Trace Forensics introduces a framework for inferring latent system states from multiple partial, noisy, and potentially adversarial witnesses. Rather than relying on direct observation, the model aggregates indirect evidence through witness transformations, confidence weighting, and probabilistic aggregation. The framework formalizes inference as an optimization over a confidence function constructed from heterogeneous inputs, incorporating uncertainty, adversarial bounds, and information-theoretic limits.Fisher information and the Cramér–Rao bound are used to characterize the fundamental limits of inference accuracy. The framework further supports computational staging, iterative refinement, and both online and offline inference modes, enabling real-time approximation as well as deferred high-precision reconstruction. Offline inference permits increased computational budgets, finer discretization, and multi-pass refinement strategies.