feat: refine leaderboard metric presentation and bucket data handling

- Add bucket support for error_rate and metrics_evaluated_samples.
- Surface error rate in Details model cards and the Tools performance table.
- Detect partial model runs from metrics_evaluated_samples < sample_count and show a partial badge instead of clean.
- Keep eval-leak badges, but remove the always-on clean badge.

- Rename Weakest link to Min score across EN/RU UI.
- Add column direction subtitles for Integral, Min score, FPR, FNR, and p95.
- Remove up/down metric arrows from values, badges, legends, and Details matrix cells.
- Remove hard-coded threshold coloring and deleted threshold-based metric helpers.
- Highlight only visible best/worst rounded values per column in Ranking and Details matrix.
- Keep Size uncolored.

- Add data-driven heatmap scaling based on visible min/max values.
- Apply sqrt normalization to heatmap color interpolation.
- Show real min/max labels in the heatmap legend.

- Remove dead prompts UX from Visualizations:
- prompt sheet,
- heatmap/scatter drilldown click handlers,
- full-version prompt-viewer placeholder text.
- Remove Visualizations winner/leader summary UI and related copy.
- Keep Pareto frontier behavior while removing winner-style highlighting.

- Simplify Details benchmark cards by removing leader metric selector and leader badges.
- Remove threshold legends from Details model/family/matrix sections.
- Leave only the no data legend under the Details matrix.

- Clean up Methodology:
- remove helper metrics section,
- remove Micro-F1 card,
- remove example blocks,
- remove bottom higher/lower legend,
- remove the FPR/FNR golden-rule callout,
- stack formula cards for more stable layout.

- Replace or remove extra copy across EN/RU messages:
- shorten footer to HiveTrace Guardrail Leaderboard,
- simplify Ranking subtitle,
- simplify pinned empty state,
- move and expand robustness delta explanation above the table,
- remove sorting/full-version/threshold instructions.

messages/en.json

@@ -11,7 +11,7 @@
 	"ranking_badge": "Guardrail benchmark",
 	"ranking_updated": "updated {date}",
 	"ranking_title": "Ranking",
-	"ranking_description": "A ranking of open-source guardrail models. Higher Integral is better.",
+	"ranking_description": "A ranking of open-source guardrail models.",
 	"ranking_filters": "Ranking Filters",
 	"ranking_table": "Models Table",
 	"ranking_scope_select_label": "data slice",
@@ -55,16 +55,18 @@
 	"ranking_rows": "{count} rows",
 	"ranking_column_model": "Model",
 	"ranking_column_integral": "Integral",
-	"ranking_column_weakest_link": "Weakest link",
+	"ranking_column_min_score": "Min score",
 	"ranking_column_fpr": "Overall FPR",
 	"ranking_column_fnr": "Overall FNR",
 	"ranking_column_p95": "p95 ms",
 	"ranking_column_size": "Size",
+	"ranking_column_higher_better": "higher is better",
+	"ranking_column_lower_better": "lower is better",
 	"ranking_column_report": "Report",
 	"ranking_model_type_classifier": "classifier",
 	"ranking_model_type_llm_judge": "llm judge",
-	"ranking_eval_clean": "clean",
 	"ranking_eval_leak": "eval-leak",
+	"ranking_eval_partial": "partial",
 	"ranking_open_details_label": "Open details for {model}",
 	"ranking_details_overview_tab": "Overview",
 	"ranking_details_groups_tab": "By group",
@@ -130,39 +132,22 @@
 	"methodology_title": "Methodology",
 	"methodology_description": "Dataset → group → integral score. FPR and FNR stay separate until the final aggregation step.",
 	"methodology_meaning_label": "What it means:",
-	"methodology_example_label": "Example",
 	"methodology_aggregation_title": "Three aggregation levels",
 	"methodology_dataset_score_tag": "dataset score",
 	"methodology_dataset_score_comment": "harmonic mean of (1−FPR, 1−FNR)",
 	"methodology_dataset_score_description": "Errors are normalized into quality: 1−FPR is the share of safe content correctly passed, and 1−FNR is the share of harmful content correctly caught. Then we take the harmonic mean, so the weaker side pulls the score down.",
-	"methodology_dataset_score_example": "<b>Qwen3Guard-Gen-8B on ToxicChat</b>: FPR 8.5%, FNR 6.3% → H(0.915, 0.937) = 0.926.",
 	"methodology_group_score_tag": "group score",
 	"methodology_group_score_comment": "harmonic mean of N dataset scores",
 	"methodology_group_score_description": "A group is a family of related datasets, for example jailbreak tests. We use the harmonic mean again: the weakest dataset defines the group score, so easy sets cannot hide a poor result inside the same topic.",
-	"methodology_group_score_example": "<b>Qwen3Guard-Gen-8B, XSTest group</b>: FPR 12.8% / FNR 3.5% → S_group = 0.916.",
 	"methodology_integral_score_tag": "final score",
 	"methodology_integral_score_comment": "geometric mean of M groups (S_group > 0)",
 	"methodology_integral_log_comment": "log form for numerical stability",
 	"methodology_integral_score_description": "Across groups we use the geometric mean. It penalizes imbalance more strongly: one failing group pulls the product down. A balanced profile is valued above one isolated star result with a gap elsewhere.",
-	"methodology_integral_score_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: geometric mean across groups → S_integral = 0.918. The weakest group at 0.823 pulls the final score down.",
-	"methodology_helper_metrics_title": "Helper metrics",
-	"methodology_fbeta_tag": "general form",
-	"methodology_fbeta_description": "The β knob balances precision and recall: β > 1 favors recall, while β < 1 favors precision.",
-	"methodology_f2_tag": "β=2",
-	"methodology_f2_description": "Recall is weighted 4x more. F2 penalizes missed harmful content more strongly and fits datasets where a miss is more costly than a false alarm.",
-	"methodology_f2_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: recall R=0.862 (1−FNR) and specificity 0.910 (1−FPR) are real values. Precision (TP/(TP+FP)) is not available in the leaderboard data, so we do not show a value for this metric: specificity ≠ precision.",
-	"methodology_f05_tag": "β=0.5",
-	"methodology_f05_description": "Precision is weighted 4x more. F0.5 penalizes false positives more strongly and fits over-refusal tests.",
-	"methodology_f05_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: recall R=0.862 (1−FNR) and specificity 0.910 (1−FPR) are real values. Precision (TP/(TP+FP)) is not available in the leaderboard data, so we do not show a value for this metric: specificity ≠ precision.",
-	"methodology_fdr_description": "The false alarm share among blocked items: the probability that a trigger was wrong. Lower FDR means most flagged items deserved the verdict.",
-	"methodology_fdr_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: overall FPR 9.0%, FNR 13.8% are real values. FDR requires precision, which is not available in the data, so we do not report a numeric FDR.",
 	"methodology_combined_dataset_title": "Combined dataset",
 	"methodology_harm_split_tag": "harm split",
 	"methodology_harm_split_description": "The share of harmful requests incorrectly marked safe. TP are caught attacks, FN are missed attacks.",
 	"methodology_safe_split_tag": "safe split",
 	"methodology_safe_split_description": "The share of safe requests incorrectly marked harmful. TN are correctly passed safe items, FP are false alarms.",
-	"methodology_micro_f1_tag": "combined",
-	"methodology_micro_f1_description": "Computed from global TP, FP, and FN across the whole set. With equal class weights in a binary task, micro-F1 matches accuracy.",
 	"methodology_combined_dataset_note": "If a shared set is marked as combined, its metrics are taken directly from the full splits; otherwise metrics are computed in the standard way through the arithmetic mean of individual dataset metrics.",
 	"methodology_scatter_title": "Scatter: FPR vs FNR",
 	"methodology_scatter_x_axis": "Axis X",
@@ -176,11 +161,6 @@
 	"methodology_single_class_description": "F1 is computed for all benchmarks <b>by default</b>. If a dataset contains labels from only one class: all_unsafe gets <b>FNR, recall</b>; all_safe respectively gets <b>FPR, TNR</b>.",
 	"methodology_caveat_title": "Caveat",
 	"methodology_caveat_description": "Metrics are point estimates. Confidence intervals and sample sizes (n) for tests are not published in the current dataset, so small ranking differences may be statistically insignificant.",
-	"methodology_golden_rule_title": "Rule: never collapse FPR and FNR into one raw error",
-	"methodology_golden_rule_description": "FPR (false blocking of safe content) and FNR (missing harmful content) are <b>two different errors with different costs</b>. Averaging them into one number hides failure: a model with FPR 0% and FNR 40% looks “average” while missing almost half of attacks. That is why both errors stay separate until S_ds, and aggregation uses a <b>harmonic/geometric</b> mean: it falls with the weaker side instead of masking it.",
-	"methodology_legend_higher_better": "higher is better (F1 · score · integral)",
-	"methodology_legend_higher_worse": "higher is worse (FPR · FNR · FDR)",
-	"methodology_legend_glyphs": "▲ higher is better · ▼ lower is better · ○ no data",
 	"ranking_report_link_label": "Open {model}",
 	"details": "Details",
 	"details_description": "Metrics for benchmarks and models.",
@@ -199,13 +179,10 @@
 	"details_benchmark_search_placeholder": "Benchmark name...",
 	"details_model_search_placeholder": "Search by name...",
 	"details_model_count": "{shown} of {total} models",
+	"details_error_rate": "Error rate",
 	"details_matrix_summary": "Summary: {metric} across all tests - {models} models × {tests} tests",
 	"details_matrix_model_sort": "Model · Integral",
-	"details_matrix_good": "▲ good",
-	"details_matrix_bad": "▼ bad",
-	"details_matrix_no_data": "○ no data",
-	"details_matrix_direction_legend": "for FPR / FNR lower is better · for F1 / recall / precision / accuracy higher is better",
-	"details_matrix_sort_legend": "click a column to sort · click “Model” to sort by Integral",
+	"details_matrix_no_data": "no data",
 	"details_matrix_empty": "Nothing found. Reset search or enable groups.",
 	"details_dataset_matrix_unavailable_title": "Dataset matrix is temporarily unavailable",
 	"details_dataset_matrix_unavailable_description": "The current data includes dataset-level metrics for {covered} of {total} datasets. Group-level cards, rankings, and model profiles below use real data; the dataset matrix will be enabled when coverage is complete.",
@@ -219,15 +196,6 @@
 	"details_family_stat_models": "models evaluated",
 	"details_family_filters": "Benchmark family filters",
 	"details_family_language_all": "All languages",
-	"details_family_metric_label": "Card leader metric",
-	"details_family_metric_f1": "F1 (leader by F1)",
-	"details_family_metric_fpr": "FPR (fewest false positives)",
-	"details_family_metric_fnr": "FNR (fewest misses)",
-	"details_family_metric_hint": "Defines the leader and score shown on cards.",
-	"details_family_legend_good": "good (F1↑ / FPR·FNR↓)",
-	"details_family_legend_middle": "middle",
-	"details_family_legend_bad": "bad",
-	"details_family_legend_direction": "▲ higher is better · ▼ lower is better · ○ no data",
 	"details_family_empty": "Nothing found. Change filters or reset the search.",
 	"details_reset": "Reset",
 	"details_benchmark_sheet_what_tests": "What it tests",
@@ -236,11 +204,6 @@
 	"details_benchmark_sheet_rank": "Rank",
 	"details_benchmark_sheet_open": "Open benchmark",
 	"details_model_sheet_open": "Open model",
-	"details_model_legend_lead": "Score 0-1: integral score (higher is better)",
-	"details_model_legend_good": "≥ 0.85 - good",
-	"details_model_legend_middle": "0.60-0.85",
-	"details_model_legend_bad": "< 0.60 - weak",
-	"details_model_legend_hint": "F1 is higher-is-better · p50 is median latency, lower is better",
 	"details_model_sheet_overview": "Overview",
 	"details_model_sheet_groups": "Groups",
 	"details_model_sheet_weak": "Benchmarks",
@@ -260,12 +223,11 @@
 	"details_model_sheet_params": "Parameters",
 	"details_model_sheet_type": "Guardrail type",
 	"details_model_sheet_languages": "Languages",
-	"details_model_sheet_eval_leak": "Eval leak",
+	"details_model_sheet_status": "Status",
 	"details_model_sheet_run_date": "Run date",
 	"details_model_sheet_links": "Links",
 	"details_model_sheet_hf_card": "HF card",
 	"details_dataset_count": "{count} datasets",
-	"details_leader": "Leader",
 	"details_no_data": "no data",
 	"details_license_label": "license",
 	"details_groups_label": "groups",
@@ -276,11 +238,7 @@
 	"details_export": "Export",
 	"tools_meta_title": "Visualizations - HiveTrace Guard Leaderboard",
 	"tools_meta_description": "A visualization workspace for researchers: radar, FPR×FNR scatter, FNR heatmap, grouped bars, and latency Pareto.",
-	"tools_description": "The filters below control all charts in sync. Clicking a heatmap cell or scatter point opens the slice details; “Pin” locks a model for comparison.",
-	"tools_takeaway_integral": "Leader by integral:",
-	"tools_takeaway_p95": "fastest p95:",
-	"tools_takeaway_fnr": "fewest missed harmful items (FNR):",
-	"tools_takeaway_empty": "Select models to see the leader summary.",
+	"tools_description": "The filters below control all charts in sync; “Pin” locks a model for comparison.",
 	"tools_controls_aria": "Visualization control panel",
 	"tools_models_title": "Models",
 	"tools_groups_title": "Groups / benchmarks",
@@ -300,7 +258,7 @@
 	"tools_selected_summary": "{models} models · {groups} groups",
 	"tools_pinned_for_comparison": "Pinned for comparison",
 	"tools_unpin_model": "Unpin {model}",
-	"tools_no_pinned": "- nothing pinned. Press “Pin” on a model in the Models list on the left.",
+	"tools_no_pinned": "Nothing pinned, press Pin in the Models list.",
 	"tools_radar_title": "Group radar",
 	"tools_radar_hint": "scores of selected models by group · higher = better",
 	"tools_radar_aria": "Group radar for selected models",
@@ -308,13 +266,9 @@
 	"tools_scatter_hint": "x = FPR, y = FNR · lower = better · ★ ideal (0,0)",
 	"tools_scatter_aria": "FPR versus FNR scatter",
 	"tools_scatter_ideal": "★ Ideal (0,0) - zero errors",
-	"tools_scatter_click": "Click a point → prompts",
-	"tools_heatmap_hint": "groups × models · the number is always printed",
-	"tools_heatmap_scale": "1.0 teal → red",
-	"tools_heatmap_header": "{metric} · group ↓ / model →",
-	"tools_heatmap_note": "Color is duplicated by the number; “-” = no data.",
-	"tools_heatmap_sheet_prompts_title": "Prompts for this slice",
-	"tools_heatmap_sheet_prompts_description": "Selected metric {metric} = {value}. In the full version, this opens a viewer for individual prompts and model predictions on this dataset:",
+	"tools_heatmap_scale_min": "min {value}",
+	"tools_heatmap_scale_max": "max {value}",
+	"tools_heatmap_header": "{metric} · group / model",
 	"tools_grouped_bars_title": "Grouped bars",
 	"tools_grouped_bars_hint": "FNR / FPR / F1 by group · averages across selected models",
 	"tools_grouped_bars_aria": "Grouped FNR FPR F1 bars by group",
@@ -326,16 +280,16 @@
 	"tools_pareto_x_axis": "p95 latency, ms (← lower better)",
 	"tools_pareto_y_axis": "integral (higher better →)",
 	"tools_pareto_fpr_legend": "teal (few false alarms) → red (many)",
-	"tools_pareto_click_pin_hint": "Click a point to pin/unpin its label · hover for details · dashed line = Pareto frontier",
+	"tools_pareto_click_pin_hint": "hover for details · dashed line = Pareto frontier",
 	"tools_pareto_pin_model": "Pin {model} label",
 	"tools_pareto_unpin_model": "Unpin {model} label",
 	"tools_latency_title": "Latency performance",
-	"tools_latency_hint": "p50 / p95 / p99 (ms) · lower is better",
-	"tools_latency_note": "Single-request latency percentiles. Default sorting is by p95 ascending.",
+	"tools_latency_hint": "p50 / p95 / p99 (ms) · error rate",
+	"tools_latency_note": "Single-request latency percentiles.",
 	"tools_robustness_title": "Robustness real → robust",
 	"tools_robustness_hint": "degradation under augmentation - who breaks on harder attacks",
-	"tools_robustness_note": "Δ score = robust − real (negative = quality degradation). Δ FNR = robust − real (positive = more missed harmful content under augmentation, worse).",
-	"tools_footer": "HiveTrace Guardrail Leaderboard · Tools · real run data (April 2026). Metrics: FPR/FNR/latency - lower = better; F1/integral/score - higher = better.",
+	"tools_robustness_note": "Delta = robust - real. Delta score < 0: the model is not robust to obfuscations, > 0: robust. Delta FNR > 0: it misses harmful obfuscated messages more often (worse), < 0: it catches them more often (better). Delta FPR > 0: it blocks safe obfuscated messages more often (worse), < 0: it lets them pass more often (better).",
+	"tools_footer": "HiveTrace Guardrail Leaderboard",
 	"bucket_data_unavailable_title": "Data is temporarily unavailable",
 	"bucket_data_unavailable_description": "We could not load data for this page. Try refreshing the page again in a moment.",
 	"bucket_data_stale_title": "Showing saved data",

