DoCoreAI: Metadata-Sufficiency for Privacy-Preserving LLM Cost Governance and Predictive Budgeting

Production AI systems face a fundamental tension: effective cost governance requires observability, but observability has traditionally required logging prompt and response content — creating compliance liability for regulated industries. We present DoCoreAI, a production LLM governance system that resolves this tension through a central thesis we call metadata sufficiency: that prompt content is unnecessary for accurate output token prediction and autonomous budget management.

Through three coordinated layers —

(1) metadata-only telemetry collected via transparent SDK interception without modifying application code,

(2) a LightGBM quantile regression model trained exclusively on structural metadata signals, and

(3) an autonomous governance layer operating on non-reversible semantic fingerprints — DoCoreAI achieves pre-call token prediction, real-time hourly budget pacing, PII detection, and policy enforcement without storing, transmitting, or processing prompt or response text at any point in the pipeline.

The system supports six major LLM providers and operates as a process-local sidecar that patches SDK call methods at Python startup, requiring no application code changes. This paper describes the system architecture, the metadata feature set, the ML training and drift management pipeline, the governance model, the threat model, and the current implementation scope with an honest account of roadmap items in active development.