{# Reusable badge/chip macros #} {% macro pass_rate_badge(rate) -%} {%- set pct = rate * 100 -%} {%- if pct >= 80 -%}{{ "%.1f"|format(pct) }}% {%- elif pct >= 50 -%}{{ "%.1f"|format(pct) }}% {%- elif pct >= 20 -%}{{ "%.1f"|format(pct) }}% {%- elif pct > 0 -%}{{ "%.1f"|format(pct) }}% {%- else -%}0% {%- endif -%} {%- endmacro %} {% macro score_badge(score) -%} {%- if score is none -%} {%- elif score != score or score == score * 2 + 1 -%}overflow {%- elif score >= 1e6 or score <= -1e6 or (score != 0 and -1e-3 < score < 1e-3) -%}{{ "%.3e"|format(score) }} {%- else -%}{{ "%.4g"|format(score) }} {%- endif -%} {%- endmacro %} {% macro best_badge(run) -%} {%- if run.has_score and run.best_score is not none -%}{{ score_badge(run.best_score) }} {%- else -%}{{ pass_rate_badge(run.best_pass_rate) }} {%- endif -%} {%- endmacro %} {# Rescaled 0-100 score chip (higher-is-better; graded like pass rate). #} {% macro rescaled_badge(v) -%} {%- if v is none -%} {%- elif v != v -%}nan {%- elif v >= 80 -%}{{ "%.1f"|format(v) }} {%- elif v >= 50 -%}{{ "%.1f"|format(v) }} {%- elif v >= 20 -%}{{ "%.1f"|format(v) }} {%- elif v > 0 -%}{{ "%.1f"|format(v) }} {%- else -%}0.0 {%- endif -%} {%- endmacro %} {# Best badge that shows rescaled (0-100) by default with raw underneath; the shared score-view toggle swaps which one is visible. Falls back to the plain raw badge when the run has no rescaled score. #} {% macro best_badge_dual(run) -%} {%- if run.has_rescaled and run.best_score_0_100 is not none -%} {{ rescaled_badge(run.best_score_0_100) }} {%- else -%}{{ best_badge(run) }} {%- endif -%} {%- endmacro %} {# Segmented Rescaled | Raw control. Wired by the shared script in base.html. #} {% macro score_view_toggle() -%} Score {%- endmacro %} {% macro submission_badge(sub, run=none) -%} {%- if run is not none and run.has_score and sub.score is not none -%}{{ score_badge(sub.score) }} {%- else -%}{{ pass_rate_badge(sub.pass_rate) }} {%- endif -%} {%- endmacro %} {# Submission badge that shows rescaled (0-100) by default with raw underneath, swapped by the shared score-view toggle. Falls back to the plain badge when this submission has no rescaled score. #} {% macro submission_badge_dual(sub, run=none) -%} {%- if run is not none and run.has_rescaled and sub.score_0_100 is not none -%} {{ rescaled_badge(sub.score_0_100) }} {%- else -%}{{ submission_badge(sub, run) }} {%- endif -%} {%- endmacro %} {% macro agent_chip(agent) -%} {%- if agent -%}{{ agent }} {%- else -%} {%- endif -%} {%- endmacro %} {% macro model_chip(model) -%} {%- if model -%}{{ model }} {%- else -%} {%- endif -%} {%- endmacro %} {% macro status_badge(status) -%} {%- if status == "PASSED" -%}PASSED {%- elif status == "FAILED" -%}FAILED {%- elif status == "ERROR" -%}ERROR {%- elif status == "SKIPPED" -%}SKIPPED {%- else -%}{{ status }} {%- endif -%} {%- endmacro %} {% macro run_status(run) -%} {%- if run.infra_error -%}⚠ infra-error {%- elif run.aborted -%}aborted {%- elif not run.has_final -%}running {%- elif run.timed_out -%}timed-out {%- else -%}done {%- endif -%} {%- endmacro %} {% macro download_icon(href, label, tooltip="") -%} {{ label }} {%- endmacro %}