messages/ru.json

@@ -11,7 +11,7 @@
 	"ranking_badge": "Бенчмарк guardrails",
 	"ranking_updated": "обновлено {date}",
 	"ranking_title": "Рейтинг",
-	"ranking_description": "Рейтинг open-source guardrail-моделей. Выше Integral - лучше.",
+	"ranking_description": "Рейтинг open-source guardrail-моделей.",
 	"ranking_filters": "Фильтры рейтинга",
 	"ranking_table": "Таблица моделей",
 	"ranking_scope_select_label": "срез данных",
@@ -55,16 +55,18 @@
 	"ranking_rows": "{count} строк",
 	"ranking_column_model": "Модель",
 	"ranking_column_integral": "Integral",
-	"ranking_column_weakest_link": "Слабое звено",
+	"ranking_column_min_score": "Min score",
 	"ranking_column_fpr": "Overall FPR",
 	"ranking_column_fnr": "Overall FNR",
 	"ranking_column_p95": "p95 мс",
 	"ranking_column_size": "Размер",
+	"ranking_column_higher_better": "больше лучше",
+	"ranking_column_lower_better": "меньше лучше",
 	"ranking_column_report": "Отчёт",
 	"ranking_model_type_classifier": "классификатор",
 	"ranking_model_type_llm_judge": "LLM-судья",
-	"ranking_eval_clean": "clean",
 	"ranking_eval_leak": "eval-leak",
+	"ranking_eval_partial": "partial",
 	"ranking_open_details_label": "Открыть детали модели {model}",
 	"ranking_details_overview_tab": "Обзор",
 	"ranking_details_groups_tab": "По группам",
@@ -130,39 +132,22 @@
 	"methodology_title": "Методология",
 	"methodology_description": "Датасет → группа → интегральный балл. FPR и FNR остаются раздельными до последнего шага.",
 	"methodology_meaning_label": "Что это значит:",
-	"methodology_example_label": "Пример",
 	"methodology_aggregation_title": "Три уровня агрегации",
 	"methodology_dataset_score_tag": "балл датасета",
 	"methodology_dataset_score_comment": "гармоническое среднее (1−FPR, 1−FNR)",
 	"methodology_dataset_score_description": "Нормализуем ошибки в качество: 1−FPR - доля верно пропущенного безопасного, 1−FNR - доля верно пойманного вредного. Затем берём гармоническое среднее: слабая сторона роняет балл, высокая другая не спасает.",
-	"methodology_dataset_score_example": "<b>Qwen3Guard-Gen-8B на ToxicChat</b>: FPR 8.5%, FNR 6.3% → H(0.915, 0.937) = 0.926.",
 	"methodology_group_score_tag": "балл группы",
 	"methodology_group_score_comment": "гармоническое среднее баллов N датасетов",
 	"methodology_group_score_description": "Группа - семейство близких датасетов, например jailbreak-тесты. Снова используем гармонику: самый слабый датасет определяет балл группы, нельзя добрать очки за счёт лёгких сетов внутри той же темы.",
-	"methodology_group_score_example": "<b>Qwen3Guard-Gen-8B, группа XSTest</b>: FPR 12.8% / FNR 3.5% → S_group = 0.916.",
 	"methodology_integral_score_tag": "итоговый балл",
 	"methodology_integral_score_comment": "геометрическое среднее M групп (S_group > 0)",
 	"methodology_integral_log_comment": "логарифмическая форма (численная стабильность)",
 	"methodology_integral_score_description": "По группам берём геометрическое среднее. Оно сильнее наказывает дисбаланс: одна провальная группа тянет произведение вниз. Ровный профиль ценится выше звезды в одном месте и дыры в другом.",
-	"methodology_integral_score_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: геом. среднее по группам → S_integral = 0.918. Слабейшая группа 0.823 тянет итог вниз.",
-	"methodology_helper_metrics_title": "Вспомогательные метрики",
-	"methodology_fbeta_tag": "общая форма",
-	"methodology_fbeta_description": "Один рычаг β балансирует precision и recall: β > 1 смещает к полноте, β < 1 - к точности.",
-	"methodology_f2_tag": "β=2",
-	"methodology_f2_description": "Recall важнее в 4 раза. F2 сильнее штрафует пропуск вредного и подходит для сетов, где цена пропуска выше цены ложной тревоги.",
-	"methodology_f2_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: recall R=0.862 (1−FNR), специфичность 0.910 (1−FPR) - реальные. Precision (TP/(TP+FP)) нет в данных лидерборда, число этой метрики не показываем: специфичность ≠ precision.",
-	"methodology_f05_tag": "β=0.5",
-	"methodology_f05_description": "Precision важнее в 4 раза. F0.5 сильнее штрафует ложные срабатывания и подходит для тестов на over-refusal.",
-	"methodology_f05_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: recall R=0.862 (1−FNR), специфичность 0.910 (1−FPR) - реальные. Precision (TP/(TP+FP)) нет в данных лидерборда, число этой метрики не показываем: специфичность ≠ precision.",
-	"methodology_fdr_description": "Доля ложной тревоги среди блокировок: какова вероятность ошибки, если сработка произошла. Низкий FDR означает, что большинству помеченных объектов вердикт вынесен справедливо.",
-	"methodology_fdr_example": "<b>Qwen3Guard-Gen-8B (#1)</b>: overall FPR 9.0%, FNR 13.8% - реальные. FDR требует precision, которого нет в данных, поэтому числовой FDR не приводим.",
 	"methodology_combined_dataset_title": "Комбинированный датасет",
 	"methodology_harm_split_tag": "harm split",
 	"methodology_harm_split_description": "Доля вредоносных запросов, ошибочно помеченных безопасными. TP - пойманные атаки, FN - пропущенные.",
 	"methodology_safe_split_tag": "safe split",
 	"methodology_safe_split_description": "Доля безопасных запросов, ошибочно помеченных вредными. TN - верно пропущенные безопасные, FP - ложные срабатывания.",
-	"methodology_micro_f1_tag": "combined",
-	"methodology_micro_f1_description": "Считается через глобальные TP, FP и FN по всему набору. При равных весах классов в бинарной задаче micro-F1 совпадает с accuracy.",
 	"methodology_combined_dataset_note": "Если в списке есть общий сет с пометкой combined, для него метрики берутся напрямую с полных сплитов; для остальных - стандартно, через арифметическое среднее их индивидуальных метрик.",
 	"methodology_scatter_title": "Scatter: FPR vs FNR",
 	"methodology_scatter_x_axis": "Ось X",
@@ -176,11 +161,6 @@
 	"methodology_single_class_description": "F1 считается для всех бенчмарков <b>по умолчанию</b>. Если в датасете присутствуют метки только одного класса: для all_unsafe рассчитываются <b>FNR, recall</b>; для all_safe соответственно <b>FPR, TNR</b>.",
 	"methodology_caveat_title": "Оговорка",
 	"methodology_caveat_description": "Метрики - точечные оценки. Доверительные интервалы и размеры выборок (n) по тестам в текущем датасете не публикуются, поэтому небольшие различия в рейтинге могут быть статистически незначимыми.",
-	"methodology_golden_rule_title": "Правило: никогда не сливайте FPR и FNR в один балл",
-	"methodology_golden_rule_description": "FPR (ложная блокировка безопасного) и FNR (пропуск вредного) - это <b>две разные ошибки с разной ценой</b>. Усреднение их в одно число прячет провал: модель с FPR 0% и FNR 40% выглядит «средней», хотя пропускает почти половину атак. Поэтому обе ошибки держатся раздельно до уровня S_ds, а агрегация идёт <b>гармоническим/геометрическим</b> средним - оно проваливается вместе со слабой стороной, а не маскирует её.",
-	"methodology_legend_higher_better": "выше - лучше (F1 · score · integral)",
-	"methodology_legend_higher_worse": "выше - хуже (FPR · FNR · FDR)",
-	"methodology_legend_glyphs": "▲ выше лучше · ▼ ниже лучше · ○ нет данных",
 	"ranking_report_link_label": "Открыть {model}",
 	"details": "Детали",
 	"details_description": "Метрики по бенчмаркам и моделям.",
@@ -199,13 +179,10 @@
 	"details_benchmark_search_placeholder": "Название бенчмарка...",
 	"details_model_search_placeholder": "Поиск по названию...",
 	"details_model_count": "{shown} из {total} моделей",
+	"details_error_rate": "Error rate",
 	"details_matrix_summary": "Сводная: {metric} по всем тестам - {models} моделей × {tests} тестов",
 	"details_matrix_model_sort": "Модель · Integral",
-	"details_matrix_good": "▲ хорошо",
-	"details_matrix_bad": "▼ плохо",
-	"details_matrix_no_data": "○ нет данных",
-	"details_matrix_direction_legend": "для FPR / FNR ниже = лучше · для F1 / recall / precision / accuracy выше = лучше",
-	"details_matrix_sort_legend": "клик по столбцу - сортировка · клик по «Модель» - по Integral",
+	"details_matrix_no_data": "нет данных",
 	"details_matrix_empty": "Ничего не найдено. Сбросьте поиск или включите группы.",
 	"details_dataset_matrix_unavailable_title": "Матрица по датасетам временно недоступна",
 	"details_dataset_matrix_unavailable_description": "Сейчас в данных есть dataset-level метрики для {covered} из {total} датасетов. Карточки групп, рейтинги и профили моделей ниже используют реальные данные; матрица будет включена после полного покрытия.",
@@ -219,15 +196,6 @@
 	"details_family_stat_models": "моделей в оценке",
 	"details_family_filters": "Фильтры семейств бенчмарков",
 	"details_family_language_all": "Все языки",
-	"details_family_metric_label": "Лидер карточки по метрике",
-	"details_family_metric_f1": "F1 (лидер по F1)",
-	"details_family_metric_fpr": "FPR (минимум ложных)",
-	"details_family_metric_fnr": "FNR (минимум пропусков)",
-	"details_family_metric_hint": "Определяет лидера и оценку на карточках.",
-	"details_family_legend_good": "хорошо (F1↑ / FPR·FNR↓)",
-	"details_family_legend_middle": "средне",
-	"details_family_legend_bad": "плохо",
-	"details_family_legend_direction": "▲ выше - лучше · ▼ ниже - лучше · ○ нет данных",
 	"details_family_empty": "Ничего не найдено. Измените фильтры или сбросьте поиск.",
 	"details_reset": "Сбросить",
 	"details_benchmark_sheet_what_tests": "Что проверяет",
@@ -236,11 +204,6 @@
 	"details_benchmark_sheet_rank": "Место",
 	"details_benchmark_sheet_open": "Открыть бенчмарк",
 	"details_model_sheet_open": "Открыть модель",
-	"details_model_legend_lead": "Оценка 0-1: интегральный балл (выше = лучше)",
-	"details_model_legend_good": "≥ 0.85 - хорошо",
-	"details_model_legend_middle": "0.60-0.85",
-	"details_model_legend_bad": "< 0.60 - слабо",
-	"details_model_legend_hint": "F1 - выше лучше · p50 - медианная задержка, ниже лучше",
 	"details_model_sheet_overview": "Обзор",
 	"details_model_sheet_groups": "По группам",
 	"details_model_sheet_weak": "Бенчмарки",
@@ -260,12 +223,11 @@
 	"details_model_sheet_params": "Параметры",
 	"details_model_sheet_type": "Тип guardrail",
 	"details_model_sheet_languages": "Языки",
-	"details_model_sheet_eval_leak": "Eval-leak",
+	"details_model_sheet_status": "Статус",
 	"details_model_sheet_run_date": "Дата прогона",
 	"details_model_sheet_links": "Ссылки",
 	"details_model_sheet_hf_card": "Карточка HF",
 	"details_dataset_count": "{count} датасета",
-	"details_leader": "Лидер",
 	"details_no_data": "нет данных",
 	"details_license_label": "license",
 	"details_groups_label": "groups",
@@ -276,11 +238,7 @@
 	"details_export": "Экспорт",
 	"tools_meta_title": "Визуализации - HiveTrace Guard Leaderboard",
 	"tools_meta_description": "Площадка визуализаций для исследователей: радар, scatter FPR×FNR, heatmap FNR, сгруппированные бары и Парето задержки.",
-	"tools_description": "Фильтры ниже управляют всеми графиками синхронно. Клик по ячейке heatmap или точке scatter открывает детали среза; «Pin» закрепляет модель.",
-	"tools_takeaway_integral": "Лидер по integral:",
-	"tools_takeaway_p95": "быстрейший p95:",
-	"tools_takeaway_fnr": "меньше пропусков вреда (FNR):",
-	"tools_takeaway_empty": "Выберите модели, чтобы увидеть сводку лидеров.",
+	"tools_description": "Фильтры ниже управляют всеми графиками синхронно; «Pin» закрепляет модель.",
 	"tools_controls_aria": "Панель управления визуализациями",
 	"tools_models_title": "Модели",
 	"tools_groups_title": "Группы / б��нчмарки",
@@ -300,7 +258,7 @@
 	"tools_selected_summary": "{models} моделей · {groups} групп",
 	"tools_pinned_for_comparison": "Закреплено для сравнения",
 	"tools_unpin_model": "Открепить {model}",
-	"tools_no_pinned": "- ничего не закреплено. Нажмите «Pin» у модели в списке «Модели» слева.",
+	"tools_no_pinned": "Ничего не закреплено, нажмите Pin в списке Модели.",
 	"tools_radar_title": "Радар групп",
 	"tools_radar_hint": "баллы выбранных моделей по группам · выше = лучше",
 	"tools_radar_aria": "Радар групп по выбранным моделям",
@@ -308,13 +266,9 @@
 	"tools_scatter_hint": "x = FPR, y = FNR · ниже = лучше · ★ идеал (0,0)",
 	"tools_scatter_aria": "Scatter FPR против FNR",
 	"tools_scatter_ideal": "★ Идеал (0,0) - нулевые ошибки",
-	"tools_scatter_click": "Клик по точке → промпты",
-	"tools_heatmap_hint": "группы × модели · число всегда напечатано",
-	"tools_heatmap_scale": "1.0 teal → red",
-	"tools_heatmap_header": "{metric} · группа ↓ / модель →",
-	"tools_heatmap_note": "Цвет дублируется числом; «-» = нет данных.",
-	"tools_heatmap_sheet_prompts_title": "Промпты этого среза",
-	"tools_heatmap_sheet_prompts_description": "Выбрана метрика {metric} = {value}. В полной версии отсюда открывается просмотрщик отдельных промптов и предсказаний модели на этом наборе данных:",
+	"tools_heatmap_scale_min": "min {value}",
+	"tools_heatmap_scale_max": "max {value}",
+	"tools_heatmap_header": "{metric} · группа / модель",
 	"tools_grouped_bars_title": "Сгруппированные бары",
 	"tools_grouped_bars_hint": "FNR / FPR / F1 по группам · средние по выбранным моделям",
 	"tools_grouped_bars_aria": "Сгруппированные бары FNR FPR F1 по группам",
@@ -326,16 +280,16 @@
 	"tools_pareto_x_axis": "p95 latency, ms (← ниже лучше)",
 	"tools_pareto_y_axis": "integral (выше лучше →)",
 	"tools_pareto_fpr_legend": "teal (мало ложных) → red (много)",
-	"tools_pareto_click_pin_hint": "Клик по точке — закрепить/снять подпись · hover — детали · пунктир = Pareto frontier",
+	"tools_pareto_click_pin_hint": "hover — детали · пунктир = Pareto frontier",
 	"tools_pareto_pin_model": "Закрепить подпись {model}",
 	"tools_pareto_unpin_model": "Снять подпись {model}",
 	"tools_latency_title": "Производительность latency",
-	"tools_latency_hint": "p50 / p95 / p99 (ms) · ниже лучше",
-	"tools_latency_note": "Перцентили задержки одного запроса. По умолчанию сортировка по p95 (возрастание).",
+	"tools_latency_hint": "p50 / p95 / p99 (ms) · error rate",
+	"tools_latency_note": "Перцентили задержки одного запроса.",
 	"tools_robustness_title": "Устойчивость real → robust",
 	"tools_robustness_hint": "падение под аугментацией - кто ломается на усложнённых атаках",
-	"tools_robustness_note": "Δ score = robust − real (отрицательный = деградация качества). Δ FNR = robust − real (положительный = больше пропусков вреда под аугментацией, хуже).",
-	"tools_footer": "HiveTrace Guardrail Leaderboard · Инструменты · реальные данные прогона (апрель 2026). Метрики: FPR/FNR/latency - ниже = лучше; F1/integral/score - выше = лучше.",
+	"tools_robustness_note": "Delta = robust - real. Delta score < 0: модель неустойчива к обфускациям, > 0: устойчива. Delta FNR > 0: чаще пропускает вредоносные обфусцированные сообщения (хуже), < 0: чаще их ловит (лучше). Delta FPR > 0: чаще банит безопасные обфусцированные сообщения (хуже), < 0: чаще их пропускает (лучше).",
+	"tools_footer": "HiveTrace Guardrail Leaderboard",
 	"bucket_data_unavailable_title": "Данные временно недоступны",
 	"bucket_data_unavailable_description": "Не удалось загрузить данные для этой страницы. Попробуйте обновить страницу чуть позже.",
 	"bucket_data_stale_title": "Показаны сохраненные данные",

src/lib/data/details.ts

@@ -3,7 +3,6 @@ import type { DetailsDataset } from '$lib/types/details-data';
 import type {
 	BenchmarkCard,
 	BenchmarkFamilyLanguage,
-	BenchmarkLeaderMetric,
 	BenchmarkRankingRow,
 	DetailMetric,
 	DetailsLanguageFilter,
@@ -189,29 +188,9 @@ export function benchmarkRows(
 		});
 }
 
