NERVE — Network Earnings & Risk Visibility Engine

Project NERVE v14-AI | Model v13 + AI Analytics | n=26,167 | CRITICAL 95.6% precision | HIGH 93.8% | Combined 92.8% (3mo grace) | June 2026

Metrics & Formulas Reference

Complete reference of every metric used in Project NERVE. Includes formula, calculation method, and business meaning.

CategoryMetricFormula / ThresholdWhat It Means
PnL — Revenue & Coststransport_savings_mo$1.15 x shipped_unitsDollar value of delivery savings. Primary revenue source.
PnL — Revenue & Costsad_revenue_moActual per-KID from Kiosk Attribution / PRUNE / FinancialsAd campaign revenue. $0 if no campaign running.
PnL — Revenue & Costslease_revenue_morenewal_price / term_months OR original_price / 60Fee partner pays Amazon to host locker. Residential only. $17.6M/yr.
PnL — Revenue & Costsconcession_cost_mo$32 x conceded_unitsRefund cost per customer concession.
PnL — Revenue & Costsrent_moActual rent from rent fileRent Amazon pays to store partner. Commercial only.
PnL — Revenue & Costsad_wrap_cost_mo$209 if has_actual_ads else $0Wrap cost. ONLY charged when campaign is running.
PnL — Revenue & Costsdepreciation_mo$105 Core/Odin, $99 Apartment, $18 DobbyMonthly hardware depreciation.
PnL — Revenue & Costsmaint_fixed_mo3-month rolling avg of vendor invoice costsActual maintenance spend smoothed over 3 months.
PnL — Revenue & Costsother_fixed_mo$30 Core/Odin, $15 Apartment, $5 DobbyMiscellaneous operational costs.
PnL — Revenue & Costsnet_savings_moSum of revenue - sum of costs (3-month avg)Net monthly PnL. Positive = profitable.
PnL — Revenue & Costsestimated_2026_pnlYTD actuals + net_savings_mo x remaining monthsFull-year projection.
Performanceshipped_moSum of daily shipped / months (3-month avg)Monthly package volume processed.
Performanceconceded_moSum of daily conceded / months (3-month avg)Monthly delivery failures.
Performancedpmo(conceded / shipped) x 1,000,000 (3-month avg)Defects per million. <8K good, >15K critical.
Performancethroughputpackages / (active_days x slot_count) (3-month avg)Utilization. >0.15 healthy, <0.08 critical.
Performancetput_velocityOLS slope of throughput over 6 monthsRate of volume change. Negative = eroding demand.
Tickets & Maintenancetickets_ytdCount of unique IssueIds in current yearComplaints filed. >5 = high risk.
Tickets & Maintenancetickets_6mCount of unique tickets in last 6 monthsRecent complaint activity.
Tickets & Maintenancedispatches_6mCount of months with vendor dispatches in 6 monthsTechnician visits. 3+ = alert.
Offlineoffline_days_ytdCount of days where offline_days > 0 in 2026Total days locker was down YTD.
Offlineoffline_ytd_pctoffline_days_ytd / total_days_in_period x 100% of year locker has been offline.
Offlineconsecutive_offline_daysCount backward from latest date while offlineCurrent continuous outage duration.
Model — Slopenet_slopeOLS regression slope over 6 months ($/mo per mo)Rate of financial deterioration.
Model — Slopenet_slope_dtOLS slope of deseasonalized PnL over 6 monthsConfirms trend is real, not seasonal.
Model — Sloper_squaredR-squared of 6-month OLS regression (0-1)Confidence in slope. Higher = more linear decline.
Model — Slopeseasonal_idxavg_month_pnl / overall_avg_pnl (clipped 0.3-3.0)Per-locker seasonal adjustment factor.
Model — TiersEMERGING RISK (P0)Positive today + 4+ months consecutive declineStill profitable but degrading fast.
Model — TiersCRITICAL (P1)Persistent negative + steep slope OR maintenance spike95.6% precision. Near-certainty of continued loss.
Model — TiersHIGH (P2)Includes 'Run Ads' quality gate candidates93.8% precision. Operational fix may resolve.
Model — TiersELEVATED (P3)Moderate decline detected91.1% precision.
Model — TiersLOW (P4)Early throughput erosion87.2% precision.
Model — TiersSTABLENo decline signal. Flat or improving PnL slope.Not flagged by the model.
Model — Tiersclassification_pathslope | maintenance | run_ads | emerging | N/AWhich mechanism flagged this locker.
EWSews_vs_netComposite: how much worse than network avg (0-100)Relative underperformance vs peers.
EWSews_vs_selfComposite: how much worse than own baseline (0-100)Deterioration vs own history.
EWSrisk_levelHIGH if both >= 40. MODERATE if one >= 40.AND logic. Both axes must be bad for HIGH.
EWS vs TierSTABLE + EWS HIGHLocker not declining (flat) but underperformingProactive watchlist. Tier won't catch until slope starts.
Cluster (DBSCAN)cluster_healthBad: 75%+ neg AND 2x+ worse. Good: DPMO<8K, tput>0.15.Zone-level classification (500m radius).
Cluster (DBSCAN)multipliermax(DPMO/avg_DPMO, avg_tput/cluster_tput)How many times worse than network average.
Proximitydistance_mHaversine straight-line distance (meters)Physical proximity between two lockers.
Proximityboth_negBoth lockers in pair have net_status = negativeConsolidation candidate if also <500m.
DFRtotal_dfrSum of all failure reason counts (last 90 days)Pre-delivery failures (different from DPMO/concessions).
Alert ThresholdsMaintenance Spikedispatches_6m >= 3 OR maint_fixed_mo > $50Hardware degradation signal.
Alert ThresholdsDPMO Spiraldpmo > 15,000Quality critically impaired.
Alert ThresholdsVolume Droptput_velocity < -0.01 AND throughput < 0.15Demand eroding.
Alert ThresholdsHigh Ticketstickets_ytd > 5High complaint volume. Partner dissatisfaction risk.
Alert ThresholdsCurrently Offlineoffline_days > 0 on latest dateLocker not processing packages now.

Key Principles

Ad Wrap Cost Rule

$209/mo charged ONLY when has_actual_ads = True (campaign running with revenue > $0). NOT when wrap is merely installed. Violating this inflates losses by ~$15M.

Tickets Deduplication

Source file contains ~51% duplicate rows. Pipeline deduplicates by IssueId before counting. Always use unique ticket counts.

Tier vs EWS

Tier = Is PnL slope declining? (trend-based)
EWS = Is locker underperforming vs peers AND vs self? (relative comparison)
A STABLE locker with EWS HIGH = consistently bad but not getting worse. Proactive watchlist signal.

3-Month Averaging

PnL, volume, DPMO, throughput, maintenance all use 3-month rolling averages. Model targets PERSISTENT problems, not one-month dips. Tickets, offline days, dispatches use exact counts (no averaging).

Lease Revenue = Income

Partner pays Amazon. This is REVENUE, not cost. "Billing annually" means price is annual (divide by 12). Expired + not renewed = $0 rent = free transport savings. Do NOT disturb.

Maintenance Sources

Actual vendor invoices (B&H + Velociti), NOT annual/12 flat estimates. A locker can have $0 for 9 months then $500 spike. 3-month rolling average preserves trends while smoothing one-offs.