Fix Gradio bindings: remove gr.Request from inputs; rely on implicit request injection. Docs already updated earlier.

Login required + per-session batch (20), min raters filter (20), periodic reload, spaced repeats, and metrics logging (duration, flat scores); add hidden states and CSV fields; enforce HF-auth annotator id

Annotator ID from HF user: remove manual input, extract from gr.Request (X-Forwarded-User/username); wire through handlers; keep per-user dedup + global balancing

Per-user scheduling + global balancing: add Annotator ID, per-user dedup, count-based prioritization, CSV adds annotator_id; update UI bindings accordingly

Fix indentation inside Blocks: move event bindings and controls under gr.Blocks context; py_compile passes

Ensure no duplicate evaluations: idempotent submit, update pair_state to remove evaluated pair, and clear UI on completion; add graceful returns when no pairs

Fix: robust empty-pairs handling, CSV trimming, and Python 3.8+ type hints; improve slider bounds and graceful no-op when no pending pairs

problem fix

UI: switch to English-only labels/messages; minor cleanup. No logic changes.

Validation: skip pairs with missing image paths; docs: add Updates section describing pairing/scheduling, storage, fairness, and config (AGENTS.md updated but intentionally ignored)

Pairing improvements: filter already-evaluated pairs from /data, round-robin schedule across test_ids, alternate A/B order per pair; ensure submit maps scores to correct model columns and auto-advance

Improve Hub upload robustness: return detailed error messages to UI and print exceptions to logs

Persist evals to /data CSV and upload per-submission JSONL to dataset repo (peiranli0930/VisEval); add UI feedback

Persist evaluations: write CSV locally and optionally push to Space repo when HF_TOKEN/SPACE_ID available

UI update: mask model info, new layout (original on top, A/B bottom), per-image 4 scores, and CSV schema update

Add Scene Composition & Object Insertion evaluation UI