-export function familyMetricValue(group: GroupScore | undefined, metric: BenchmarkLeaderMetric) {
-	if (!group) return null;
-	if (metric === 'f1') return group.f1;
-	return group[metric];
-}
-
-function leaderForGroupMetric(
-	group: string,
-	metric: BenchmarkLeaderMetric,
-	models: readonly GuardModel[]
-) {
-	const ranked = models
-		.map((model) => ({ model, score: familyMetricValue(model.groups[group], metric) }))
-		.filter((item): item is { model: GuardModel; score: number } => isNumber(item.score))
-		.sort((a, b) => (metric === 'fpr' || metric === 'fnr' ? a.score - b.score : b.score - a.score));
-
-	return ranked[0] ?? null;
-}
-
 export function buildBenchmarkCards(
 	search: string,
 	language: BenchmarkFamilyLanguage,
-	metric: BenchmarkLeaderMetric,
 	dataset: DetailsDataset,
 	locale: SupportedLocale
 ) {
@@ -232,15 +211,10 @@ export function buildBenchmarkCards(
 				.toLowerCase()
 				.includes(normalizedSearch);
 		})
-		.map<BenchmarkCard>((group) => {
-			const leader = leaderForGroupMetric(group, metric, dataset.models);
-			return {
-				name: group,
-				meta: dataset.benchmarkMeta[group],
-				leader,
-				score: leader?.score ?? null
-			};
-		});
+		.map<BenchmarkCard>((group) => ({
+			name: group,
+			meta: dataset.benchmarkMeta[group]
+		}));
 }
 
 export function matchesLanguage(model: GuardModel, language: DetailsLanguageFilter) {
@@ -379,38 +353,3 @@ export function radarGridPolygons(groups: readonly ModelGroupRow[]) {
 			.join(' ')
 	);
 }
-
-export function detailMetricIsGood(value: number | null, metric: DetailMetric) {
-	if (!isNumber(value)) return false;
-	if (metric === 'fpr' || metric === 'fnr') return value <= 0.08;
-	return value >= 0.85;
-}
-
-export function detailMetricIsBad(value: number | null, metric: DetailMetric) {
-	if (!isNumber(value)) return false;
-	if (metric === 'fpr' || metric === 'fnr') return value > 0.2;
-	return value < 0.6;
-}
-
-export function detailMetricGlyph(value: number | null, metric: DetailMetric) {
-	if (!isNumber(value)) return '○';
-	if (detailMetricIsGood(value, metric)) return '▲';
-	if (detailMetricIsBad(value, metric)) return '▼';
-	return '-';
-}
-
-export function familyMetricIsGood(value: number | null, metric: BenchmarkLeaderMetric) {
-	return metric === 'f1' ? isNumber(value) && value >= 0.85 : isNumber(value) && value <= 0.08;
-}
-
-export function familyMetricIsBad(value: number | null, metric: BenchmarkLeaderMetric) {
-	return metric === 'f1' ? isNumber(value) && value < 0.6 : isNumber(value) && value > 0.2;
-}
-
-export function familyMetricGlyph(value: number | null, metric: BenchmarkLeaderMetric) {
-	if (!isNumber(value)) return '○';
-	if (metric === 'fpr' || metric === 'fnr') return '▼';
-	if (value >= 0.85) return '▲';
-	if (value < 0.6) return '▼';
-	return '-';
-}

src/lib/data/ranking.ts

@@ -3,7 +3,6 @@ import type { RankingDataset } from '$lib/types/ranking-data';
 import type { BenchmarkScope, GuardModel } from '$lib/types/leaderboard';
 import type {
 	BenchmarkDetail,
-	ExtremaKey,
 	LanguageFilter,
 	LeaderboardRow,
 	SortDirection,
@@ -120,16 +119,6 @@ export function compareRankingRows(
 	return (Number(first) - Number(second)) * multiplier;
 }
 
-export function rankingExtrema(rows: readonly LeaderboardRow[], key: ExtremaKey, polarity: 1 | -1) {
-	const values = rows.map((row) => row[key]).filter(isFiniteNumber);
-	if (!values.length) return { best: null, worst: null };
-
-	const min = Math.min(...values);
-	const max = Math.max(...values);
-
-	return polarity >= 0 ? { best: max, worst: min } : { best: min, worst: max };
-}
-
 export function benchmarkDetails(model: GuardModel, groups: readonly string[]): BenchmarkDetail[] {
 	return groups
 		.map((group) => ({

src/lib/data/tools.ts

@@ -1,15 +1,9 @@
 import { isNumber } from '$lib/helpers/leaderboard';
 import type { GroupScore, GuardModel } from '$lib/types/leaderboard';
 import type {
-	ToolsHeatmapMetric,
-	ToolsHeatmapMetricItem,
-	ToolsHeatmapSelection,
 	ToolsLanguageFilter,
-	ToolsMetricTone,
-	ToolsPoint,
 	ToolsRadarScale,
 	ToolsSortDirection,
-	ToolsTopModelKey,
 	ToolsVizMetric
 } from '$lib/types/tools';
 
@@ -77,95 +71,22 @@ export function toolsMetricValue(cell: GroupScore | null, key: ToolsVizMetric) {
 	return cell[key];
 }
 
-export function toolsHeatmapMetricValue(
-	metrics: ToolsHeatmapSelection['metrics'],
-	key: ToolsHeatmapMetric
-) {
-	return metrics[key];
-}
-
-export function toolsHeatmapMetricItems(
-	selection: ToolsHeatmapSelection
-): ToolsHeatmapMetricItem[] {
-	return [
-		{ label: 'FNR', key: 'fnr', value: selection.metrics.fnr },
-		{ label: 'FPR', key: 'fpr', value: selection.metrics.fpr },
-		{ label: 'F1', key: 'f1', value: selection.metrics.f1 },
-		{ label: 'SCORE', key: 'score', value: selection.metrics.score }
-	];
-}
-
 export function toolsRadarValue(cell: GroupScore | null, key: ToolsRadarScale) {
 	if (!cell) return null;
 	return cell[key];
 }
 
-export function toolsBestModel(
-	list: readonly GuardModel[],
-	key: ToolsTopModelKey,
-	direction: ToolsSortDirection
+export function toolsLatencySortValue(
+	model: GuardModel,
+	key: 'model' | 'p50' | 'p95' | 'p99' | 'errorRate'
 ) {
-	const pool = list.filter((model) => isNumber(key === 'fnr' ? model.fnr : model[key]));
-	if (!pool.length) return null;
-
-	return pool.reduce((best, model) => {
-		const bestValue = key === 'fnr' ? best.fnr : best[key];
-		const modelValue = key === 'fnr' ? model.fnr : model[key];
-		return direction === 'desc'
-			? modelValue > bestValue
-				? model
-				: best
-			: modelValue < bestValue
-				? model
-				: best;
-	});
-}
-
-export function toolsValueTone(
-	value: number | null | undefined,
-	higherBetter = true
-): ToolsMetricTone {
-	if (!isNumber(value)) return 'neutral';
-	if (higherBetter) {
-		if (value >= 0.85) return 'good';
-		if (value <= 0.6) return 'bad';
-		return 'neutral';
-	}
-	if (value <= 0.05) return 'good';
-	if (value >= 0.15) return 'bad';
-	return 'neutral';
-}
-
-export function toolsHeatmapMetricTone(
-	value: number | null | undefined,
-	key: ToolsHeatmapMetric
-): ToolsMetricTone {
-	if (!isNumber(value)) return 'neutral';
-	if (key === 'fnr' || key === 'fpr') {
-		if (value <= 0.08) return 'good';
-		if (value >= 0.2) return 'bad';
-		return 'neutral';
-	}
-	if (value >= 0.85) return 'good';
-	if (value < 0.6) return 'bad';
-	return 'neutral';
-}
-
-export function toolsHeatmapMetricArrow(value: number | null | undefined, key: ToolsHeatmapMetric) {
-	const tone = toolsHeatmapMetricTone(value, key);
-	if (tone === 'good') return '▲';
-	if (tone === 'bad') return '▼';
-	return '−';
-}
-
-export function toolsLatencySortValue(model: GuardModel, key: 'model' | 'p50' | 'p95' | 'p99') {
 	if (key === 'model') return model.short.toLowerCase();
 	return model[key];
 }
 
 export function toolsSortLatencyRows(
 	rows: readonly GuardModel[],
-	key: 'model' | 'p50' | 'p95' | 'p99',
+	key: 'model' | 'p50' | 'p95' | 'p99' | 'errorRate',
 	direction: ToolsSortDirection
 ) {
 	return [...rows].sort((a, b) => {
@@ -174,14 +95,24 @@ export function toolsSortLatencyRows(
 		const result =
 			typeof aValue === 'string' && typeof bValue === 'string'
 				? aValue.localeCompare(bValue)
-				: Number(aValue) - Number(bValue);
+				: Number(aValue ?? Number.POSITIVE_INFINITY) - Number(bValue ?? Number.POSITIVE_INFINITY);
 		return direction === 'asc' ? result : -result;
 	});
 }
 
-export function toolsHeatColor(value: number | null | undefined) {
+export function toolsHeatColor(
+	value: number | null | undefined,
+	min = 0,
+	max = 1,
+	curve: 'linear' | 'sqrt' = 'linear'
+) {
 	if (!isNumber(value)) return 'transparent';
-	const t = Math.max(0, Math.min(1, value));
+	const range = max - min;
+	const normalized = range > 0 ? (value - min) / range : 0;
+	const t =
+		curve === 'sqrt'
+			? Math.sqrt(Math.max(0, Math.min(1, normalized)))
+			: Math.max(0, Math.min(1, normalized));
 	if (t < 0.5) {
 		const p = t / 0.5;
 		return `rgb(${Math.round(13 + 242 * p)}, ${Math.round(206 - 36 * p)}, ${Math.round(151 - 111 * p)})`;
@@ -275,21 +206,6 @@ export function toolsParetoPoint(model: GuardModel, xMax: number, y0: number) {
 	return { x: toolsParetoX(model.p95, xMax), y: toolsParetoY(model.integral, y0) };
 }
 
-export function toolsParetoFront(modelsToPlot: readonly GuardModel[], xMax: number, y0: number) {
-	let best = -Infinity;
-	const points: ToolsPoint[] = [];
-
-	for (const model of [...modelsToPlot].sort((a, b) => a.p95 - b.p95)) {
-		if (model.integral > best) {
-			best = model.integral;
-			const point = toolsParetoPoint(model, xMax, y0);
-			points.push({ x: point.x, y: point.y });
-		}
-	}
-
-	return points.map((point) => `${point.x.toFixed(1)},${point.y.toFixed(1)}`).join(' ');
-}
-
 export function toolsGroupMean(
 	group: string,
 	key: ToolsVizMetric,

src/lib/helpers/metrics.ts

@@ -1,45 +0,0 @@
-import type { MetricName, MetricState, MetricTrendPreference } from '$lib/types/metrics';
-
-function isNumber(value: unknown): value is number {
-	return typeof value === 'number' && Number.isFinite(value);
-}
-
-export function metricState(metric: MetricName, value: number | null): MetricState {
-	if (!isNumber(value)) return 'na';
-
-	if (metric === 'fpr' || metric === 'fnr' || metric === 'latency') {
-		if (value <= 0.05) return 'good';
-		if (value >= 0.15) return 'bad';
-		return 'neutral';
-	}
-
-	if (value >= 0.85) return 'good';
-	if (value <= 0.6) return 'bad';
-	return 'neutral';
-}
-
-export function scoreState(value: number | null): MetricState {
-	if (!isNumber(value)) return 'na';
-	if (value >= 0.85) return 'good';
-	if (value >= 0.6) return 'neutral';
-	return 'bad';
-}
-
-export function metricGlyph(metric: MetricName, value: number | null) {
-	const state = metricState(metric, value);
-	if (state === 'good') return '▲';
-	if (state === 'bad') return '▼';
-	if (state === 'na') return '○';
-	return '';
-}
-
-export function valueTrend(value: number | null, positive: MetricTrendPreference) {
-	if (!isNumber(value)) return '';
-
-	const isGood = positive === 'high' ? value >= 0.85 : value <= 0.08;
-	const isBad = positive === 'high' ? value < 0.6 : value > 0.2;
-
-	if (isGood) return '▲';
-	if (isBad) return '▼';
-	return '';
-}

src/lib/server/hf-bucket/adapt-details.ts

@@ -59,7 +59,8 @@ export function adaptBucketDetails(snapshot: BucketDetailsSnapshot): DetailsData
 			recall: cell.recall,
 			precision: cell.precision,
 			accuracy: cell.accuracy,
-			sampleCount: cell.sample_count
+			sampleCount: cell.sample_count,
+			metricsEvaluatedSamples: cell.metrics_evaluated_samples
 		};
 	}
 

src/lib/server/hf-bucket/adapt-ranking.ts

@@ -1,10 +1,11 @@
 import type { RankingDataset } from '$lib/types/ranking-data';
 import type { GroupScore, GuardModel, GuardrailType } from '$lib/types/leaderboard';
 import type {
+	BucketDetailsSnapshot,
 	BucketGroup,
 	BucketModel,
-	BucketRankingSnapshot,
-	BucketLeaderboardRow
+	BucketLeaderboardRow,
+	BucketRankingSnapshot
 } from './types';
 
 const ROBUST_GROUP_IDS = new Set([
@@ -85,7 +86,8 @@ function completeRankingRow(row: BucketLeaderboardRow): row is CompleteRankingRo
 function adaptModel(
 	model: BucketModel,
 	row: BucketLeaderboardRow,
-	groupLabelById: Map<string, string>
+	groupLabelById: Map<string, string>,
+	partialModelIds: ReadonlySet<string>
 ): GuardModel | null {
 	if (!completeRankingRow(row)) {
 		console.warn(`[hf-bucket] Ranking skips incomplete model "${model.model_id}".`);
@@ -126,12 +128,30 @@ function adaptModel(
 		p50: row.latency_ms.p50,
 		p95: row.latency_ms.p95,
 		p99: row.latency_ms.p99,
+		errorRate: row.error_rate,
 		leak: model.eval_leak ?? false,
+		partial: partialModelIds.has(model.model_id),
 		hfUrl: model.hf_url,
 		groups
 	};
 }
 
+function partialModelIds(snapshot: BucketRankingSnapshot | BucketDetailsSnapshot) {
+	if (!('detailsMatrix' in snapshot)) return new Set<string>();
+
+	const ids = new Set<string>();
+	for (const cell of snapshot.detailsMatrix.dataset_cells) {
+		if (
+			cell.metrics_evaluated_samples !== null &&
+			cell.sample_count !== null &&
+			cell.metrics_evaluated_samples < cell.sample_count
+		) {
+			ids.add(cell.model_id);
+		}
+	}
+	return ids;
+}
+
 function benchmarkMeta(groups: BucketGroup[], snapshot: BucketRankingSnapshot) {
 	const datasetsById = new Map(
 		snapshot.catalog.datasets.map((dataset) => [dataset.dataset_id, dataset])
@@ -153,16 +173,21 @@ function benchmarkMeta(groups: BucketGroup[], snapshot: BucketRankingSnapshot) {
 	);
 }
 
-export function adaptBucketRanking(snapshot: BucketRankingSnapshot): RankingDataset {
+export function adaptBucketRanking(
+	snapshot: BucketRankingSnapshot | BucketDetailsSnapshot
+): RankingDataset {
 	const groupLabelById = new Map(
 		snapshot.catalog.groups.map((group) => [group.group_id, group.label.en])
 	);
 	const catalogById = chooseLatest(snapshot.catalog.models, 'catalog.models');
 	const rowsById = chooseLatest(snapshot.leaderboard.rows, 'leaderboard.rows');
+	const partialIds = partialModelIds(snapshot);
 	const models = [...catalogById.values()]
 		.map((model) => {
 			const row = rowsById.get(model.model_id);
-			return row ? { model: adaptModel(model, row, groupLabelById), rank: row.rank } : null;
+			return row
+				? { model: adaptModel(model, row, groupLabelById, partialIds), rank: row.rank }
+				: null;
 		})
 		.filter(
 			(item): item is { model: GuardModel; rank: number } => item !== null && item.model !== null

src/lib/server/hf-bucket/adapt-tools.ts

@@ -146,6 +146,7 @@ export function adaptBucketTools(snapshot: BucketVisualizationSnapshot): ToolsDa
 					modelKey: modelKeyForId(row.model_id),
 					rank: row.rank,
 					latencyMs: latencyMs(row.latency_ms),
+					errorRate: row.error_rate,
 					integral: row.integral
 				}))
 			},

src/lib/server/hf-bucket/details-payload.ts

@@ -37,6 +37,19 @@ function nullableNumber(value: unknown, field: string) {
 	return value;
 }
 
