Prospective Boundary Prediction from Cumulative Typed Trajectory Transitions

Trajectory Graph Monitoring is a prospective runtime-evaluation study testing whether structural information from an incomplete execution trajectory can improve prediction of a boundary violation before the violation occurs. The frozen benchmark contains 150 episodes and 522 prospective prefixes across three simulated workflow families generated with Qwen2.5-1.5B-Instruct. Every evaluated prefix terminates strictly before any violating action. Using five-fold episode-grouped cross-validation, a model augmented with cumulative typed-transition counts achieved an AUPRC of 0.4524 compared with 0.3683 for the semantic baseline and 0.3510 for telemetry alone in the frozen primary evaluation. A separate grouped-split robustness run reproduced the direction of the result and included typed-count ablations, randomization and noise controls, prediction-horizon sensitivity, held-out-family evaluation, calibration analysis, and coefficient diagnostics. The release includes the technical report, frozen episode and prefix datasets, executable evaluation notebooks, out-of-fold predictions and metrics, bootstrap outputs, diagnostic analyses, publication figures, manifests, hashes, and an independent-replication packet. The result is intentionally bounded to the evaluated benchmark and does not establish causal mechanism, cross-model generalization, deployment readiness, or intervention effectiveness. Independent external replication remains pending.