Claimability Index: An AI Framework for Predicting Successful Construction Claims

Abstract This paper proposes a probability-based decision-support framework for early stage construction claim governance, in which claim strength is quantified as a claimability index (CI) ranging from 0 to 1 and interpreted as an estimated probability of claim success. A structured claim-level data set was developed, and multiple probabilistic learning models were evaluated under both a conservative escalation policy with a decision threshold of τ = 0.75 and model-specific threshold search. Using Platt-calibrated probabilities with sigmoid calibration and five-fold cross-validation (CV) on the training data, random forest showed the strongest discrimination, with an area under the receiver operating characteristic curve (ROC-AUC) of 0.958 and an area under the precision-recall curve (PR-AUC) of 0.977; at τ = 0.75 , it achieved a precision of 1.000, a recall of 0.710, and an F1 score of 0.830. Logistic regression showed slightly lower discrimination, with a ROC-AUC of 0.924 and a PR-AUC of 0.958, together with a precision of 1.000, a recall of 0.645, and an F1 score of 0.784. Threshold search showed that F1-optimal operating points generally fall in the τ ≈ 0.50 – 0.60 range; for example, random forest peaked at τ = 0.60 with an F1 score of 0.900. This result demonstrates that threshold selection is a governance parameter rather than a purely technical setting. Calibration evidence, including the Brier score and reliability diagnostics, supports the use of calibrated CI values for transparent triage, evidence-first strengthening of mid-CI claims, and risk-aligned escalation decisions across negotiation, alternative dispute resolution (ADR), and formal proceedings.