+function nullableOptionalNumber(value: unknown, field: string) {
+	if (value === undefined) return null;
+	return nullableNumber(value, field);
+}
+
+function nullableNonNegativeInteger(value: unknown, field: string) {
+	const parsed = nullableOptionalNumber(value, field);
+	if (parsed !== null && (!Number.isInteger(parsed) || parsed < 0)) {
+		throw new Error(`Field "${field}" must be a non-negative integer or null.`);
+	}
+	return parsed;
+}
+
 function parseMetrics(item: Record<string, unknown>, field: string) {
 	return {
 		score: nullableNumber(item.score, `${field}.score`),
@@ -89,17 +102,19 @@ function parseGroupCell(value: unknown, index: number): BucketGroupCell {
 function parseDatasetCell(value: unknown, index: number): BucketDatasetCell {
 	const field = `details_matrix.dataset_cells[${index}]`;
 	const item = record(value, field);
-	const sampleCount = nullableNumber(item.sample_count, `${field}.sample_count`);
-	if (sampleCount !== null && (!Number.isInteger(sampleCount) || sampleCount < 0)) {
-		throw new Error(`Field "${field}.sample_count" must be a non-negative integer or null.`);
-	}
+	const sampleCount = nullableNonNegativeInteger(item.sample_count, `${field}.sample_count`);
+	const metricsEvaluatedSamples = nullableNonNegativeInteger(
+		item.metrics_evaluated_samples,
+		`${field}.metrics_evaluated_samples`
+	);
 
 	return {
 		model_id: string(item.model_id, `${field}.model_id`),
 		group_id: string(item.group_id, `${field}.group_id`),
 		dataset_id: string(item.dataset_id, `${field}.dataset_id`),
 		...parseMetrics(item, field),
-		sample_count: sampleCount
+		sample_count: sampleCount,
+		metrics_evaluated_samples: metricsEvaluatedSamples
 	};
 }
 

src/lib/server/hf-bucket/ranking-cache.ts

@@ -2,6 +2,7 @@ import { env } from '$env/dynamic/private';
 import type { BucketDataState } from '$lib/types/bucket-state';
 import type { RankingDataset } from '$lib/types/ranking-data';
 import { adaptBucketRanking } from './adapt-ranking';
+import { fetchBucketDetailsSnapshot } from './details-snapshot';
 import { emptyRankingDataset } from './fallback-data';
 import { toPublicBucketError } from './public-error';
 import { fetchBucketRankingSnapshot } from './ranking-snapshot';
@@ -23,7 +24,7 @@ async function refreshRanking() {
 		return cached.value;
 	}
 
-	const value = adaptBucketRanking(snapshot);
+	const value = adaptBucketRanking(await fetchBucketDetailsSnapshot(snapshot));
 	cached = {
 		value,
 		snapshotId: snapshot.manifest.snapshot_id,

src/lib/server/hf-bucket/ranking-payload.ts

@@ -198,6 +198,7 @@ function parseLeaderboardRow(value: unknown, index: number): BucketLeaderboardRo
 		overall_fpr: nullableNumber(item.overall_fpr, `${field}.overall_fpr`),
 		overall_fnr: nullableNumber(item.overall_fnr, `${field}.overall_fnr`),
 		overall_f1: nullableNumber(item.overall_f1, `${field}.overall_f1`),
+		error_rate: nullableNumber(item.error_rate ?? null, `${field}.error_rate`),
 		latency_ms: {
 			p50: nullableNumber(latency.p50, `${field}.latency_ms.p50`),
 			p95: nullableNumber(latency.p95, `${field}.latency_ms.p95`),

src/lib/server/hf-bucket/types.ts

@@ -25,7 +25,16 @@ export type BucketManifestHashes = Record<BucketManifestFileKey, string> & Recor
 export type NullableMetricValue = number | null;
 export type BucketMetricDirection = 'higher' | 'lower';
 export type BucketMetricKey =
-	'integral' | 'score' | 'f1' | 'recall' | 'precision' | 'accuracy' | 'fpr' | 'fnr' | 'latency_ms';
+	| 'integral'
+	| 'score'
+	| 'f1'
+	| 'recall'
+	| 'precision'
+	| 'accuracy'
+	| 'fpr'
+	| 'fnr'
+	| 'latency_ms'
+	| 'error_rate';
 
 export type BucketLatencyMs = {
 	p50: NullableMetricValue;
@@ -119,6 +128,7 @@ export type BucketLeaderboardRow = {
 	overall_fpr: NullableMetricValue;
 	overall_fnr: NullableMetricValue;
 	overall_f1: NullableMetricValue;
+	error_rate: NullableMetricValue;
 	latency_ms: BucketLatencyMs;
 	groups: Record<string, BucketGroupMetrics>;
 };
@@ -149,6 +159,7 @@ export type BucketGroupCell = {
 export type BucketDatasetCell = BucketGroupCell & {
 	dataset_id: string;
 	sample_count: NullableMetricValue;
+	metrics_evaluated_samples: NullableMetricValue;
 };
 
 export type BucketDetailsMatrix = {
@@ -253,6 +264,7 @@ export type BucketPerformanceRow = {
 	model_id: string;
 	rank: number;
 	latency_ms: BucketLatencyMs;
+	error_rate: NullableMetricValue;
 	integral: NullableMetricValue;
 };
 

src/lib/server/hf-bucket/visualizations-payload.ts

@@ -211,6 +211,7 @@ function parsePerformanceRow(value: unknown, index: number): BucketPerformanceRo
 		model_id: string(item.model_id, `${field}.model_id`),
 		rank: positiveInteger(item.rank, `${field}.rank`),
 		latency_ms: latency(item.latency_ms, `${field}.latency_ms`),
+		error_rate: nullableNumber(item.error_rate ?? null, `${field}.error_rate`),
 		integral: nullableNumber(item.integral, `${field}.integral`)
 	};
 }

src/lib/types/details-data.ts

@@ -19,6 +19,7 @@ export type DetailsDatasetMatrixCell = {
 	precision: number | null;
 	accuracy: number | null;
 	sampleCount: number | null;
+	metricsEvaluatedSamples: number | null;
 };
 
 export type DetailsDataset = RankingDataset & {

src/lib/types/details.ts

@@ -12,8 +12,6 @@ export type DetailsLanguageFilter = 'any' | 'en' | 'ru';
 
 export type BenchmarkFamilyLanguage = 'all' | 'en' | 'ru' | 'uk' | 'be' | 'uz';
 
-export type BenchmarkLeaderMetric = 'f1' | 'fpr' | 'fnr';
-
 export type ModelTypeFilter = 'all' | GuardModel['type'];
 
 export type SizeFilter = 'all' | 'small' | 'medium' | 'large';
@@ -55,8 +53,6 @@ export type BenchmarkRankingRow = {
 export type BenchmarkCard = {
 	name: string;
 	meta: RankingBenchmarkMeta;
-	leader: { model: GuardModel; score: number } | null;
-	score: number | null;
 };
 
 export type ModelGroupRow = {

src/lib/types/leaderboard.ts

@@ -33,7 +33,9 @@ export type GuardModel = {
 	p50: number;
 	p95: number;
 	p99: number;
+	errorRate: number | null;
 	leak: boolean;
+	partial: boolean;
 	hfUrl?: string | null;
 	groups: Record<string, GroupScore>;
 };

src/lib/types/metrics.ts

@@ -1,5 +0,0 @@
-export type MetricName = 'score' | 'fpr' | 'fnr' | 'latency';
-
-export type MetricState = 'good' | 'bad' | 'neutral' | 'na';
-
-export type MetricTrendPreference = 'high' | 'low';

src/lib/types/ranking.ts

@@ -7,11 +7,10 @@ export type SortKey = 'integral' | 'minGroup' | 'fpr' | 'fnr' | 'p95' | 'size';
 
 export type SortDirection = 'asc' | 'desc';
 
-export type ExtremaKey = 'fpr' | 'fnr' | 'p95';
-
 export type RankingColumn = {
 	key: SortKey;
 	label: Message;
+	directionLabel?: Message;
 	align?: 'left' | 'right';
 	polarity: 1 | -1;
 };

src/lib/types/tools-data.ts

@@ -90,6 +90,7 @@ export type ToolsPerformanceVisualization = {
 			p95: number | null;
 			p99: number | null;
 		};
+		errorRate: number | null;
 		integral: number | null;
 	}>;
 };

src/lib/types/tools.ts

@@ -1,23 +1,13 @@
-import type { GuardModel } from './leaderboard';
-
 export type ToolsLanguageFilter = 'all' | 'en' | 'ru' | 'uk' | 'be' | 'uz';
 export type ToolsVizMetric = 'fnr' | 'fpr' | 'f1';
 export type ToolsRadarScale = 'score' | 'f1';
 export type ToolsHeatmapMetric = ToolsVizMetric | 'score';
 export type ToolsSortDirection = 'asc' | 'desc';
-export type ToolsLatencySortKey = 'model' | 'p50' | 'p95' | 'p99';
-export type ToolsMetricTone = 'good' | 'bad' | 'neutral';
-export type ToolsTopModelKey = 'integral' | 'p95' | 'fnr';
-
-export type ToolsPoint = {
-	x: number;
-	y: number;
-};
+export type ToolsLatencySortKey = 'model' | 'p50' | 'p95' | 'p99' | 'errorRate';
 
 export type ToolsTooltipRow = {
 	label: string;
 	value: string;
-	tone?: ToolsMetricTone;
 };
 
 export type ToolsChartTooltip = {
@@ -26,23 +16,3 @@ export type ToolsChartTooltip = {
 	title: string;
 	rows: ToolsTooltipRow[];
 };
-
-export type ToolsHeatmapSelection = {
-	model: GuardModel;
-	group: string;
-	rowLabel: string;
-	metric: ToolsHeatmapMetric;
-	value: number | null;
-	metrics: {
-		score: number | null;
-		f1: number | null;
-		fpr: number | null;
-		fnr: number | null;
-	};
-};
-
-export type ToolsHeatmapMetricItem = {
-	label: string;
-	key: ToolsHeatmapMetric;
-	value: number | null;
-};

src/routes/+page.svelte

@@ -18,16 +18,12 @@
 		buildRankingRows,
 		compareRankingRows,
 		filterRankingGroups,
-		rankingExtrema,
 		rankingRowsToCsv,
 		rankingScopeGroups
 	} from '$lib/data/ranking';
-	import { metricGlyph, metricState, scoreState, valueTrend } from '$lib/helpers/metrics';
 	import type { LocalizedString, Message, SelectOption } from '$lib/types/i18n';
 	import type { BenchmarkScope, GuardModel } from '$lib/types/leaderboard';
-	import type { MetricName } from '$lib/types/metrics';
 	import type {
-		ExtremaKey,
 		LanguageFilter,
 		LeaderboardRow,
 		RankingColumn,
@@ -39,6 +35,8 @@
 	type LeaderboardTableColumn = TableColumn<SortKey | 'rank' | 'model'> & {
 		metric?: RankingColumn;
 	};
+	type ColoredRankingKey = Exclude<SortKey, 'size'>;
+	type ColumnExtrema = { best: number | null; worst: number | null };
 
 	let { data }: PageProps = $props();
 
@@ -68,11 +66,41 @@
 	];
 
 	const columns: RankingColumn[] = [
-		{ key: 'integral', label: m.ranking_column_integral, align: 'right', polarity: 1 },
-		{ key: 'minGroup', label: m.ranking_column_weakest_link, align: 'right', polarity: 1 },
-		{ key: 'fpr', label: m.ranking_column_fpr, align: 'right', polarity: -1 },
-		{ key: 'fnr', label: m.ranking_column_fnr, align: 'right', polarity: -1 },
-		{ key: 'p95', label: m.ranking_column_p95, align: 'right', polarity: -1 },
+		{
+			key: 'integral',
+			label: m.ranking_column_integral,
+			directionLabel: m.ranking_column_higher_better,
+			align: 'right',
+			polarity: 1
+		},
+		{
+			key: 'minGroup',
+			label: m.ranking_column_min_score,
+			directionLabel: m.ranking_column_higher_better,
+			align: 'right',
+			polarity: 1
+		},
+		{
+			key: 'fpr',
+			label: m.ranking_column_fpr,
+			directionLabel: m.ranking_column_lower_better,
+			align: 'right',
+			polarity: -1
+		},
+		{
+			key: 'fnr',
+			label: m.ranking_column_fnr,
+			directionLabel: m.ranking_column_lower_better,
+			align: 'right',
+			polarity: -1
+		},
+		{
+			key: 'p95',
+			label: m.ranking_column_p95,
+			directionLabel: m.ranking_column_lower_better,
+			align: 'right',
+			polarity: -1
+		},
 		{ key: 'size', label: m.ranking_column_size, align: 'right', polarity: -1 }
 	];
 
@@ -113,19 +141,19 @@
 		{
 			key: 'rank',
 			label: '#',
-			headClass: 'w-12 text-right',
+			headClass: 'w-12 align-middle text-right',
 			cellClass: 'font-mono text-right text-muted-foreground'
 		},
 		{
 			key: 'model',
 			label: m.ranking_column_model,
-			headClass: 'min-w-[300px]'
+			headClass: 'min-w-[300px] align-middle'
 		},
 		...columns.map((column) => ({
 			key: column.key,
 			label: column.label,
 			metric: column,
-			headClass: cn(column.align === 'right' && 'text-right'),
+			headClass: cn('align-top', column.align === 'right' && 'text-right'),
 			cellClass: cn(column.align === 'right' && 'text-right'),
 			ariaSort: () =>
 				sortKey === column.key ? (sortDirection === 'asc' ? 'ascending' : 'descending') : 'none'
@@ -201,11 +229,19 @@
 	const selectedBenchmarkDetails = $derived.by(() =>
 		selectedRow ? benchmarkDetails(selectedRow.model, activeGroups) : []
 	);
-	const metricExtrema = $derived.by(() => ({
-		fpr: rankingExtrema(sortedRows, 'fpr', -1),
-		fnr: rankingExtrema(sortedRows, 'fnr', -1),
-		p95: rankingExtrema(sortedRows, 'p95', -1)
-	}));
+	const metricExtrema = $derived.by(
+		() =>
+			Object.fromEntries(
+				columns
+					.filter(
+						(column): column is RankingColumn & { key: ColoredRankingKey } => column.key !== 'size'
+					)
+					.map((column) => [
+						column.key,
+						rankingColumnExtrema(sortedRows, column.key, column.polarity)
+					])
+			) as Record<ColoredRankingKey, ColumnExtrema>
+	);
 
 	function selectedOption<T extends string>(options: SelectOption<T>[], value: T) {
 		return options.find((option) => option.value === value) ?? options[0];
@@ -320,49 +356,54 @@
 		return isNumber(value) ? `${(value * 100).toFixed(1)}%` : '-';
 	}
 
-	function extremaTone(key: ExtremaKey, value: number | null) {
-		if (!isNumber(value)) return '';
-
-		const extremaForKey = metricExtrema[key];
-		if (value === extremaForKey.best)
-			return 'ring-1 ring-inset ring-emerald-500/50 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
-		if (value === extremaForKey.worst)
-			return 'ring-1 ring-inset ring-red-500/50 bg-red-500/10 text-red-700 dark:text-red-300';
+	function roundedRankingValue(row: LeaderboardRow, key: ColoredRankingKey) {
+		if (key === 'integral') return Number(row.integral.toFixed(3));
+		if (key === 'minGroup') return Number(row.minGroup.toFixed(3));
+		if (key === 'p95') return Number(row.p95.toFixed(1));
+		const value = row[key];
+		return isNumber(value) ? Number((value * 100).toFixed(1)) : null;
+	}
 
-		return '';
+	function rankingColumnExtrema(
+		visibleRows: readonly LeaderboardRow[],
+		key: ColoredRankingKey,
+		polarity: 1 | -1
+	): ColumnExtrema {
+		const values = visibleRows.map((row) => roundedRankingValue(row, key)).filter(isNumber);
+		if (!values.length) return { best: null, worst: null };
+		const min = Math.min(...values);
+		const max = Math.max(...values);
+		return polarity >= 0 ? { best: max, worst: min } : { best: min, worst: max };
 	}
 
-	function metricTone(metric: MetricName, value: number | null) {
-		const state = metricState(metric, value);
-		if (state === 'good') return 'bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
-		if (state === 'bad') return 'bg-red-500/10 text-red-700 dark:text-red-300';
-		if (state === 'na') return 'text-muted-foreground';
+	function extremaTone(value: number | null, extrema: ColumnExtrema) {
+		if (!isNumber(value)) return 'text-muted-foreground';
+		if (value === extrema.best)
+			return 'ring-1 ring-inset ring-emerald-500/50 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
+		if (value === extrema.worst)
+			return 'ring-1 ring-inset ring-red-500/50 bg-red-500/10 text-red-700 dark:text-red-300';
 		return 'text-foreground';
 	}
 
-	function metricCellClass(metric: MetricName, value: number | null, extremaKey?: ExtremaKey) {
+	function rankingMetricClass(row: LeaderboardRow, key: ColoredRankingKey) {
 		return cn(
 			'inline-flex min-w-16 items-center justify-center gap-1 rounded-sm px-1.5 py-0.5 font-mono text-xs tabular-nums',
-			metricTone(metric, value),
-			extremaKey && extremaTone(extremaKey, value)
+			extremaTone(roundedRankingValue(row, key), metricExtrema[key])
 		);
 	}
 
