Skip to content

Why tracing exists

Tracing is not optional overhead; it is the point of the project.

What tracing enables

  • Auditability: decisions are accompanied by structured provenance.
  • Regression detection: compare fingerprints and outcomes across versions.
  • Replay validation: verify that a recorded run satisfies the trace contract.
  • Debugging: correlate phases, agents, and failures without guessing.

What tracing does not claim

  • Tracing does not make a model “correct”.
  • A REPLAYABLE trace does not guarantee the provider will return identical tokens in the future.
  • A NON_REPLAYABLE trace can still be useful; it just must be labeled honestly.

Practical implication

If you remove trace metadata, you remove the system’s ability to prove what happened. The code is expected to fail fast rather than emit unverifiable artifacts.