0.8 and above denies, 0.2 to 0.8 goes to review, below 0.2 allows. The
band lives in models.json as decision_bands, so another model can take
one without touching the page, and the card's colour comes from whichever
rule the model actually defines rather than a fixed visual scale.
Also widen the headline: the page inspects what a tool returns as well as
what it is about to do.
Space the task form on the wrapper Gradio actually nests
Renjun An
20d10c9
Both columns were meant to share a 16px rhythm, but the rail gets it from
its column's gap while the task form's margins were reset away by Gradio,
so the attack chips sat 15px above the model cards. Gradio nests the real
children two wrappers down, in a .styler and a .form with a 1px gap of
their own; put the gap there and the two columns line up row for row.
Make a model card and an attack chip the same box
Renjun An
2b527a6
Fix the review's findings: feedback, focus, headings
Renjun An
a95b50c
Running a cold checkpoint takes tens of seconds and the only sign was a
grey "processing" line in the corner, behind the Gradio footer. The run
button now disables and reads "Running…", and the progress text is
readable where the result will appear.
The focus ring on the chips and the run button was the browser default:
a 1px near-black outline, invisible on a dark chip and on the accent
button. Both get an explicit ring.
The two column headings were different type at different heights, so the
first card and first chip did not start level. One treatment for every
column heading, an exact line box, and both rows now begin at the same
pixel. Also darken the JSON copy button, the last surface that stayed
white in either colour scheme.
Cut the explanatory copy and enlarge the header
Renjun An
f9871b2
Align the columns and tidy what the review turned up
Renjun An
477fe83
Gradio pads .block and gr.HTML's .prose but not rows or columns, so the
page carried a 12px inconsistency: headers, field labels and the result
heading sat inset from the buttons beside them, three different left
edges per column. Unpad both, and each column now has one edge.
Also from the review: band a score by the model's own threshold, or an
amber bar sat next to a red "Above reference threshold" pill; colour the
span summary by whether anything was marked rather than always amber; a
chip is a block, since text-overflow never applied to the flex box it
was; drop two dead custom properties and the column scale/min_width the
stylesheet has been overriding; merge the duplicated responsive rules
and hoist wire_cases beside the other helpers.
Give the layout a fixed width so results cannot resize it
Renjun An
869bcb8
Pin textarea heights so cases do not resize the form
Renjun An
9e99762
Drop the footer
Renjun An
2864fee
Highlight the case that is loaded
Renjun An
82a0ece
Stop the result and input boxes growing with their content
Renjun An
8f34c0d
Five cases per model and a wider layout
Renjun An
2de5d65
Use real cases from each model's training corpus
Renjun An
132d576
Attack-only cases, a custom-case button, and clearer model names