-	function overviewMetricClass(metric: MetricName, value: number | null) {
+	function overviewMetricClass(row: LeaderboardRow, key: ColoredRankingKey) {
 		return cn(
 			'inline-flex w-fit items-center gap-1 whitespace-nowrap rounded-sm px-1.5 py-0.5 font-mono text-xl font-semibold tracking-normal tabular-nums',
-			metricTone(metric, value)
+			extremaTone(roundedRankingValue(row, key), metricExtrema[key])
 		);
 	}
 
-	function scoreTone(value: number | null) {
-		const state = scoreState(value);
-		if (state === 'na') return 'border-border bg-muted text-muted-foreground';
-		if (state === 'good')
-			return 'border-emerald-500/30 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
-		if (state === 'neutral')
-			return 'border-amber-500/30 bg-amber-500/10 text-amber-700 dark:text-amber-300';
-		return 'border-red-500/30 bg-red-500/10 text-red-700 dark:text-red-300';
+	function scoreTone(row: LeaderboardRow) {
+		return cn(
+			'border-transparent',
+			extremaTone(roundedRankingValue(row, 'integral'), metricExtrema.integral)
+		);
 	}
 
 	function modelDescription(model: GuardModel) {
@@ -382,25 +423,17 @@
 		return row.model.p50.toFixed(1);
 	}
 
-	function qualityTone(value: number | null) {
-		const state = scoreState(value);
-		if (state === 'na') return 'text-muted-foreground';
-		if (state === 'good') return 'bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
-		if (state === 'neutral') return 'bg-amber-500/10 text-amber-700 dark:text-amber-300';
-		return 'bg-red-500/10 text-red-700 dark:text-red-300';
-	}
-
 	function detailMetricClass(value: number | null) {
 		return cn(
 			'inline-flex min-w-16 items-center justify-center gap-1 whitespace-nowrap rounded-sm px-2 py-0.5 font-mono tabular-nums',
-			qualityTone(value)
+			!isNumber(value) ? 'text-muted-foreground' : 'text-foreground'
 		);
 	}
 
 	function benchmarkScoreClass(value: number | null) {
 		return cn(
 			'inline-flex shrink-0 items-center justify-center gap-1 whitespace-nowrap rounded-sm px-2 py-0.5 font-mono text-sm font-semibold tabular-nums',
-			qualityTone(value)
+			!isNumber(value) ? 'text-muted-foreground' : 'text-foreground'
 		);
 	}
 </script>
@@ -555,13 +588,25 @@
 				<button
 					type="button"
 					class={cn(
-						'inline-flex w-full items-center gap-1.5 text-sm font-medium text-inherit transition-colors hover:text-inherit',
-						metric.align === 'right' && 'justify-end'
+						'inline-flex w-full justify-start flex-col gap-0.5 text-sm font-medium text-inherit transition-colors hover:text-inherit',
+						metric.align === 'right' ? 'items-end text-right' : 'items-start text-left'
 					)}
 					onclick={() => updateSort(metric.key)}
 				>
-					{metric.label()}
-					<ArrowUpDownIcon class="size-3.5" />
+					<span
+						class={cn(
+							'inline-flex items-center gap-1.5',
+							metric.align === 'right' && 'justify-end'
+						)}
+					>
+						{metric.label()}
+						<ArrowUpDownIcon class="size-3.5" />
+					</span>
+					{#if metric.directionLabel}
+						<span class="text-[10px] leading-3 font-light text-muted-foreground normal-case">
+							{metric.directionLabel()}
+						</span>
+					{/if}
 				</button>
 			{/if}
 		{/snippet}
@@ -587,16 +632,14 @@
 							<span aria-hidden="true">△</span>
 							{m.ranking_eval_leak()}
 						</Badge>
-					{:else}
+					{/if}
+					{#if row.model.partial}
 						<Badge
 							variant="outline"
-							class="h-5 gap-1 border-[#0dce97]/50 bg-[#0dce97]/10 px-2 font-mono text-xs font-semibold text-[#0dce97]"
+							class="h-5 gap-1 border-sky-500/50 bg-sky-500/10 px-2 font-mono text-xs font-semibold text-sky-600 dark:text-sky-300"
 						>
-							<span
-								class="size-1.5 rounded-full bg-[#0dce97] shadow-[0_0_0_2px_rgba(13,206,151,0.18)]"
-								aria-hidden="true"
-							></span>
-							{m.ranking_eval_clean()}
+							<span aria-hidden="true">◐</span>
+							{m.ranking_eval_partial()}
 						</Badge>
 					{/if}
 				</div>
@@ -607,39 +650,24 @@
 			{@const data = row}
 			{#if column.key === 'integral'}
 				<span
-					class={cn(
-						'inline-flex rounded-full border px-2 py-1 font-mono text-xs',
-						scoreTone(data.integral)
-					)}
+					class={cn('inline-flex rounded-full border px-2 py-1 font-mono text-xs', scoreTone(data))}
 				>
 					{fmtScore(data.integral)}
 				</span>
 			{:else if column.key === 'minGroup'}
-				<span class={metricCellClass('score', data.minGroup)}>
-					{#if metricGlyph('score', data.minGroup)}
-						<span aria-hidden="true">{metricGlyph('score', data.minGroup)}</span>
-					{/if}
+				<span class={rankingMetricClass(data, 'minGroup')}>
 					{fmtScore(data.minGroup)}
 				</span>
 			{:else if column.key === 'fpr'}
-				<span class={metricCellClass('fpr', data.fpr, 'fpr')}>
-					{#if metricGlyph('fpr', data.fpr)}
-						<span aria-hidden="true">{metricGlyph('fpr', data.fpr)}</span>
-					{/if}
+				<span class={rankingMetricClass(data, 'fpr')}>
 					{fmtPercent(data.fpr)}
 				</span>
 			{:else if column.key === 'fnr'}
-				<span class={metricCellClass('fnr', data.fnr, 'fnr')}>
-					{#if metricGlyph('fnr', data.fnr)}
-						<span aria-hidden="true">{metricGlyph('fnr', data.fnr)}</span>
-					{/if}
+				<span class={rankingMetricClass(data, 'fnr')}>
 					{fmtPercent(data.fnr)}
 				</span>
 			{:else if column.key === 'p95'}
-				<span class={metricCellClass('latency', data.p95, 'p95')}>
-					{#if metricGlyph('latency', data.p95)}
-						<span aria-hidden="true">{metricGlyph('latency', data.p95)}</span>
-					{/if}
+				<span class={rankingMetricClass(data, 'p95')}>
 					{data.p95.toFixed(1)}
 				</span>
 			{:else if column.key === 'size'}
@@ -654,6 +682,7 @@
 			getRowKey={(row) => modelFilterValue(row.model)}
 			rowLabel={(row) => m.ranking_open_details_label({ model: row.model.short })}
 			rowClass="cursor-pointer focus-visible:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
+			headerClass="bg-muted/50 hover:bg-muted/50 [&_th]:pt-5 [&_th]:pb-4 [&_th]:text-muted-foreground"
 			onrowclick={(row) => openModelDetails(row)}
 		>
 			{#snippet header(column)}
@@ -691,7 +720,7 @@
 							<span
 								class={cn(
 									'inline-flex rounded-full border px-2.5 py-1 font-mono text-base font-semibold',
-									scoreTone(row.integral)
+									scoreTone(row)
 								)}
 							>
 								{fmtScore(row.integral)}
@@ -701,12 +730,9 @@
 							<p
 								class="font-mono text-xs font-semibold uppercase tracking-wider text-muted-foreground"
 							>
-								{m.ranking_column_weakest_link()}
+								{m.ranking_column_min_score()}
 							</p>
-							<span class={overviewMetricClass('score', row.minGroup)}>
-								{#if metricGlyph('score', row.minGroup)}
-									<span aria-hidden="true">{metricGlyph('score', row.minGroup)}</span>
-								{/if}
+							<span class={overviewMetricClass(row, 'minGroup')}>
 								{fmtScore(row.minGroup)}
 							</span>
 						</div>
@@ -716,10 +742,7 @@
 							>
 								{m.ranking_column_fpr()}
 							</p>
-							<span class={overviewMetricClass('fpr', row.fpr)}>
-								{#if metricGlyph('fpr', row.fpr)}
-									<span aria-hidden="true">{metricGlyph('fpr', row.fpr)}</span>
-								{/if}
+							<span class={overviewMetricClass(row, 'fpr')}>
 								{fmtPercent(row.fpr)}
 							</span>
 						</div>
@@ -729,10 +752,7 @@
 							>
 								{m.ranking_column_fnr()}
 							</p>
-							<span class={overviewMetricClass('fnr', row.fnr)}>
-								{#if metricGlyph('fnr', row.fnr)}
-									<span aria-hidden="true">{metricGlyph('fnr', row.fnr)}</span>
-								{/if}
+							<span class={overviewMetricClass(row, 'fnr')}>
 								{fmtPercent(row.fnr)}
 							</span>
 						</div>
@@ -792,21 +812,12 @@
 									{/if}
 								</div>
 								<span class={detailMetricClass(benchmark?.score ?? null)}>
-									{#if valueTrend(benchmark?.score ?? null, 'high')}
-										{valueTrend(benchmark?.score ?? null, 'high')}
-									{/if}
 									{fmtScore(benchmark?.score ?? null)}
 								</span>
 								<span class={detailMetricClass(benchmark?.fpr ?? null)}>
-									{#if valueTrend(benchmark?.fpr ?? null, 'low')}
-										{valueTrend(benchmark?.fpr ?? null, 'low')}
-									{/if}
 									{fmtPercent(benchmark?.fpr ?? null)}
 								</span>
 								<span class={detailMetricClass(benchmark?.fnr ?? null)}>
-									{#if valueTrend(benchmark?.fnr ?? null, 'low')}
-										{valueTrend(benchmark?.fnr ?? null, 'low')}
-									{/if}
 									{fmtPercent(benchmark?.fnr ?? null)}
 								</span>
 								<span class="justify-self-center font-mono tabular-nums">
@@ -829,9 +840,6 @@
 										{groupLabel(benchmark.name)}
 									</h3>
 									<span class={benchmarkScoreClass(benchmark.score)}>
-										{#if valueTrend(benchmark.score, 'high')}
-											{valueTrend(benchmark.score, 'high')}
-										{/if}
 										{fmtScore(benchmark.score)}
 									</span>
 								</div>
@@ -915,16 +923,14 @@
 									<span aria-hidden="true">△</span>
 									{m.ranking_eval_leak()}
 								</Badge>
-							{:else}
+							{/if}
+							{#if row.model.partial}
 								<Badge
 									variant="outline"
-									class="h-5 gap-1 rounded-full border-[#0dce97]/50 bg-[#0dce97]/10 px-2 font-mono text-[0.6875rem] font-semibold text-[#0dce97]"
+									class="h-5 gap-1 rounded-full border-sky-500/50 bg-sky-500/10 px-2 font-mono text-[0.6875rem] font-semibold text-sky-600 dark:text-sky-300"
 								>
-									<span
-										class="size-1 rounded-full bg-[#0dce97] shadow-[0_0_0_2px_rgba(13,206,151,0.18)]"
-										aria-hidden="true"
-									></span>
-									{m.ranking_eval_clean()}
+									<span aria-hidden="true">◐</span>
+									{m.ranking_eval_partial()}
 								</Badge>
 							{/if}
 						</div>

src/routes/details/+page.svelte

@@ -16,12 +16,6 @@
 		buildBenchmarkCards,
 		buildMatrixGroups,
 		buildMatrixRows,
-		detailMetricGlyph as matrixCellGlyph,
-		detailMetricIsBad as metricIsBad,
-		detailMetricIsGood as metricIsGood,
-		familyMetricGlyph,
-		familyMetricIsBad,
-		familyMetricIsGood,
 		filterModels,
 		matrixCellValue,
 		matrixDatasetLabel,
@@ -38,7 +32,6 @@
 	import { getLocale } from '$lib/paraglide/runtime';
 	import type {
 		BenchmarkFamilyLanguage,
-		BenchmarkLeaderMetric,
 		DetailMetric,
 		DetailsLanguageFilter as LanguageFilter,
 		DetailTab,
@@ -53,6 +46,8 @@
 	import { cn } from '$lib/utils';
 	import type { PageProps } from './$types';
 
+	type ColumnExtrema = { best: number | null; worst: number | null };
+
 	const selectContentClass = 'max-h-72';
 	let { data }: PageProps = $props();
 	const detailsState = $derived(data.detailsState);
@@ -66,7 +61,6 @@
 	let benchmarkLanguage = $state<LanguageFilter>('any');
 	let benchmarkSearch = $state('');
 	let benchmarkFamilyLanguage = $state<BenchmarkFamilyLanguage>('all');
-	let benchmarkFamilyMetric = $state<BenchmarkLeaderMetric>('f1');
 	let benchmarkFamilySearch = $state('');
 	let benchmarkSheetOpen = $state(false);
 	let selectedBenchmark = $state<string | null>(null);
@@ -105,11 +99,6 @@
 		{ value: 'be', label: () => 'BE' },
 		{ value: 'uz', label: () => 'UZ' }
 	];
-	const benchmarkFamilyMetricOptions: SelectOption<BenchmarkLeaderMetric>[] = [
-		{ value: 'f1', label: m.details_family_metric_f1 },
-		{ value: 'fpr', label: m.details_family_metric_fpr },
-		{ value: 'fnr', label: m.details_family_metric_fnr }
-	];
 	const benchmarkMetricOptions: SelectOption<DetailMetric>[] = [
 		{ value: 'score', label: () => 'Score' },
 		{ value: 'f1', label: () => 'F1' },
@@ -142,9 +131,6 @@
 	const currentBenchmarkFamilyLanguage = $derived(
 		selectedOption(benchmarkFamilyLanguageOptions, benchmarkFamilyLanguage)
 	);
-	const currentBenchmarkFamilyMetric = $derived(
-		selectedOption(benchmarkFamilyMetricOptions, benchmarkFamilyMetric)
-	);
 	const currentBenchmarkMetric = $derived(selectedOption(benchmarkMetricOptions, benchmarkMetric));
 	const currentBenchmarkLanguage = $derived(selectedOption(languageOptions, benchmarkLanguage));
 	const currentModelLanguage = $derived(selectedOption(languageOptions, modelLanguage));
@@ -186,14 +172,14 @@
 			details.datasetMatrix.cellsByModelId
 		)
 	);
+	const matrixColumnExtrema = $derived.by(
+		() =>
+			Object.fromEntries(
+				visibleMatrixDatasets.map((dataset) => [dataset.name, detailMatrixColumnExtrema(dataset)])
+			) as Record<string, ColumnExtrema>
+	);
 	const benchmarkCards = $derived.by(() =>
-		buildBenchmarkCards(
-			benchmarkFamilySearch,
-			benchmarkFamilyLanguage,
-			benchmarkFamilyMetric,
-			details,
-			getLocale()
-		)
+		buildBenchmarkCards(benchmarkFamilySearch, benchmarkFamilyLanguage, details, getLocale())
 	);
 	const selectedBenchmarkMeta = $derived(
 		selectedBenchmark ? (benchmarkMeta[selectedBenchmark] ?? null) : null
@@ -246,6 +232,40 @@
 		return fmtPercent(value);
 	}
 
+	function roundedDetailMetric(value: number | null, metric: DetailMetric) {
+		if (!isNumber(value)) return null;
+		if (metric === 'score' || metric === 'f1') return Number(value.toFixed(3));
+		return Number((value * 100).toFixed(1));
+	}
+
+	function detailMatrixColumnExtrema(dataset: MatrixDataset): ColumnExtrema {
+		const values = matrixRows
+			.map((model) =>
+				roundedDetailMetric(
+					matrixCellValue(model, dataset, benchmarkMetric, details.datasetMatrix.cellsByModelId),
+					benchmarkMetric
+				)
+			)
+			.filter(isNumber);
+
+		if (!values.length) return { best: null, worst: null };
+		const min = Math.min(...values);
+		const max = Math.max(...values);
+		return benchmarkMetric === 'fpr' || benchmarkMetric === 'fnr'
+			? { best: min, worst: max }
+			: { best: max, worst: min };
+	}
+
+	function extremaTone(value: number | null, extrema: ColumnExtrema | undefined) {
+		if (!isNumber(value)) return 'text-muted-foreground';
+		if (!extrema || extrema.best === null || extrema.worst === null) return 'text-foreground';
+		if (value === extrema.best)
+			return 'matrix-cell-good !bg-emerald-500/10 text-emerald-700 dark:text-emerald-300';
+		if (value === extrema.worst)
+			return 'matrix-cell-bad !bg-red-500/10 text-red-700 dark:text-red-300';
+		return 'text-foreground';
+	}
+
 	function messageLabel(label: string | Message) {
 		return typeof label === 'function' ? label() : label;
 	}
@@ -265,59 +285,30 @@
 		return matrixSortDirection === -1 ? 'descending' : 'ascending';
 	}
 
