Asha's Escalation-Driver Confound Check — Chapter 10 Worked Example ========================================================================= The Monday-morning Key Influencers verification trace behind Asha’s 90-minute escalation review in Chapter 10 of The Defensible Decision. Three drivers on the AI output; three verification moves; three different verdicts. Wholly synthetic. Any resemblance to a specific real Key Influencers output is coincidental; the confound-check pattern is drawn from typical practitioner-observed behaviour on support-ticket data. Files ----- - asha-escalation-drivers.csv — flat rows: one per driver with raw multiplier, verification move, verdict, notes, and action - asha-escalation-drivers.json — canonical scenario contract with source hash What it represents ------------------ Power BI Key Influencers ranks three drivers of escalation on Asha’s support-tickets model: ticket age over 7 days (4.1x), Enterprise tier (2.3x), Identity product line (1.8x). All three look defensible on the raw AI output. Asha runs the confound-check verification move per driver: driver 1 (age) survives clean; driver 2 (Enterprise tier) collapses to a modest ~1.2x effect once customer-size is added as a co-Explain-by field (Enterprise-tier was serving as a proxy for size); driver 3 (Identity) survives when scoped correctly to the SSO/federation feature area but not to the product line as a whole. The Smart Narrative gets two edits; the Anomaly Detection flag gets suppressed as a known business-calendar event; the Decomposition Tree walk carries a caveat because its path includes the Enterprise-tier confound. Invariants (locked) ------------------- The scenario carries executable invariants that Chapter 10’s release contract enforces: - Exactly 3 drivers with ids 1..3 and no gaps. - Every driver’s raw effect multiplier > 1.0 (Key Influencers surfacing threshold). - Each driver names a verification move and a verdict from {survives, survives_when_scoped, confound}. - Exactly one driver survives clean, one survives scoped, one collapses as confound. - Ticket-age driver has the highest raw multiplier and survives. - Enterprise-tier driver is the confound; carries a specific action ("do not surface"). - Smart Narrative edits: exactly 2 (remove Enterprise, rescope Identity). - Anomaly Detection flag matches a business-calendar event (monthly integration rollout). - Decomposition Tree AI-selected path includes the Enterprise-tier confound (so the tree walk carries the same caveat as the Smart Narrative). - Review summary counts reconcile with per-driver verdicts. Why the invariants matter ------------------------- Chapter 10’s pedagogical point is per-visual triage — the four AI visuals require different verification moves because they use different algorithms. Key Influencers needs the confound check; Anomaly Detection needs the business-calendar check; Decomposition Tree needs the filter-context check; Smart Narrative needs the BRD-alignment check. Asha’s trace is the concrete case that shows all four moves on one dataset. If the driver verdicts collapsed (all survive, or all fail), the trace would stop teaching per-visual triage as a discipline. What it does NOT represent -------------------------- - Actual support-ticket data with real customer identities. - The full Explain-by field set (only the two fields that produce the confound — tier and customer size — are modelled explicitly). - Statistical significance thresholds (Key Influencers surfaces effects it considers meaningful; the discipline is verification, not p-value gate-keeping). Reproducibility --------------- The dataset ships from the book source at AlexBooks/applied-ai/ai-directed-viz/data/asha-escalation-drivers.json and is re-emitted here by AlexBooks/applied-ai/ai-directed-viz/scripts/generate-asha-drivers.mjs --site-root . License ------- Released for teaching use alongside The Defensible Decision. Use freely.