Benchmark: add eval set provenance banner (821 = 100% PETIMOT test set, fair comparison)

Fix CDF: full color palette for all 7 models, MODEL_ORDER worst→best, solid/thick for PETIMOT family

ATLAS integration: RCSB on-demand PDB + pair metadata (query/target conformations)

Benchmark: split filter (train/val/test) + 5-fold generalisation table + eval set viewer

Benchmark: median markers on CDF + evaluation set viewer (821 proteins)

Fix: lazy torch import in inference.py — page loads without torch installed

Remove torch (~800MB) from build — use pickle for .pt files — build time: 10min→2min

Fix Statistics: remove duplicate margin kwarg in fig_imp.update_layout

Fix CDF: filter to 821 common proteins so PETIMOT comparison is fair

Fix: ROOT NameError + bench cards layout + Statistics TypeError + bio_api sequences

Fix UnboundLocalError: initialize pdb_text=None before PDB path check

Restore profiles in Explorer: disp_profiles.json with 36k sparklines

[10:35 UTC] AA sequence viewer + rich correlation + fix fillcolor/statsmodels/disp_profile

Fix: rgba fillcolor bug + replace statsmodels lowess with scipy binned trendline

Fix all imports: PETIMOT_ROOT, remove render_eigenvalue_spectrum [00:59 UTC]

Fix: export PETIMOT_ROOT from data_loader.py

Fix healthcheck: start-period=120s to let app finish loading before health probes

Fix crash loop: lazy zip loading in check_data_status (no full scan at startup) [16:08 UTC]

Add Protein Detail page: 7k proteins, scatter correlations, multi-model RMSIP², success rate analysis

Optimize build: remove heavy deps (torch_scatter/transformers/biopython) — ETA <5min [2026-03-27 14:29 UTC]

data_loader: auto-download ground_truth from HF Dataset at startup

Add ensure_ground_truth() - auto-extract ground_truth.zip on HF Space startup

Force Space restart to pick up updated app/utils/data_loader.py [2026-03-27 13:51 UTC]

Add 36k NMA metrics + IS/SS/LS PETIMOT ablations + fix NaN correlation

Fix: missing embedding_viewer, NaN in correlation matrix, upload all components

Upload predictions.zip with huggingface_hub

Update: Benchmark page, mode comparison, RMSIP analysis, 36k predictions

Add LineChart and Progress configurations to table

Add disp_profile sampling for sparklines

Upgrade: violins, heatmap, leaderboards, length-stratified analysis

Upgrade: fuzzy search, compare mode, CSV export, stats cards

Upgrade: hero section, quick search, sparklines, featured proteins

Fix: handle empty index.json in check_data_status

Fix: handle empty index.json, fallback to zip scan

Fix: use direct HTTP requests for predictions.zip to avoid hf_hub_download locking on LFS pointer

Fix Dockerfile: add curl for HEALTHCHECK and configure UID 1000 permissions

Fix: requirements.txt — add huggingface_hub + auto-download

Fix: app/utils/download.py — add huggingface_hub + auto-download

Fix: auto-download predictions.zip from HF if not found or LFS pointer

Fix: LFS pointer detection + debug logging for predictions

Fix: auto-download predictions.zip from HF repo at runtime

Add LFS pointer detection + debug logging

Fix: robust zip reading + handle empty DataFrame

Update download check for predictions.zip

Update data_loader to read directly from zip

Remove broken broken 10k files folder, replacing with zip

Add predictions (zipped fast)

Add 36127 predictions (zipped)

Add predictions batch 2/8

Add predictions batch 1/8