Modern AI systems expose rich internal signals, gradients, activations, logits, but PyTorch doesn't provide structured interfaces to surface them. Teams repeatedly reimplement ad hoc instrumentation for stability, fairness, provenance, & logging.

We present a concrete, code-level analysis of missing observability primitives in PyTorch, derived from building a compliance and monitoring system using only core APIs (hooks, autograd, param.grad). We identify 4 recurring gaps: lack of training provenance, absence of dataset-level semantics (e.g. sensitive attributes), no structured hook outputs, & zero visibility into optimiser-level dynamics.

We then generalise these gaps to LLM inference & agent systems, where gradients are replaced by token probabilities, and optimiser steps by multi-step decisions. The same structural problem persists: signals exist but aren't standardised or exposed.

We propose framework-level primitives, structured hook outputs, gradient health reports, & run-scoped audit contexts that enable interoperable tooling without expanding PyTorch scope. This is a discussion grounded in implementation details, API design trade-offs, & reproducible engineering patterns.