When Accurate Fits Cannot Certify Deployment: Deployment Non-Identifiability in AI-Assisted Chip Reliability

AI-based reliability predictors are increasingly attractive for chip-design and sign-off workflows, but accelerated electromigration (EM) testing requires extrapolating to use conditions far outside the tested envelope. A model can fit the accelerated regime accurately while its deployment prediction remains unresolved. We introduce EMBench-Synth, a controlled EM benchmark in which the true generating process is known, and use it to quantify this deployment non-identifiability: an optimizer-independent construction reaches $15.97\times$ hidden deployment disagreement while remaining within $0.15$ log-lifetime units of the accelerated data. Black-box ML predictors (MLP, KAN, symbolic regression) inherit the hazard, while a physics-regularized ML variant substantially improves deployment transfer under the benchmark's $2\times$ criterion. A robust compatible-set certificate starts at $91.8\times$ median worst-case deployment disagreement and falls only to $88.4\times$ after the best one-batch test-design policy, with all 20 seeds correctly returning ABSTAIN under a $2\times$ sign-off limit. The results suggest that AI-assisted chip-reliability systems require deployment-aware identifiability tests and explicit physics constraints rather than relying on accelerated-window fit quality alone. We release the benchmark, procedures, figures, and machine-readable results.