LizzyDocs

Stage 05 · Prove

Read the evidence, not just the final score.

Inspect holdout metrics, reward curves, regressions, curated examples, spend, and the exact student artifact produced by a run.

Start with the run timeline

The run resource is the source of truth for stage order, timestamps, terminal status, metered totals, dataset version, and recipe. A copilot workflow stores the same run ID so a reconnect can resume inspection.

GET/v1/distill/runs/{run_id}

Inspect the run and its stages.

GET/v1/distill/runs/{run_id}/logs

Read bounded logs. Filter by stage or cursor instead of replaying an unbounded stream.

Read normalized metrics

Use GET /v1/distill/runs/{run_id}/metrics for per-step series such as loss, reward mean and variance, KL, latency, and cost. Compare holdout baseline and final scores; training reward alone is not a deployment claim.

Quality

Holdout deltaDid unseen examples improve?

Reliability

RegressionsWhich cohorts or rewards got worse?

Efficiency

Cost & latencyDoes the student meet the original objective?

Inspect the evaluation report

Reports combine summary metrics, verdict, regressions, and curated examples. The copilot may summarize evidence and link the report, but creating a public share link is human-only.

GET/v1/distill/reports/{report_id}

Read the verdict and evaluation summary.

GET/v1/distill/reports/{report_id}/examples

Page through curated examples; retrieve one example for deeper review.

Decision checklist

  • Holdout metrics meet the goal and uncertainty is acceptable.
  • No high-severity regression is hidden by an average.
  • Reward behavior still matches the intended product outcome.
  • The published model ID belongs to this run and report.
  • Serving fallback points to a known-good teacher.
Built for humans and copilots.

Every risky action has an explicit handoff, validation, or approval boundary.

Recovery guide