Keep What the data shows / What this means labels in a single card,
drop author-facing Hub path notes, and only render Statistics sections
that have data.
Add view tabs so the Space stops being one long scroll.
Keep Refresh/Export with the title; put MediCarePlus/CityGov above the best-method and What the data shows row, with spaced ladder/comparator/edge copy.
27 days ago
Space toolbar button icons with the same 8px gap as card titles.
Always show the matched-primary trio plus ModernBERT; put best-method cards in a narrow left column beside What the data shows; refresh methodology, implications, blog placeholder, and Published models Hub links.
Add query-type slice heatmap and RFDT vs Encoder disagreement panels.
Pull by_query_type accuracy and paired predictions.jsonl from Hub so the Space can show where methods diverge on denied slices, not only aggregate macro-F1.
Drop Runs card so the summary strip is one row of five.
Keep the generation ladder matched-8B-only; surface Encoder-FT in its own chart, summary strip, findings, and implications so it is not mistaken for a ladder rung.
Discover results/*/metrics.json on the experiments dataset so Refresh picks up
rfdt_base/rfdt_lora (and future slots) without another Space code change.
Clarify COMPASS as evaluation; restyle implications like other cards.
Use a single What the data shows paragraph above the charts, place
Implications after All published metrics, and stick to fa-solid icons
that actually resolve (chart-line, compass).
Tighten explorer chrome: topbar controls, leaner footer copy.
Drop the redundant status line and Hub/plaintext wording, keep summary
cards at 2x2 until narrow, and credit the researcher with cleaned
resources (methodology, dataset, coming-soon links).
Polish explorer layout: About first, blue theme, solid page edges.
Single lightbulb About card above controls, methodology in a footer
modal, shared page column and plot margins so panel edges align, and
blues restored after the purple trial.
Add outcome-aware info cards; fix lift chart orientation.
Drop run IDs from the table, default Matched 8B off, put backbone on the
Lift x-axis like the learned chart, and render framing plus data-conditioned
implication cards from the loaded metrics.
Drop the metric picker and resource cards; keep domain/matched controls,
always-on ladder/lift/edge figures, a sortable metrics table, and a thin
footer. Monochrome Font Awesome icons on a blue-on-navy palette.
Fix Space boot: drop invalid html.Div ml prop and bind Dash 4 callbacks.
