Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention
Abstract
Changing a deployment policy can invalidate a world model's uncertainty while leaving the physical transition mechanism and predictor unchanged. For one-step prediction, if source and target laws share the conditional mechanism P(Y|X)=Q(Y|X), then the intervention is identified by the pre-outcome change of measure w(X)=dQX/dPX. For policy-controlled dynamics this ratio factors into state-occupancy and action-propensity terms. Weighted conformal calibration transports marginal coverage to the target law; with an estimated ratio ŵ, PQ{Y ∈ Cŵ(X)} ≥ 1−α−TV(Q,Q̂) ≥ 1−α−EP|w−ŵ|. Population log-loss regret for a source-versus-target domain classifier bounds the same ratio error. If h is its clipped target posterior, ρ is the target-domain prior, and Δ(h) is population log-loss regret, then EP|w−ŵh| ≤ (1/(ργ))√(Δ(h)/2). An exact cancellation improves the dependence from a generic odds-map bound of order γ−2 to (ργ)−1. Our experiments use a released DINO-feature world model and physical Franka trajectories. Under the strongest controlled state-action shift, ordinary 90% conformal sets cover 55.0%, whereas oracle and learned-ratio calibration cover 91.0% and 90.9%. The learned surrogate remains within 0.9% total variation of the target law. Calibration is relative to the deployment law: the model can remain fixed while calibration follows the intervention. We machine-check in Lean 4 the finite-support algebraic and causal identities used in the classifier-regret argument, including its coverage-composition step.
Related
- Conformal Policy Selection: When to Deploy AI Decisions for Targeted Intervention
- Reference-Null Calibrated Thresholds for E-Processes with Applications to Conformal Martingales
- Moment-Calibrated Conformal Prediction with Side Information
- SCOPE Before You Sample: Conformal Scores Guide Posterior Exploration for Fast and Calibrated Bayesian Predictive Inference
- Bagged Martingale Posteriors: Calibrated Uncertainty Quantification for Predictive Resampling
BibTeX
@misc{mittal2026calibrated,
title = {{Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention}},
author = {Ayush Mittal and Dhruv Gupta},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.35877},
eprint = {2610.35877},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.35877},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention
AU - Ayush Mittal
AU - Dhruv Gupta
PY - 2026
DA - 2026/10/01
PB - arXiv
DO - 10.48550/arxiv.2610.35877
UR - https://arxiv.org/abs/2610.35877
AB - Changing a deployment policy can invalidate a world model's uncertainty while leaving the physical transition mechanism and predictor unchanged. For one-step prediction, if source and target laws share the conditional mechanism P(Y|X)=Q(Y|X), then the intervention is identified by the pre-outcome change of measure w(X)=dQ_X/dP_X. For policy-controlled dynamics this ratio factors into state-occupancy and action-propensity terms. Weighted conformal calibration transports marginal coverage to the target law; with an estimated ratio ŵ, P_Q{Y ∈ C_ŵ(X)} ≥ 1−α−TV(Q,Q̂) ≥ 1−α−E_P|w−ŵ|. Population log-loss regret for a source-versus-target domain classifier bounds the same ratio error. If h is its clipped target posterior, ρ is the target-domain prior, and Δ(h) is population log-loss regret, then E_P|w−ŵ_h| ≤ (1/(ργ))√(Δ(h)/2). An exact cancellation improves the dependence from a generic odds-map bound of order γ^−2 to (ργ)^−1. Our experiments use a released DINO-feature world model and physical Franka trajectories. Under the strongest controlled state-action shift, ordinary 90% conformal sets cover 55.0%, whereas oracle and learned-ratio calibration cover 91.0% and 90.9%. The learned surrogate remains within 0.9% total variation of the target law. Calibration is relative to the deployment law: the model can remain fixed while calibration follows the intervention. We machine-check in Lean 4 the finite-support algebraic and causal identities used in the classifier-regret argument, including its coverage-composition step.
ER -