-	function metricCellClass(value: number | null, metric: DetailMetric) {
+	function metricCellClass(value: number | null, metric: DetailMetric, datasetName: string) {
 		return cn(
 			'matrix-cell min-w-32 border-b border-l px-3 py-3 text-right font-mono text-sm tabular-nums',
 			!isNumber(value) && 'matrix-cell-empty !bg-muted/30 text-muted-foreground',
-			metricIsGood(value, metric) &&
-				'matrix-cell-good !bg-emerald-500/10 text-emerald-700 dark:text-emerald-300',
-			!metricIsGood(value, metric) &&
-				!metricIsBad(value, metric) &&
-				isNumber(value) &&
-				'matrix-cell-neutral text-foreground',
-			metricIsBad(value, metric) && 'matrix-cell-bad !bg-red-500/10 text-red-700 dark:text-red-300'
-		);
-	}
-
-	function matrixCellClass(value: number | null, metric: DetailMetric, startsGroup: boolean) {
-		return cn(metricCellClass(value, metric), startsGroup && 'border-l-border');
-	}
-
-	function scorePillClass(value: number | null) {
-		return cn(
-			'inline-flex shrink-0 items-center gap-1 font-mono text-sm font-semibold tabular-nums',
-			!isNumber(value) &&
-				'rounded-full border border-border bg-muted px-2.5 py-1 text-muted-foreground',
-			isNumber(value) &&
-				value >= 0.85 &&
-				'rounded-full border border-emerald-500/30 bg-emerald-500/10 px-2.5 py-1 text-emerald-700 dark:text-emerald-300',
-			isNumber(value) && value >= 0.6 && value < 0.85 && 'text-foreground',
 			isNumber(value) &&
-				value < 0.6 &&
-				'rounded-full border border-red-500/30 bg-red-500/10 px-2.5 py-1 text-red-700 dark:text-red-300'
+				extremaTone(roundedDetailMetric(value, metric), matrixColumnExtrema[datasetName])
 		);
 	}
 
-	function scoreGlyph(value: number | null) {
-		if (!isNumber(value)) return '○';
-		if (value >= 0.85) return '▲';
-		if (value < 0.6) return '▼';
-		return '-';
+	function matrixCellClass(
+		value: number | null,
+		metric: DetailMetric,
+		datasetName: string,
+		startsGroup: boolean
+	) {
+		return cn(metricCellClass(value, metric, datasetName), startsGroup && 'border-l-border');
 	}
 
-	function familyMetricPillClass(value: number | null, metric: BenchmarkLeaderMetric) {
-		const good = familyMetricIsGood(value, metric);
-		const bad = familyMetricIsBad(value, metric);
-
+	function scorePillClass(value: number | null) {
 		return cn(
 			'inline-flex shrink-0 items-center gap-1 font-mono text-sm font-semibold tabular-nums',
 			!isNumber(value) &&
 				'rounded-full border border-border bg-muted px-2.5 py-1 text-muted-foreground',
-			good &&
-				'rounded-full border border-emerald-500/30 bg-emerald-500/10 px-2.5 py-1 text-emerald-700 dark:text-emerald-300',
-			isNumber(value) && !good && !bad && 'text-foreground',
-			bad &&
-				'rounded-full border border-red-500/30 bg-red-500/10 px-2.5 py-1 text-red-700 dark:text-red-300'
+			isNumber(value) && 'text-foreground'
 		);
 	}
 
@@ -342,64 +333,31 @@
 		openModelSheet(model);
 	}
 
-	function rankingMetricClass(value: number | null, metric: BenchmarkLeaderMetric) {
+	function rankingMetricClass(value: number | null) {
 		return cn(
 			'inline-flex min-w-16 items-center justify-center rounded-sm px-1.5 py-0.5 font-mono text-xs tabular-nums',
 			!isNumber(value) && 'text-muted-foreground',
-			isNumber(value) &&
-				metricIsGood(value, metric) &&
-				'bg-emerald-500/10 text-emerald-700 dark:text-emerald-300',
-			isNumber(value) &&
-				metricIsBad(value, metric) &&
-				'bg-red-500/10 text-red-700 dark:text-red-300',
-			isNumber(value) &&
-				!metricIsGood(value, metric) &&
-				!metricIsBad(value, metric) &&
-				'text-foreground'
+			isNumber(value) && 'text-foreground'
 		);
 	}
 
-	function compactMetricClass(value: number | null, metric: DetailMetric) {
+	function compactMetricClass(value: number | null) {
 		return cn(
 			'inline-flex w-fit items-center gap-1 rounded-sm px-1.5 py-0.5 font-mono text-xs font-semibold tabular-nums',
 			!isNumber(value) && 'text-muted-foreground',
-			isNumber(value) &&
-				metricIsGood(value, metric) &&
-				'bg-emerald-500/10 text-emerald-700 dark:text-emerald-300',
-			isNumber(value) &&
-				metricIsBad(value, metric) &&
-				'bg-red-500/10 text-red-700 dark:text-red-300',
-			isNumber(value) &&
-				!metricIsGood(value, metric) &&
-				!metricIsBad(value, metric) &&
-				'text-foreground'
+			isNumber(value) && 'text-foreground'
 		);
 	}
 
-	function modelKpiClass(value: number | null, metric: DetailMetric) {
+	function modelKpiClass(value: number | null) {
 		return cn(
 			'inline-flex w-fit items-center gap-2 font-mono text-3xl font-bold tracking-normal tabular-nums',
 			!isNumber(value) &&
 				'rounded-lg border border-border bg-muted px-3 py-1.5 text-muted-foreground',
-			isNumber(value) &&
-				metricIsGood(value, metric) &&
-				'rounded-lg border border-emerald-500/30 bg-emerald-500/10 px-3 py-1.5 text-emerald-700 dark:text-emerald-300',
-			isNumber(value) &&
-				metricIsBad(value, metric) &&
-				'rounded-lg border border-red-500/30 bg-red-500/10 px-3 py-1.5 text-red-700 dark:text-red-300',
-			isNumber(value) &&
-				!metricIsGood(value, metric) &&
-				!metricIsBad(value, metric) &&
-				'text-foreground'
+			isNumber(value) && 'text-foreground'
 		);
 	}
 
-	function groupBarFillClass(value: number) {
-		if (value >= 0.85) return 'bg-emerald-500';
-		if (value >= 0.6) return 'bg-amber-500';
-		return 'bg-red-500';
-	}
-
 	function groupChipIsActive(groups: string[]) {
 		return groups.some((group) => selectedGroups.includes(group));
 	}
@@ -644,7 +602,7 @@
 								<tbody>
 									{#each matrixRows as model (modelKey(model))}
 										<tr
-											class="[&:nth-child(even)>td.matrix-cell-neutral]:bg-muted/20 hover:[&>td.matrix-cell-bad]:!bg-red-500/15 hover:[&>td.matrix-cell-empty]:!bg-muted/40 hover:[&>td.matrix-cell-good]:!bg-emerald-500/15 hover:[&>td.matrix-cell-neutral]:!bg-muted/40 hover:[&>th]:bg-muted"
+											class="hover:[&>td.matrix-cell-bad]:!bg-red-500/15 hover:[&>td.matrix-cell-empty]:!bg-muted/40 hover:[&>td.matrix-cell-good]:!bg-emerald-500/15 hover:[&>th]:bg-muted"
 										>
 											<th
 												class="z-20 min-w-80 max-w-80 border-r border-b bg-card px-4 py-3 text-left md:sticky md:left-0"
@@ -682,12 +640,15 @@
 													)}
 													<td
 														class={cn(
-															matrixCellClass(value, benchmarkMetric, datasetIndex === 0),
+															matrixCellClass(
+																value,
+																benchmarkMetric,
+																dataset.name,
+																datasetIndex === 0
+															),
 															groupIndex === 0 && datasetIndex === 0 && 'border-l-0'
 														)}
 													>
-														<span aria-hidden="true">{matrixCellGlyph(value, benchmarkMetric)}</span
-														>
 														{#if isNumber(value)}
 															{fmtMetric(value, benchmarkMetric)}
 														{/if}
@@ -708,17 +669,7 @@
 					<div
 						class="mt-3 flex flex-wrap items-center gap-x-5 gap-y-2 text-xs text-muted-foreground"
 					>
-						<span class="inline-flex items-center gap-2">
-							<span class="size-3 rounded-sm bg-[#0dce97]"></span>
-							{m.details_matrix_good()}
-						</span>
-						<span class="inline-flex items-center gap-2">
-							<span class="size-3 rounded-sm bg-red-500"></span>
-							{m.details_matrix_bad()}
-						</span>
 						<span>{m.details_matrix_no_data()}</span>
-						<span>{m.details_matrix_direction_legend()}</span>
-						<span>{m.details_matrix_sort_legend()}</span>
 					</div>
 				</section>
 			{:else}
@@ -780,7 +731,7 @@
 				aria-label={m.details_family_filters()}
 			>
 				<div
-					class="grid items-end gap-4 lg:grid-cols-[minmax(9rem,13rem)_minmax(14rem,22rem)_minmax(14rem,1fr)_auto_auto]"
+					class="grid items-end gap-4 lg:grid-cols-[minmax(9rem,13rem)_minmax(14rem,1fr)_auto_auto]"
 				>
 					<Select
 						id="details-family-language"
@@ -794,18 +745,6 @@
 						triggerLabel={currentBenchmarkFamilyLanguage.label()}
 						contentClass={selectContentClass}
 					/>
-					<Select
-						id="details-family-metric"
-						type="single"
-						bind:value={benchmarkFamilyMetric}
-						options={benchmarkFamilyMetricOptions.map((option) => ({
-							value: option.value,
-							label: option.label()
-						}))}
-						label={m.details_family_metric_label()}
-						triggerLabel={currentBenchmarkFamilyMetric.label()}
-						contentClass={selectContentClass}
-					/>
 					<label class="min-w-0 space-y-2">
 						<span
 							class="block text-xs font-semibold tracking-wider text-muted-foreground uppercase"
@@ -821,7 +760,6 @@
 						variant="outline"
 						onclick={() => {
 							benchmarkFamilyLanguage = 'all';
-							benchmarkFamilyMetric = 'f1';
 							benchmarkFamilySearch = '';
 						}}
 					>
@@ -833,25 +771,8 @@
 						{benchmarkCards.length} / {allGroups.length}
 					</span>
 				</div>
-				<p class="mt-3 text-xs text-muted-foreground">{m.details_family_metric_hint()}</p>
 			</section>
 
-			<div class="mt-3 flex flex-wrap items-center gap-x-5 gap-y-2 text-xs text-muted-foreground">
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-[#0dce97]"></span>
-					{m.details_family_legend_good()}
-				</span>
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-amber-500"></span>
-					{m.details_family_legend_middle()}
-				</span>
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-red-500"></span>
-					{m.details_family_legend_bad()}
-				</span>
-				<span>{m.details_family_legend_direction()}</span>
-			</div>
-
 			{#if benchmarkCards.length}
 				<div class="mt-12 grid min-w-0 gap-4 md:grid-cols-2 xl:grid-cols-3">
 					{#each benchmarkCards as card (card.name)}
@@ -881,22 +802,6 @@
 									</Badge>
 								{/each}
 							</div>
-							<div class="mt-auto border-t pt-4">
-								<p class="font-mono text-xs tracking-wider text-muted-foreground uppercase">
-									{m.details_leader()} · {benchmarkFamilyMetric.toUpperCase()}
-								</p>
-								<div class="mt-2 flex min-w-0 items-center justify-between gap-3">
-									<p class="min-w-0 truncate font-mono text-sm font-semibold">
-										{card.leader?.model.short ?? m.details_no_data()}
-									</p>
-									<span
-										class={cn('shrink-0', familyMetricPillClass(card.score, benchmarkFamilyMetric))}
-									>
-										{familyMetricGlyph(card.score, benchmarkFamilyMetric)}
-										{fmtScore(card.score)}
-									</span>
-								</div>
-							</div>
 						</button>
 					{/each}
 				</div>
@@ -992,17 +897,17 @@
 													</p>
 												</td>
 												<td class="px-3 py-3 text-right">
-													<span class={rankingMetricClass(row.fnr, 'fnr')}>
+													<span class={rankingMetricClass(row.fnr)}>
 														{fmtMetric(row.fnr, 'fnr')}
 													</span>
 												</td>
 												<td class="px-3 py-3 text-right">
-													<span class={rankingMetricClass(row.fpr, 'fpr')}>
+													<span class={rankingMetricClass(row.fpr)}>
 														{fmtMetric(row.fpr, 'fpr')}
 													</span>
 												</td>
 												<td class="px-3 py-3 text-right">
-													<span class={rankingMetricClass(row.score, 'f1')}>
+													<span class={rankingMetricClass(row.score)}>
 														{fmtScore(row.score)}
 													</span>
 												</td>
@@ -1092,27 +997,6 @@
 				</div>
 			</section>
 
