Legal regulations, financial criticality, and low trust in the evolving use of AI in engineering project management
make Explainable AI (XAI) essential. Current research in XAI is focused on feature-attribution tools like SHapley
Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Their outputs align with
the expectations of data scientists but are not readily understandable to most engineering project managers, who work
with Gantt charts and earned value scores. This paper introduces the Tri-Modal Explanation Architecture (TMEA) that
generates three coordinated outputs from a single decision-tree surrogate of a Proximal Policy Optimization (PPO)
control policy: a Gantt schedule overlay, a KPI partial-derivative panel, and a Structured Decision Rationale (SDR).
Temporal, magnitude, and causal alignment between the three outputs is enforced by a Cross-Modal Consistency
Mechanism (CMCM); explanations that fail alignment are withheld rather than released. A simulation environment built
from an empirical project database (187 real projects) produces scenarios that are statistically indistinguishable from
real earned-value trajectories, confirmed by K-S tests (D ≤ 0.032 , p ≥ 0.419 ). Across 1,400 bootstrapped scenarios
and a comparison against GPT-4o with and without retrieval augmentation (RAG), the Tri-Modal configuration delivers
higher surrogate fidelity and cross-modal alignment than any single- or bi-modal alternative