How Flagging Works

Partner Health & Churn Early Warning System · Sensing & Reasoning Layers

1. What the system watches

Every partner is scored monthly using four signal categories:

Signal	What it means	Why it matters
Slots listed	Bookable slots the partner has made available this cycle	Often the first sign of disengagement, before payout even moves
Slots booked	Listed slots actually booked by users	Utilization rate = booked ÷ listed. Low utilization can be a partner or a demand problem
Attendance confirmed	Bookings that converted to a real check-in	Separates "booked but ghosted" from real usage
Payout per cycle	Amount paid to the partner, and its trend over 3 cycles	The commercial signal, ties directly to why a partner stays or leaves

From these, the system derives utilization rate, attendance rate, payout trend, listing trend, and a city demand index used to tell partner-specific decline apart from a market-wide dip.

2. How a partner gets flagged

Step 1: Detection

A partner is provisionally flagged if any of these hold, measured against their own trailing 3-month baseline, not a fleet-wide average:

Utilization rate down more than 25%
Payout declining for 2 or more consecutive cycles
Slots listed down more than 20%

Step 2: Verification

Before a human ever sees it, each flag is checked against context:

City demand check: if the whole city softened this cycle, the flag is annotated as "city softness" rather than escalated as a partner problem
Persistence check: a one-cycle dip is logged, not escalated, unless it repeats or the partner is Tier A/B
Confidence score: high, medium, or low, based on how many signals fired together

3. How partners are tiered A–E

Tiering combines two inputs: payout per slot (commercial value) and city demand density (is this a high or low demand market). A high payout-per-slot partner in a quiet city carries different risk than the same partner in a saturated metro, so both are weighed together rather than payout alone.

Open input: exact tier boundaries (what payout-per-slot range defines Tier A vs. B, etc.) require real FITPASS payout data before this ships. Not fabricated for this mock.

4. What's deliberately out of scope, v1

No ML churn scoring. Detection is rule-based against trailing baselines. This keeps it explainable and doesn't require a data science build to ship.

No adaptive/self-learning thresholds. The monthly before/after check, did contacted partners recover, is a human-reviewed comparison. If thresholds need adjusting, a PM or analyst does it manually. This is a deliberate scope cut for v1, not a gap.