scaling-group/eve/inspect-population

Inspect a run's population

Read what a run actually evolved: the best candidates, their code, and how scores progressed. Source of truth is the run's lineage DBs + artifacts (local; no wandb needed). run_root layout is in docs/using-eve.md.

The candidates

Each run has <run_root>/solver_lineage.db and <run_root>/optimizer_lineage.db. The eve populations live in the eve_population_entries table, one row per candidate:

sqlite3 <run_root>/solver_lineage.db \
  "SELECT entry_id, created_at FROM eve_population_entries ORDER BY created_at;"

Swap in optimizer_lineage.db for optimizers. (Note: the DB also has generic nodes / edges tables, but the eve loop does not populate them, do not query those.)

Scores (rank best-first)

Scores are not stored inline in the DB; each entry references a flat score artifact ({status, score, summary}) at artifacts/<run-id>_solver/state/<...>__<entry_id>_score.yaml. Rank candidates best-first:

for f in <run_root>/artifacts/*_solver/state/*_score.yaml; do
  s=$(awk -F': ' '/^score:/{print $2; exit}' "$f")
  printf '%s\t%s\n' "$s" "$(basename "$f")"
done | sort -rn | head

The filename carries the <entry_id>; top line = best so far. (The same flat score is also at each candidate's <workspace>/logs/evaluate/score.yaml. Do not parse the workspace-root score.yaml, that one is a different, nested phase-2 record, not the flat score.)

Smoke runs typically show score: 0.0 with a status: error summary; that is the known circle_packing smoke payload error, not a skill problem.

Read a candidate's evolved code

The agent's written solution is under that candidate's workspace: <run_root>/solver_workspaces/<...>/solver/. The archived copy is also stored inline as JSON in the entry's files_ref_json artifact under <run_root>/artifacts/<run-id>_solver/state/.

Improvement over the run

Workspace directories are named ..._step_<n>_...; compare the best score.yaml per step to see the trajectory. The optimizer log trees under <run_root>/artifacts/<run-id>_optimizer/ show how the guidance evolved.

Notes

• Completed-iteration count: grep -c "Phase 3: updated Elo" <run_root>/runner.log.
• To understand why a candidate behaved a certain way (not just its score/code), read its agent session: ../debug-agent-session/SKILL.md.