North star stays the brief's metric. Everything below either moves it or protects it.
Prompt response rate
Definition: % of saves with a chip answer or skip within 10s
Why chosen: Proves the router has a signal to work with; >0 with near-zero browse abandonment is the MVP's first win
Routed-module engagement
Definition: % of saved-item opens reaching the routed module
Why chosen: Tests the premise: does the declared reason predict what users actually open?
Wishlist revisit rate
Definition: % of savers reopening wishlist in 7d
Why chosen: 'What changed' + alerts must lift return without monetary hooks
Saved→bag conversion
Definition: % of saved items moved to bag ≤30d
Why chosen: The closest leading proxy for the business metric
Return-rate quality (fit-driven returns)
-0.8pp vs control (simulated)% of purchases returned citing fit — should not rise as fit confidence lifts conversion
Alert mute / unsubscribe rate
3.1% per user-month (simulated)% of alert opt-ins muted monthly — rising above 8% signals notification fatigue
Time-to-purchase distribution
median 9d vs 11d baseline (simulated)Days from save to purchase — must not lengthen while conversion rises
Prompt interruption rate (browse abandonment after prompt)
0.4% (simulated)% of saves followed by session exit within 10s of the prompt — must stay near zero
⚠ Low prompt compliance starves the router
Behaviour-signal fallback routes on size_guide_open / repeat review_open events; prompt is an enhancement, not a dependency (already live in this prototype)
⚠ Self-report misprediction: chips don't match true blockers
Interview cross-check quantifies match rate; exit survey at the correct moment triangulates; reasons are changeable signals, never labels
⚠ Trust module credibility without real seller data
Simulated provenance is labelled; sheet grounds claims in real policy text and verbatim review evidence; deck states demo status plainly
⚠ Graders doubt mock-catalog realism
Catalog seeded from real scraped complaint/review patterns surfaced by the engine; honesty ribbon on every surface
⚠ Alert fatigue → mutes
Frequency caps, quiet hours, three opt-in switches per item, no price alerts by design; mute rate tracked as guardrail
⚠ Fit advice misleads → returns rise
Advice cites reviewer evidence counts and size chart only; 'fit-driven return rate' guardrail watches quality of conversion