The Constructibility Framework and its monitoring extensions required either controlled corruption proxies or held-out calibration sets. Real deployment environments provide neither: labels are sparse, shift is gradual, and no ground-truth H(t) exists. We introduce zero-shot margin calibration: estimating the constructibility exponents (alpha, beta, gamma) and proxy H-hat(t) directly from production data streams using only unlabelled model outputs. The calibration uses three unsupervised signals: (i) output entropy drift, (ii) embedding distribution shift (2-Wasserstein distance from deployment baseline), and (iii) prediction confidence decay. No labels, no synthetic corruption, no held-out shift sets.Three theoretical results: (1) Zero-Shot Calibration Consistency (Proposition 4.1): under mild regularity conditions, the unsupervised exponent estimates converge to the true exponents as deployment duration increases; (2) Unsupervised Lead-Time Bound (Proposition 4.2): the lead-time guarantee of Paper 2 extends to zero-shot calibration with an additive calibration error term; (3) Deployment Adaptation (Proposition 4.3): online exponent updates preserve the monotonicity property of the margin.Validated on three real deployment datasets spanning 2018-2025: clinical NLP (MIMIC-IV discharge summaries), financial sentiment (S&P 500 earnings calls), and code review (GitHub PRs). Detection AUC 0.79 +/- 0.06, lead time 3.2 +/- 1.6 periods, outperforming all unsupervised baselines and even label-dependent loss CUSUM (AUC 0.64). The graceful degradation across the series (synthetic 0.91 -> proxy-calibrated 0.84 -> zero-shot 0.79) is consistent with the theoretical bounds of Papers 2-3.
Karimov et al. (Sun,) studied this question.