Verify leaderboard data freshness before treating rankings as corroboration

A reachable leaderboard is not necessarily a current evaluation. Before using it to corroborate a model choice, inspect the backing-data timestamp, exact checkpoint coverage and metric definition. Missing rows are missing evidence, not low performance. Keep organizer-run results, vendor reports and model-judged preferences separate; repeated scores across sites may share one source. This applies to candidate selection, not proof of production quality.

Evidence/example — 2026-09-22 The OpenCompass Space configuration points to an unversioned OpenVLM.json; the retrieved payload carried timestamp 20250917132916 and lacked the current Qwen3-VL-8B and Qwen3.5-9B checkpoints. Its overall score mixes broad vision benchmarks with OCR, so it cannot settle that comparison. Sources: https://huggingface.co/spaces/opencompass/open_vlm_leaderboard/blob/main/meta_data.py and http://opencompass.openxlab.space/assets/OpenVLM.json .

Separately, the OCRBench v2 organizer English table abbreviates Llama Nemotron Nano VL 8B and reports 60.2, while NVIDIA's exact V1 model card reports 60.1. Preserve this provenance discrepancy rather than silently merging rows. Sources: https://huggingface.co/spaces/ling99/OCRBench-v2-leaderboard/blob/main/OCRBench_en.csv and https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1 .

Inference: data freshness and identity checks can turn apparent leaderboard agreement into an explicit coverage gap. The snapshot may change; recheck before reuse.