Causal · Counterfactual · RLHF Research Console
Tenantconsolidated · 14 tenants
Window
live · refreshed 14s ago
Causal ROICounterfactualsSimulation LabLearning TimelineRec QualityPolicy EvolutionTreatment vs Control
Causal ROI Attribution
Doubly-robust ATE · external-factor adjusted · $4.82M AI-attributed (30d)
live · 28d window · n=18,420
AI-attributed savings
$4.82M+18.4%
30d · CI ±$420k
Causal ATE (margin)
+1.62ppp<0.001
DR-learner · n=18,420
Counterfactual lift
+12.8%vs control
stratified by lane
Policy effectiveness
0.847+0.06 wk
expected reward / call
Would-have-saved
$1.14Mmissed
blocked recs · 30d
Learning velocity
+9.2%/wkstable
policy gradient norm
Causal ATE rose to +1.62pp
Doubly-robust estimator over 18.4k shipments. Effect concentrated in Reefer·West and Intermodal lanes; LTL·Southeast remains non-significant.
External tailwinds explain 24% of P&L delta
Fuel, seasonality, and volume mix accounted for $1.60M of $6.42M total movement. AI-attributed net = $4.82M (CI ±$420k).
$1.14M would-have-saved blocked by guardrails
Counterfactual replay of 312 blocked recs at conf 0.70-0.85. Recommend lowering threshold for Carrier-mix domain (low realized risk).
Policy v4.5 reward +6.1% vs production
Shadow eval over 7d. Counterfactual regularizer reduces exploration variance by 38% with no acceptance drop.
Causal Graph
treatment → mediator → outcome · effect & CI on edgesSavings Decomposition
P&L → AI-attributed · would-have-happened removedReported P&L delta
+$6.42M
− Fuel tailwind
$-0.84M
− Seasonality
$-0.31M
− Volume mix
$-0.45M
AI-attributed (net)
+$4.82M
· Carrier recs
+$2.10M
· Mode shifts
+$1.34M
· Accessorial avoidance
+$0.78M
· Lane consolidation
+$0.60M