Document research evidence and drift decision logic - #23

Summary

• replace redundant badge blocks with workflow-status and source-backed evidence badges
• add an explicit PSI/KS drift-decision specification mapped to the C++ implementation
• report the committed benchmark results with workload and synthetic-data scope
• correct the quickstart to use the repository's Python benchmark and C++ build command

Evidence boundaries

The reported latency and throughput are for the seeded Python reference workload. The F1 result is limited to deliberately separated synthetic perturbations; it is not presented as production accuracy.

Validation

Verified README badge references, display-math delimiters, thresholds, formulas, and reported p95/F1 values against drift-engine/drift_engine.cpp and benchmarks/latest.json.

Summary by CodeRabbit

• Documentation
  • Updated the README with a reproducible drift-monitoring overview.
  • Documented PSI and KS calculations, drift thresholds, benchmark scope, evidence, research questions, and limitations.
  • Revised local validation instructions to run the Python benchmark and compile the C++ engine directly.
  • Clarified the architecture as separating inference and telemetry layers locally.