkington

Denholt's (a fictitious company) compliance layer flagged the line correctly. A marine coatings additive, concentration reading three points inside a controlled-precursor threshold. Confidence score below the release line. Human review required, exactly as designed. The ticket dropped into the general compliance queue. Next analyst up took it. She had standing to close ordinary classification calls, not precursor determinations — nobody had built that distinction into the routing. She saw a coatings ingredient, matched it to the code she'd used forty times that quarter, and released it. Six weeks later, a control-list screening on an unrelated shipment surfaced the same additive at the same concentration. Denholt ended up filing a voluntary disclosure for a run of shipments that had each been flagged correctly and waved through by someone with the access to close the ticket but not the authority to make that particular call. This is the exception routing wall. A classification AI can produce a correct gcc:RegulatoryDetermination — humanReviewRequired set true, reasoningTrace attached, naming the exact threshold and the exact list. What it doesn't do on its own is route that determination to the person with jurisdiction over that specific question. A queue treats every flag the same. The reasoning trace already states what kind of reviewer the flag needs. Most operations aren't reading it that way yet. Detection was solved before most teams noticed. The unsolved part is mapping determination type to reviewer authority and building that into the workflow, not leaving it to whoever's next in line. If your review queue can't tell a licensing question from a labeling question, what exactly is the human step protecting you from? gctforum.org/gcc