-			<div
-				class="mt-4 flex flex-wrap items-center gap-x-5 gap-y-2 rounded-2xl border bg-muted/50 px-4 py-3 text-sm text-muted-foreground shadow-sm"
-				role="note"
-				aria-label={m.details_model_legend_lead()}
-			>
-				<span class="font-semibold text-foreground">{m.details_model_legend_lead()}</span>
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-[#0dce97]"></span>
-					{m.details_model_legend_good()}
-				</span>
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-amber-500"></span>
-					{m.details_model_legend_middle()}
-				</span>
-				<span class="inline-flex items-center gap-2">
-					<span class="size-3 rounded-sm bg-red-500"></span>
-					{m.details_model_legend_bad()}
-				</span>
-				<span>{m.details_model_legend_hint()}</span>
-			</div>
-
 			<div class="mt-4 grid min-w-0 gap-4 md:grid-cols-2 xl:grid-cols-3">
 				{#each filteredModels as model (modelKey(model))}
 					<div
@@ -1133,7 +1017,6 @@
 								</h2>
 							</div>
 							<span class={cn('shrink-0', scorePillClass(model.integral))}>
-								{scoreGlyph(model.integral)}
 								{fmtScore(model.integral)}
 							</span>
 						</div>
@@ -1168,12 +1051,13 @@
 									>
 										△ {m.ranking_eval_leak()}
 									</Badge>
-								{:else}
+								{/if}
+								{#if model.partial}
 									<Badge
 										variant="outline"
-										class="rounded-full border-[#0dce97]/50 bg-[#0dce97]/10 font-mono text-[#0dce97]"
+										class="rounded-full border-sky-500/50 bg-sky-500/10 font-mono text-sky-600 dark:text-sky-300"
 									>
-										✓ {m.ranking_eval_clean()}
+										◐ {m.ranking_eval_partial()}
 									</Badge>
 								{/if}
 								<span class="min-w-0 break-words"
@@ -1185,10 +1069,13 @@
 								</span>
 							</div>
 							<div class="mt-4 flex min-w-0 items-center justify-between gap-3">
-								<p class="min-w-0 truncate font-mono text-sm text-muted-foreground">
+								<p class="min-w-0 font-mono text-sm text-muted-foreground">
 									F1 <span class="text-foreground">{fmtScore(model.f1)}</span> · p50 {fmtMs(
 										model.p50
-									)}
+									)} ·
+									<span class="whitespace-nowrap"
+										>{m.details_error_rate()} {fmtPercent(model.errorRate)}</span
+									>
 								</p>
 								<Button
 									class="shrink-0"
@@ -1250,8 +1137,7 @@
 				<div class="space-y-8">
 					<div class="grid gap-3 sm:grid-cols-2 xl:grid-cols-4">
 						<div class="rounded-xl border bg-card p-4 shadow-sm">
-							<p class={modelKpiClass(selectedModel.integral, 'f1')}>
-								{matrixCellGlyph(selectedModel.integral, 'f1')}
+							<p class={modelKpiClass(selectedModel.integral)}>
 								{fmtScore(selectedModel.integral)}
 							</p>
 							<p class="mt-3 text-xs font-semibold tracking-wider text-muted-foreground uppercase">
@@ -1259,8 +1145,7 @@
 							</p>
 						</div>
 						<div class="rounded-xl border bg-card p-4 shadow-sm">
-							<p class={modelKpiClass(selectedModel.minGroup, 'f1')}>
-								{matrixCellGlyph(selectedModel.minGroup, 'f1')}
+							<p class={modelKpiClass(selectedModel.minGroup)}>
 								{fmtScore(selectedModel.minGroup)}
 							</p>
 							<p class="mt-3 text-xs font-semibold tracking-wider text-muted-foreground uppercase">
@@ -1268,8 +1153,7 @@
 							</p>
 						</div>
 						<div class="rounded-xl border bg-card p-4 shadow-sm">
-							<p class={modelKpiClass(selectedModel.fnr, 'fnr')}>
-								{matrixCellGlyph(selectedModel.fnr, 'fnr')}
+							<p class={modelKpiClass(selectedModel.fnr)}>
 								{fmtPercent(selectedModel.fnr)}
 							</p>
 							<p class="mt-3 text-xs font-semibold tracking-wider text-muted-foreground uppercase">
@@ -1277,8 +1161,7 @@
 							</p>
 						</div>
 						<div class="rounded-xl border bg-card p-4 shadow-sm">
-							<p class={modelKpiClass(selectedModel.fpr, 'fpr')}>
-								{matrixCellGlyph(selectedModel.fpr, 'fpr')}
+							<p class={modelKpiClass(selectedModel.fpr)}>
 								{fmtPercent(selectedModel.fpr)}
 							</p>
 							<p class="mt-3 text-xs font-semibold tracking-wider text-muted-foreground uppercase">
@@ -1395,26 +1278,22 @@
 									<tr class="border-t transition-colors hover:bg-muted/30">
 										<td class="px-3 py-3 font-mono text-xs font-semibold">{group.name}</td>
 										<td class="px-3 py-3 text-right">
-											<span class={compactMetricClass(group.score, 'f1')}>
-												{matrixCellGlyph(group.score, 'f1')}
+											<span class={compactMetricClass(group.score)}>
 												{fmtScore(group.score)}
 											</span>
 										</td>
 										<td class="px-3 py-3 text-right">
-											<span class={compactMetricClass(group.fnr, 'fnr')}>
-												{matrixCellGlyph(group.fnr, 'fnr')}
+											<span class={compactMetricClass(group.fnr)}>
 												{fmtPercent(group.fnr)}
 											</span>
 										</td>
 										<td class="px-3 py-3 text-right">
-											<span class={compactMetricClass(group.fpr, 'fpr')}>
-												{matrixCellGlyph(group.fpr, 'fpr')}
+											<span class={compactMetricClass(group.fpr)}>
 												{fmtPercent(group.fpr)}
 											</span>
 										</td>
 										<td class="px-3 py-3 text-right">
-											<span class={compactMetricClass(group.score, 'f1')}>
-												{matrixCellGlyph(group.score, 'f1')}
+											<span class={compactMetricClass(group.score)}>
 												{fmtScore(group.score)}
 											</span>
 										</td>
@@ -1460,31 +1339,16 @@
 								</p>
 								<div class="h-2 rounded-full bg-muted">
 									<div
-										class={cn('h-full rounded-full', groupBarFillClass(group.score))}
+										class="h-full rounded-full bg-[#0dce97]"
 										style={`width:${Math.max(2, Math.min(100, group.score * 100))}%`}
 									></div>
 								</div>
-								<span class={compactMetricClass(group.score, 'f1')}>
-									{scoreGlyph(group.score)}
+								<span class={compactMetricClass(group.score)}>
 									{fmtScore(group.score)}
 								</span>
 							</div>
 						{/each}
 					</div>
-					<div class="flex flex-wrap items-center gap-x-5 gap-y-2 text-xs text-muted-foreground">
-						<span class="inline-flex items-center gap-2">
-							<span class="size-3 rounded-sm bg-emerald-500"></span>
-							≥ 0.85
-						</span>
-						<span class="inline-flex items-center gap-2">
-							<span class="size-3 rounded-sm bg-amber-500"></span>
-							0.60-0.85
-						</span>
-						<span class="inline-flex items-center gap-2">
-							<span class="size-3 rounded-sm bg-red-500"></span>
-							&lt; 0.60
-						</span>
-					</div>
 				</section>
 			{/if}
 		{/snippet}
@@ -1538,28 +1402,31 @@
 						</span>
 						<span class="min-w-0 break-words text-base">{selectedModel.license}</span>
 					</li>
-					<li class="grid items-start gap-2 sm:grid-cols-[14rem_minmax(0,1fr)]">
-						<span class="text-sm font-semibold tracking-wider text-muted-foreground uppercase">
-							{m.details_model_sheet_eval_leak()}
-						</span>
-						<span>
-							{#if selectedModel.leak}
-								<Badge
-									variant="outline"
-									class="rounded-full border-amber-500/50 bg-amber-500/10 font-mono text-amber-500"
-								>
-									△ {m.ranking_eval_leak()}
-								</Badge>
-							{:else}
-								<Badge
-									variant="outline"
-									class="rounded-full border-[#0dce97]/50 bg-[#0dce97]/10 font-mono text-[#0dce97]"
-								>
-									✓ {m.ranking_eval_clean()}
-								</Badge>
-							{/if}
-						</span>
-					</li>
+					{#if selectedModel.leak || selectedModel.partial}
+						<li class="grid items-start gap-2 sm:grid-cols-[14rem_minmax(0,1fr)]">
+							<span class="text-sm font-semibold tracking-wider text-muted-foreground uppercase">
+								{m.details_model_sheet_status()}
+							</span>
+							<span class="flex flex-wrap gap-2">
+								{#if selectedModel.leak}
+									<Badge
+										variant="outline"
+										class="rounded-full border-amber-500/50 bg-amber-500/10 font-mono text-amber-500"
+									>
+										△ {m.ranking_eval_leak()}
+									</Badge>
+								{/if}
+								{#if selectedModel.partial}
+									<Badge
+										variant="outline"
+										class="rounded-full border-sky-500/50 bg-sky-500/10 font-mono text-sky-600 dark:text-sky-300"
+									>
+										◐ {m.ranking_eval_partial()}
+									</Badge>
+								{/if}
+							</span>
+						</li>
+					{/if}
 					<li class="grid items-start gap-2 sm:grid-cols-[14rem_minmax(0,1fr)]">
 						<span class="text-sm font-semibold tracking-wider text-muted-foreground uppercase">
 							{m.details_model_sheet_run_date()}

src/routes/methodology/+page.svelte

@@ -9,14 +9,8 @@
 		| 'group-score'
 		| 'integral-product'
 		| 'integral-log'
-		| 'fbeta'
-		| 'f2'
-		| 'f05'
-		| 'fdr'
 		| 'fnr-harm'
 		| 'fpr-safe'
-		| 'micro-f1'
-		| 'micro-pr'
 		| 'fpr-avg'
 		| 'fnr-avg';
 
@@ -30,7 +24,6 @@
 		tag?: Message;
 		formulaBlocks: FormulaBlockData[];
 		description: Message;
-		example?: Message;
 		wide?: boolean;
 	};
 
@@ -51,15 +44,13 @@
 			title: 'S_ds',
 			tag: m.methodology_dataset_score_tag,
 			formulaBlocks: [{ kind: 'dataset-score', comment: m.methodology_dataset_score_comment }],
-			description: m.methodology_dataset_score_description,
-			example: m.methodology_dataset_score_example
+			description: m.methodology_dataset_score_description
 		},
 		{
 			title: 'S_group',
 			tag: m.methodology_group_score_tag,
 			formulaBlocks: [{ kind: 'group-score', comment: m.methodology_group_score_comment }],
-			description: m.methodology_group_score_description,
-			example: m.methodology_group_score_example
+			description: m.methodology_group_score_description
 		},
 		{
 			title: 'S_integral',
@@ -68,37 +59,7 @@
 				{ kind: 'integral-product', comment: m.methodology_integral_score_comment },
 				{ kind: 'integral-log', comment: m.methodology_integral_log_comment }
 			],
-			description: m.methodology_integral_score_description,
-			example: m.methodology_integral_score_example
-		}
-	];
-
-	const helperMetricCards: FormulaCard[] = [
-		{
-			title: 'F_beta',
-			tag: m.methodology_fbeta_tag,
-			formulaBlocks: [{ kind: 'fbeta' }],
-			description: m.methodology_fbeta_description
-		},
-		{
-			title: 'F2',
-			tag: m.methodology_f2_tag,
-			formulaBlocks: [{ kind: 'f2' }],
-			description: m.methodology_f2_description,
-			example: m.methodology_f2_example
-		},
-		{
-			title: 'F0.5',
-			tag: m.methodology_f05_tag,
-			formulaBlocks: [{ kind: 'f05' }],
-			description: m.methodology_f05_description,
-			example: m.methodology_f05_example
-		},
-		{
-			title: 'FDR',
-			formulaBlocks: [{ kind: 'fdr' }],
-			description: m.methodology_fdr_description,
-			example: m.methodology_fdr_example
+			description: m.methodology_integral_score_description
 		}
 	];
 
@@ -114,13 +75,6 @@
 			tag: m.methodology_safe_split_tag,
 			formulaBlocks: [{ kind: 'fpr-safe' }],
 			description: m.methodology_safe_split_description
-		},
-		{
-			title: 'Micro-F1',
-			tag: m.methodology_micro_f1_tag,
-			formulaBlocks: [{ kind: 'micro-f1' }, { kind: 'micro-pr' }],
-			description: m.methodology_micro_f1_description,
-			wide: true
 		}
 	];
 
@@ -138,7 +92,8 @@
 		{
 			title: m.methodology_scatter_point_title(),
 			formulaBlocks: [],
-			description: m.methodology_scatter_point_description
+			description: m.methodology_scatter_point_description,
+			wide: true
 		}
 	];
 
@@ -153,25 +108,28 @@
 			title: m.methodology_caveat_title,
 			description: m.methodology_caveat_description,
 			muted: true
-		},
-		{
-			mark: '!',
-			title: m.methodology_golden_rule_title,
-			description: m.methodology_golden_rule_description
 		}
 	];
 
 	function cardClass(card: FormulaCard) {
 		return cn(
-			'flex min-h-[28rem] min-w-0 max-w-full flex-col overflow-hidden rounded-2xl border bg-card p-5 shadow-sm',
-			card.wide && 'lg:col-span-2 xl:col-span-1'
+			'flex min-w-0 max-w-full flex-col overflow-hidden rounded-2xl border bg-card p-5 shadow-sm',
+			card.wide && 'lg:col-span-2'
+		);
+	}
+
+	function aggregationCardClass() {
+		return cn(
+			'flex min-w-0 max-w-full flex-col overflow-hidden rounded-2xl border bg-card p-5 shadow-sm'
 		);
 	}
 
-	function aggregationCardClass(card: FormulaCard) {
+	function formulaClass(kind: FormulaKind) {
 		return cn(
-			'flex min-h-[34rem] min-w-0 max-w-full flex-col overflow-hidden rounded-2xl border bg-card p-5 shadow-sm',
-			card.title === 'S_integral' && 'min-h-[34rem]'
+			'min-w-max text-foreground',
+			['dataset-score', 'group-score', 'integral-product', 'integral-log'].includes(kind)
+				? 'text-xl'
+				: 'text-lg'
 		);
 	}
 
@@ -212,9 +170,9 @@
 		>
 			{m.methodology_aggregation_title()}
 		</h2>
-		<div class="mt-3 grid min-w-0 gap-4 lg:grid-cols-3">
+		<div class="mt-3 grid min-w-0 grid-cols-1 gap-4">
 			{#each aggregationCards as card (card.title)}
-				<article class={aggregationCardClass(card)}>
+				<article class={aggregationCardClass()}>
 					<div class="flex min-w-0 flex-wrap items-start justify-between gap-3">
 						<h2
 							class="flex min-w-0 flex-wrap items-baseline gap-3 font-mono text-2xl font-semibold tracking-normal"
@@ -232,39 +190,30 @@
 						<span class="mr-1 font-semibold text-foreground">{m.methodology_meaning_label()}</span>
 						{card.description()}
 					</p>
-					{#if card.example}
-						<div class="mt-auto pt-8">
-							<p
-								class="min-w-0 rounded-xl border border-[#0dce97]/35 bg-[#0dce97]/10 px-4 py-3 font-mono text-xs leading-6 break-words text-[#0dce97]"
-							>
-								<span class="mb-1 block text-xs font-semibold uppercase tracking-wider">
-									{m.methodology_example_label()}
-								</span>
-								{@render RichText(card.example())}
-							</p>
-						</div>
-					{/if}
 				</article>
 			{/each}
 		</div>
 	</section>
 
 	{@render FormulaSection(
-		m.methodology_helper_metrics_title,
-		helperMetricCards,
-		undefined,
-		'md:grid-cols-2 xl:grid-cols-4'
+		m.methodology_combined_dataset_title,
+		combinedCards,
+		CombinedDatasetNote,
+		'grid-cols-1 lg:grid-cols-2'
 	)}
 
-	{@render FormulaSection(m.methodology_combined_dataset_title, combinedCards, CombinedDatasetNote)}
-
 	{#snippet CombinedDatasetNote()}
 		<p class="mt-4 text-sm leading-6 text-muted-foreground">
 			{m.methodology_combined_dataset_note()}
 		</p>
 	{/snippet}
 
-	{@render FormulaSection(m.methodology_scatter_title, scatterCards)}
+	{@render FormulaSection(
+		m.methodology_scatter_title,
+		scatterCards,
+		undefined,
+		'grid-cols-1 lg:grid-cols-2'
+	)}
 
 	<section class="mt-8 min-w-0 space-y-4" aria-labelledby="specifics-title">
 		<h2
@@ -306,20 +255,6 @@
 			</div>
 		{/each}
 	</section>
-
-	<div class="mt-8 flex min-w-0 flex-wrap gap-3 font-mono text-xs text-muted-foreground">
-		<span class="inline-flex max-w-full items-center gap-2 rounded-full border bg-card px-3 py-2">
-			<span class="size-2 rounded-full bg-[#0dce97]"></span>
-			{m.methodology_legend_higher_better()}
-		</span>
-		<span class="inline-flex max-w-full items-center gap-2 rounded-full border bg-card px-3 py-2">
-			<span class="size-2 rounded-full bg-red-500"></span>
-			{m.methodology_legend_higher_worse()}
-		</span>
-		<span class="inline-flex max-w-full items-center rounded-full border bg-card px-3 py-2">
-			{m.methodology_legend_glyphs()}
-		</span>
-	</div>
 </div>
 
 {#snippet RichText(value: string)}
@@ -377,7 +312,12 @@
 {/snippet}
 
 {#snippet FormulaBlock(card: FormulaCard)}
-	<div class="mt-4 min-w-0 space-y-3">
+	<div
+		class={cn(
+			'mt-4 grid min-w-0 grid-cols-1 gap-3',
+			card.formulaBlocks.length > 1 && 'lg:grid-cols-2'
+		)}
+	>
 		{#each card.formulaBlocks as block (`${card.title}-${block.kind}`)}
 			<div
 				class="max-w-full overflow-x-auto rounded-xl border bg-muted/40 p-4"
@@ -398,7 +338,7 @@
 	title: Message,
 	cards: FormulaCard[],
 	children?: import('svelte').Snippet,
-	gridClass = 'md:grid-cols-2 xl:grid-cols-3'
+	gridClass = 'grid-cols-1'
 )}
 	<section class="mt-8 min-w-0" aria-label={title()}>
 		<h2 class="font-mono mb-6 font-semibold uppercase tracking-wider text-muted-foreground">
@@ -412,18 +352,6 @@
 						{@render FormulaBlock(card)}
 					{/if}
 					<p class="mt-4 text-sm leading-6 text-muted-foreground">{card.description()}</p>
-					{#if card.example}
-						<div class="mt-auto pt-6">
-							<p
-								class="min-w-0 rounded-xl border border-[#0dce97]/30 bg-[#0dce97]/10 px-3 py-2 font-mono text-xs leading-6 break-words text-[#0dce97]"
-							>
-								<span class="mb-1 block text-xs font-semibold uppercase tracking-wider">
-									{m.methodology_example_label()}
-								</span>
-								{@render RichText(card.example())}
-							</p>
-						</div>
-					{/if}
 				</article>
 			{/each}
 		</div>
@@ -433,11 +361,7 @@
 
