EXPERIMENTS 27 + 29 + 30 + 31 + 33 · CLUSTER CONFIGURATION DIAGNOSTICS

How does clustering change review coverage?

HOW TO READ THE MAP

Generated candidates and human reviews share one map.

Compare DeepSeek V4 Flash's direct semantic partition, six Experiment 33 controlled-tag and tag/embedding partitions, the 11 foundational Experiment 29 configurations, and all 14 Experiment 31 conditions. Experiment 33 cluster labels summarize the dominant issue family, failure mode, and target tags. Blue points have no Sol-high match at the frozen 0.70 threshold; green points do. Pink diamonds are atomic ground-truth human review issues. One DeepSeek paper uses a clearly labeled diagnostic alternate because its primary response was structurally invalid.

Generated candidate Associated with ground truth Ground-truth human issue Chosen cluster representative

CLUSTER ↔ GROUND-TRUTH GRAPH

Association table

PaperNo GT matchExactly one GT matchMore than one GT matchTotal clusters

INTERACTIVE SCATTER

Candidate clusters

100%
Click a point to read it. Scroll to zoom; drag the background to pan.

CLUSTERS

Cluster association details

ClusterConceptItemsMatched itemsGT issuesRepresentative

GROUND TRUTH

Human review issues