Published example: astroakd_100_v2

The repository includes a curated public example containing generated astronomy questions, independently collected Astroquery MCP reference answers, the A/B/C baseline answer samples, scoring summaries and figures. No live OpenAI or archive calls are necessary to inspect these saved outputs.

Inspect the questions and references

The joined CSV preserves structured values as JSON inside cells. 92 of 100 questions have complete independent reference evidence; the eight literature-driven cases are excluded pending confirmed publication-to-observation associations.

Automatic archive-reference agreement

Strict reference-answer agreement for 92 evaluable questions

Outcome

A: MCP only

B: AstroAKD instructions

C: Web only

Strict reference matches

87/92

88/92

13/92

Strict agreement

94.57%

95.65%

14.13%

Partially correct

4

4

2

Interpretation: The figures measure agreement with independently retrieved Astroquery MCP reference snapshots, not independently reviewed scientific accuracy. Only fully matching answers count in the strict percentage. The single-question difference between A and B is not established evidence of a domain-instruction benefit; several borderline output-format classifications merit review.

Agreement by question type

Reference-answer agreement by question type

Original baseline runtime

Mean runtime from original common-budget runs

The operational figures use the original common-budget A/B/C executions (100 questions per arm, 300 run records), without substituting subsequent higher-turn diagnostic retries. Reference coverage and runtime are different metrics.

For the methodology and report commands, see Evaluation and plots. Additional PNG/SVG figures, archive breakdowns, and example JSONL files are listed in the public examples README.