Benchmark & Dataset
Weekly Digest

20 complete repository migrations across four categories and 520 experimental runs.

Applies Migration Audit, Behavioral Tests and adversarial differential verification in sequence.

Only 28 of 520 runs (5.4%) pass all stages.

It distinguishes real migration completion from accidental test passes or retained legacy stacks.

893 human-authored questions with expected output modalities and per-modality rubrics.

MMI Value measures content quality; MPS measures whether every expected modality is present.

It separates correct modality delivery from correct content, exposing routing and completeness failures.

1.3B video URLs, 80M downloaded videos and about 10M hours, yielding roughly 55M clips and 300M keyframes.

Validated through video-text, audio-text and image-text retrieval with scaling analyses.

It turns a shared source layer into trainable multimodal assets at several granularities.

Four synthetic companies from 12 to 10,000 employees; the largest corpus exceeds 230,000 documents.

Tests information extraction and knowledge-base querying as document scale increases.

Performance continues to fall as corpora approach real enterprise scale.

It makes document scale and temporal consistency first-class variables for enterprise agents and RAG.

30 hours of annotated two-person dialogue, 154 conversations, 106 speakers and six interaction styles.

Measures EOT / interruption recall, false-positive rate and median latency.

It makes interaction timing a comparable benchmark dimension for full-duplex voice agents.

36 hard cases with references, prompts, evaluation specs, provenance and geometric-challenge tags.

Uses a shared rerender rig and blind assembly judge across visual, semantic and geometric metrics.

Gradeability specs define single-view limits while separating visual, geometric, structural and execution quality.