tests/test_reports.py

from inferscale.reports import generate_research_report


def test_report_contains_guardrails_and_ranking():
    robust = {
        "repetitions": 4,
        "bootstrap_samples": 100,
        "robust_winner": "Top-1 decayed",
        "nominal_winner_first_seed": "Utility-aware multi-step",
        "policies": [
            {
                "robust_rank": 1,
                "label": "Top-1 decayed",
                "median_ttft_ms": 100,
                "ttft_ci95_low_ms": 95,
                "ttft_ci95_high_ms": 105,
                "ttft_win_rate": 0.75,
                "pareto_stability": 1.0,
                "median_oracle_regret_ms": 3,
                "worst_seed_ttft_ms": 120,
            }
        ],
        "oracle": {"candidate_count_per_seed": 19, "median_ttft_ms": 97, "definition": "bounded", "note": "bounded note"},
    }
    text = generate_research_report(robust)
    assert "# InferScale Research Consolidation" in text
    assert "Robust winner" in text
    assert "globally optimal" in text