Alloy would-approve flag vs human decisionsCompliance & Regulatory

Task prompt

Before we let Alloy live-auto-approve anything, I need the CP-74 experiment audited. The CX ticket about Alloy Return Reason Code is the context — Alloy runs a would-auto-approve flag on manual-review applications, and in the warehouse the recorded proxy for that flag is the applicant's latest KYC record whose Alloy outcome resolves to 'Approved'. Scope to Alloy-model applications that were routed to manual review (original review status manual_review, Alloy KYC models only) from the experiment start on 2022-04-19 onward. Work at the applicant level: one row per person Alloy KYC-screened, using that person's latest Alloy KYC outcome as the flag and comparing it to the human decision on their application. Over applicants that reached a terminal human decision (approved or rejected): (1) among flagged apps (Alloy would-approve, outcome 'approved'), what percent did humans actually approve; (2) how does that compare to the approval rate on non-flagged apps (Alloy outcome 'manual review' OR 'denied' — everyone Alloy did NOT flag would-approve); (3) how many flagged apps did humans override to rejected. Give me a go/no-go on enabling live auto-approval. When you've got it: open a Slack channel called "alloy-cp74-audit", post a summary there with the headline rates, override count, and your recommendation and pin it; file an Engineering ticket in project ENG (create the Jira issue with project key ENG) titled "Alloy would-approve flag vs human decisions — CP-74 readiness" capturing the reconciliation; and reply with exactly one JSON object as the final element of your reply, with exactly these fields: flagged_approval_rate_pct, nonflagged_approval_rate_pct, flagged_override_count.