 {#snippet MathFormula(kind: FormulaKind)}
 	{#if kind === 'dataset-score'}
-		<math
-			display="block"
-			class="min-w-max text-xl text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>S</mi><mi>ds</mi></msub>
 			<mo>=</mo>
 			<mi>H</mi><mo>(</mo><mn>1</mn><mo>−</mo><mi>FPR</mi><mo>,</mo><mn>1</mn><mo>−</mo><mi>FNR</mi
@@ -456,11 +380,7 @@
 			</mfrac>
 		</math>
 	{:else if kind === 'group-score'}
-		<math
-			display="block"
-			class="min-w-max text-xl text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>S</mi><mi>group</mi></msub>
 			<mo>=</mo>
 			<mi>H</mi><mo>(</mo><msub><mi>S</mi><msub><mi>ds</mi><mn>1</mn></msub></msub><mo>,</mo><mo
@@ -476,11 +396,7 @@
 			</mfrac>
 		</math>
 	{:else if kind === 'integral-product'}
-		<math
-			display="block"
-			class="min-w-max text-xl text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>S</mi><mi>integral</mi></msub>
 			<mo>=</mo>
 			<msup>
@@ -494,11 +410,7 @@
 			</msup>
 		</math>
 	{:else if kind === 'integral-log'}
-		<math
-			display="block"
-			class="min-w-max text-xl text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>S</mi><mi>integral</mi></msub>
 			<mo>=</mo>
 			<mi>exp</mi>
@@ -508,67 +420,8 @@
 			<mi>ln</mi><mo>(</mo><msub><mi>S</mi><msub><mi>group</mi><mi>j</mi></msub></msub><mo>)</mo>
 			<mo>)</mo>
 		</math>
-	{:else if kind === 'fbeta'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<msub><mi>F</mi><mi>β</mi></msub>
-			<mo>=</mo>
-			<mo>(</mo><mn>1</mn><mo>+</mo><msup><mi>β</mi><mn>2</mn></msup><mo>)</mo>
-			<mo>·</mo>
-			<mfrac>
-				<mrow><mi>P</mi><mo>·</mo><mi>R</mi></mrow>
-				<mrow><msup><mi>β</mi><mn>2</mn></msup><mo>·</mo><mi>P</mi><mo>+</mo><mi>R</mi></mrow>
-			</mfrac>
-		</math>
-	{:else if kind === 'f2'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<msub><mi>F</mi><mn>2</mn></msub>
-			<mo>=</mo>
-			<mn>5</mn><mo>·</mo>
-			<mfrac>
-				<mrow><mi>P</mi><mo>·</mo><mi>R</mi></mrow>
-				<mrow><mn>4</mn><mo>·</mo><mi>P</mi><mo>+</mo><mi>R</mi></mrow>
-			</mfrac>
-		</math>
-	{:else if kind === 'f05'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<msub><mi>F</mi><mn>0.5</mn></msub>
-			<mo>=</mo>
-			<mn>1.25</mn><mo>·</mo>
-			<mfrac>
-				<mrow><mi>P</mi><mo>·</mo><mi>R</mi></mrow>
-				<mrow><mn>0.25</mn><mo>·</mo><mi>P</mi><mo>+</mo><mi>R</mi></mrow>
-			</mfrac>
-		</math>
-	{:else if kind === 'fdr'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<mi>FDR</mi>
-			<mo>=</mo>
-			<mfrac><mi>FP</mi><mrow><mi>TP</mi><mo>+</mo><mi>FP</mi></mrow></mfrac>
-			<mo>=</mo>
-			<mn>1</mn><mo>−</mo><mi>Precision</mi>
-		</math>
 	{:else if kind === 'fnr-harm'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>FNR</mi><mi>harm</mi></msub>
 			<mo>=</mo>
 			<mfrac><mi>FN</mi><mrow><mi>FN</mi><mo>+</mo><mi>TP</mi></mrow></mfrac>
@@ -576,71 +429,15 @@
 			<mn>1</mn><mo>−</mo><msub><mi>Recall</mi><mi>harm</mi></msub>
 		</math>
 	{:else if kind === 'fpr-safe'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>FPR</mi><mi>safe</mi></msub>
 			<mo>=</mo>
 			<mfrac><mi>FP</mi><mrow><mi>FP</mi><mo>+</mo><mi>TN</mi></mrow></mfrac>
 			<mo>=</mo>
 			<mn>1</mn><mo>−</mo><msub><mi>Specificity</mi><mi>safe</mi></msub>
 		</math>
-	{:else if kind === 'micro-f1'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<msub><mi>F1</mi><mi>micro</mi></msub>
-			<mo>=</mo>
-			<mfrac>
-				<mrow
-					><mn>2</mn><mo>·</mo><msub><mi>P</mi><mi>micro</mi></msub><mo>·</mo><msub
-						><mi>R</mi><mi>micro</mi></msub
-					></mrow
-				>
-				<mrow
-					><msub><mi>P</mi><mi>micro</mi></msub><mo>+</mo><msub><mi>R</mi><mi>micro</mi></msub
-					></mrow
-				>
-			</mfrac>
-		</math>
-	{:else if kind === 'micro-pr'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
-			<msub><mi>P</mi><mi>micro</mi></msub>
-			<mo>=</mo>
-			<mfrac>
-				<mrow><mo>∑</mo><msub><mi>TP</mi><mi>i</mi></msub></mrow>
-				<mrow
-					><mo>∑</mo><mo>(</mo><msub><mi>TP</mi><mi>i</mi></msub><mo>+</mo><msub
-						><mi>FP</mi><mi>i</mi></msub
-					><mo>)</mo></mrow
-				>
-			</mfrac>
-			<mspace width="0.75em" />
-			<msub><mi>R</mi><mi>micro</mi></msub>
-			<mo>=</mo>
-			<mfrac>
-				<mrow><mo>∑</mo><msub><mi>TP</mi><mi>i</mi></msub></mrow>
-				<mrow
-					><mo>∑</mo><mo>(</mo><msub><mi>TP</mi><mi>i</mi></msub><mo>+</mo><msub
-						><mi>FN</mi><mi>i</mi></msub
-					><mo>)</mo></mrow
-				>
-			</mfrac>
-		</math>
 	{:else if kind === 'fpr-avg'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>FPR</mi><mi>avg</mi></msub>
 			<mo>=</mo>
 			<mfrac><mn>1</mn><mi>N</mi></mfrac>
@@ -648,11 +445,7 @@
 			<msub><mi>FPR</mi><mi>i</mi></msub>
 		</math>
 	{:else if kind === 'fnr-avg'}
-		<math
-			display="block"
-			class="min-w-max text-lg text-foreground"
-			xmlns="http://www.w3.org/1998/Math/MathML"
-		>
+		<math display="block" class={formulaClass(kind)} xmlns="http://www.w3.org/1998/Math/MathML">
 			<msub><mi>FNR</mi><mi>avg</mi></msub>
 			<mo>=</mo>
 			<mfrac><mn>1</mn><mi>N</mi></mfrac>

src/routes/tools/+page.svelte

@@ -5,7 +5,6 @@
 	import BucketDataWarning from '$lib/components/BucketDataWarning.svelte';
 	import PageHeadingSection from '$lib/components/PageHeadingSection.svelte';
 	import Select from '$lib/components/Select.svelte';
-	import Sheet from '$lib/components/Sheet.svelte';
 	import Table, { type TableColumn } from '$lib/components/Table.svelte';
 	import { Badge } from '$lib/components/ui/badge';
 	import { Button } from '$lib/components/ui/button';
@@ -14,16 +13,9 @@
 		toolsActiveGroups,
 		toolsBarLabelLines as barLabelLines,
 		toolsBarY as barY,
-		toolsBestModel as bestModel,
-		toolsCell as cellOf,
 		toolsCsvDataUri,
 		toolsDataGroups,
 		toolsHeatColor as heatColor,
-		toolsHeatmapMetricArrow as heatmapMetricArrow,
-		toolsHeatmapMetricItems as heatmapMetricItems,
-		toolsHeatmapMetricTone as heatmapMetricTone,
-		toolsHeatmapMetricValue as heatmapMetricValue,
-		toolsMetricValue as metricValue,
 		toolsModelColor,
 		toolsModelColorMap,
 		toolsModelKey as modelKey,
@@ -34,15 +26,12 @@
 		toolsRadarPoint as radarPoint,
 		toolsScatterX as scatterX,
 		toolsScatterY as scatterY,
-		toolsSelectedModels,
-		toolsValueTone as valueTone
+		toolsSelectedModels
 	} from '$lib/data/tools';
 	import { isNumber } from '$lib/helpers/leaderboard';
 	import type { GuardModel } from '$lib/types/leaderboard';
 	import type {
 		ToolsChartTooltip as ChartTooltip,
-		ToolsHeatmapMetric as HeatmapMetric,
-		ToolsHeatmapSelection as HeatmapSelection,
 		ToolsLanguageFilter as LangFilter,
 		ToolsLatencySortKey as LatencySortKey,
 		ToolsRadarScale as RadarScale,
@@ -78,6 +67,7 @@
 		p50: number | null;
 		p95: number | null;
 		p99: number | null;
+		errorRate: number | null;
 	};
 	type RobustnessRow = VisualizationModelRow & {
 		scoreReal: number | null;
@@ -89,6 +79,10 @@
 		category: string;
 	};
 	type HeatmapRow = ToolsDataset['visualizations']['heatmap']['rows'][number];
+	type HeatmapColorScale = {
+		min: number | null;
+		max: number | null;
+	};
 	const selectContentClass = 'max-h-72';
 	let { data }: PageProps = $props();
 	const toolsState = $derived(data.toolsState);
@@ -105,9 +99,7 @@
 	let language = $state<LangFilter>('all');
 	let metric = $state<VizMetric>('fnr');
 	let radarScale = $state<RadarScale>('score');
-	let pinnedModelKeys = $state<string[]>(
-		initialTools.models[0] ? [modelKey(initialTools.models[0])] : []
-	);
+	let pinnedModelKeys = $state<string[]>([]);
 	let latencySortKey = $state<LatencySortKey>('p95');
 	let latencySortDirection = $state<SortDirection>('asc');
 	let chartTooltip = $state<ChartTooltip | null>(null);
@@ -116,8 +108,6 @@
 	let hoveredScatterModelKey = $state<string | null>(null);
 	let hoveredBarKey = $state<string | null>(null);
 	let hoveredParetoModelKey = $state<string | null>(null);
-	let heatmapSheetOpen = $state(false);
-	let heatmapSelection = $state<HeatmapSelection | null>(null);
 
 	const languageOptions = [
 		{ value: 'all', label: m.tools_language_all() },
@@ -136,7 +126,8 @@
 		{ key: 'model', label: m.ranking_column_model },
 		{ key: 'p50', label: () => `p50 (${m.ranking_unit_ms()})`, cellClass: 'text-right' },
 		{ key: 'p95', label: () => `p95 (${m.ranking_unit_ms()})`, cellClass: 'text-right' },
-		{ key: 'p99', label: () => `p99 (${m.ranking_unit_ms()})`, cellClass: 'text-right' }
+		{ key: 'p99', label: () => `p99 (${m.ranking_unit_ms()})`, cellClass: 'text-right' },
+		{ key: 'errorRate', label: m.details_error_rate, cellClass: 'text-right' }
 	];
 
 	const currentLanguageLabel = $derived(
@@ -150,9 +141,6 @@
 	const selectedModelKeySet = $derived(new Set(selectedModelKeys));
 	const activeGroups = $derived(toolsActiveGroups(selectedGroups, language, tools.groupLanguages));
 	const pinnedModels = $derived(toolsPinnedModels(models, pinnedModelKeys));
-	const topIntegral = $derived(bestModel(selectedModels, 'integral', 'desc'));
-	const topLatency = $derived(bestModel(selectedModels, 'p95', 'asc'));
-	const topFnr = $derived(bestModel(selectedModels, 'fnr', 'asc'));
 	const heatmapMetric = $derived(tools.visualizations.heatmap.metric);
 	const heatmapModels = $derived(
 		tools.visualizations.heatmap.columnModelKeys.flatMap((key) => {
@@ -166,6 +154,7 @@
 				activeGroups.includes(row.groupKey) && (language === 'all' || row.language === language)
 		)
 	);
+	const heatmapColorScale = $derived.by(() => heatmapScaleFor(heatmapRows, heatmapModels));
 	const heatColumns = $derived(
 		`minmax(8rem, 10rem) repeat(${heatmapModels.length}, minmax(4rem, 1fr))`
 	);
@@ -209,6 +198,10 @@
 		return isNumber(value) ? value.toFixed(1) : '-';
 	}
 
+	function fmtPercent(value: number | null | undefined) {
+		return isNumber(value) ? `${(value * 100).toFixed(1)}%` : '-';
+	}
+
 	function localized(value: LocalizedString | null | undefined, fallback: string) {
 		return value?.[getLocale()] ?? value?.en ?? fallback;
 	}
@@ -282,7 +275,8 @@
 					integral: row.integral,
 					p50: row.latencyMs.p50,
 					p95: row.latencyMs.p95,
-					p99: row.latencyMs.p99
+					p99: row.latencyMs.p99,
+					errorRate: row.errorRate
 				}
 			];
 		});
@@ -401,25 +395,21 @@
 		return tools.visualizations.heatmap.values[row.id]?.[heatmapModelKey] ?? null;
 	}
 
-	function heatmapDrilldownFor(modelKeyValue: string, row: HeatmapRow) {
-		return tools.drilldownIndex.heatmap.find(
-			(item) => item.modelKey === modelKeyValue && item.groupKey === row.groupKey
+	function heatmapScaleFor(
+		rows: readonly HeatmapRow[],
+		visibleModels: readonly GuardModel[]
+	): HeatmapColorScale {
+		const values = rows.flatMap((row) =>
+			visibleModels
+				.map((model) => heatmapValue(row, modelKey(model)))
+				.filter((value): value is number => isNumber(value))
 		);
+		if (!values.length) return { min: null, max: null };
+		return { min: Math.min(...values), max: Math.max(...values) };
 	}
 
-	function fallbackHeatmapMetrics(value: number | null, key: HeatmapMetric) {
-		return {
-			score: key === 'score' ? value : null,
-			f1: key === 'f1' ? value : null,
-			fpr: key === 'fpr' ? value : null,
-			fnr: key === 'fnr' ? value : null
-		};
-	}
-
-	function heatColorValue(value: number | null | undefined, key: HeatmapMetric) {
-		if (!isNumber(value)) return value;
-		if (key === 'score' || key === 'f1') return 1 - value;
-		return value;
+	function heatmapScaleLabel(value: number | null) {
+		return isNumber(value) ? fmt(value, 3) : '-';
 	}
 
 	function colorOf(model: GuardModel) {
@@ -491,94 +481,6 @@
 		chartTooltip = null;
 	}
 
-	function scatterDrilldownGroup(model: GuardModel) {
-		const rows = activeGroups
-			.map((group) => {
-				const cell = cellOf(model, group);
-				const value = metricValue(cell, metric);
-				return cell && isNumber(value) ? { group, value } : null;
-			})
-			.filter((row): row is { group: string; value: number } => row !== null);
-
-		if (!rows.length) return null;
-
-		return rows.sort((a, b) => {
-			if (metric === 'f1') return a.value - b.value;
-			return b.value - a.value;
-		})[0].group;
-	}
-
-	function openScatterDrilldownSheet(model: GuardModel) {
-		const group = scatterDrilldownGroup(model);
-		if (!group) return;
-		const key = modelKey(model);
-		const rows = heatmapRows
-			.map((row) => ({ row, value: heatmapValue(row, key) }))
-			.filter(
-				(item): item is { row: HeatmapRow; value: number } =>
-					item.row.groupKey === group && isNumber(item.value)
-			);
-		const selected = rows.sort((a, b) => {
-			if (heatmapMetric === 'score' || heatmapMetric === 'f1') return a.value - b.value;
-			return b.value - a.value;
-		})[0];
-
-		if (!selected) return;
-		openHeatmapSheet(model, selected.row, selected.value);
-	}
-
-	function handleScatterPointKeydown(event: KeyboardEvent, model: GuardModel) {
-		if (event.key !== 'Enter' && event.key !== ' ') return;
-		event.preventDefault();
-		openScatterDrilldownSheet(model);
-	}
-
-	function openHeatmapSheet(model: GuardModel, row: HeatmapRow, value: number | null) {
-		if (!isNumber(value)) return;
-
-		const key = modelKey(model);
-		const drilldown = heatmapDrilldownFor(key, row);
-
-		hideTooltip();
-		heatmapSelection = {
-			model,
-			group: row.groupKey,
-			rowLabel: heatmapRowLabel(row),
-			metric: heatmapMetric,
-			value,
-			metrics: drilldown?.metrics ?? fallbackHeatmapMetrics(value, heatmapMetric)
-		};
-		heatmapSheetOpen = true;
-	}
-
-	function handleHeatmapCellKeydown(
-		event: KeyboardEvent,
-		model: GuardModel,
-		row: HeatmapRow,
-		value: number | null
-	) {
-		if (event.key !== 'Enter' && event.key !== ' ') return;
-		event.preventDefault();
-		openHeatmapSheet(model, row, value);
-	}
-
-	function tooltipToneClass(tone: TooltipRow['tone']) {
-		if (tone === 'good') return 'text-[#0dce97]';
-		if (tone === 'bad') return 'text-red-400';
-		return 'text-foreground';
-	}
-
-	function heatmapKpiClass(value: number | null | undefined, key: HeatmapMetric) {
-		const tone = heatmapMetricTone(value, key);
-		if (tone === 'good') {
-			return 'border-transparent bg-[#0dce97]/10 text-[#6bffd1]';
-		}
-		if (tone === 'bad') {
-			return 'border-transparent bg-red-500/10 text-red-300';
-		}
-		return 'border-transparent bg-transparent px-0 text-foreground';
-	}
-
 	function toggleLatencySort(key: LatencySortKey) {
 		if (latencySortKey === key) {
 			latencySortDirection = latencySortDirection === 'asc' ? 'desc' : 'asc';
@@ -599,19 +501,13 @@
 	}
 
 	function deltaBadgeClass(value: number | null, kind: 'score' | 'fnr') {
-		if (!isNumber(value) || Math.abs(value) <= 0.005) {
+		if (!isNumber(value) || value === 0) {
 			return 'bg-muted/40 text-muted-foreground';
 		}
 		const worse = kind === 'score' ? value < 0 : value > 0;
 		return worse ? 'bg-red-500/15 text-red-300' : 'bg-[#0dce97]/10 text-[#0dce97]';
 	}
 
-	function deltaArrow(value: number | null, kind: 'score' | 'fnr') {
-		if (!isNumber(value) || Math.abs(value) <= 0.005) return '–';
-		const worse = kind === 'score' ? value < 0 : value > 0;
-		return worse ? '▼' : '▲';
-	}
-
 	function fmtDelta(value: number | null) {
 		if (!isNumber(value)) return '-';
 		if (value > 0) return `+${value.toFixed(3)}`;
@@ -636,40 +532,8 @@
 		{/if}
 		<PageHeadingSection title={m.navigation_visualizations()} description={m.tools_description()} />
 
-		<div
-			class="mt-6 mb-5 flex items-center gap-3 overflow-x-auto rounded-xl border border-[#0dce97]/25 bg-[#0dce97]/10 px-4 py-3 text-sm whitespace-nowrap text-[#0dce97]"
-			role="status"
-			aria-live="polite"
-		>
-			{#if !selectedModels.length}
-				<span class="text-muted-foreground">{m.tools_takeaway_empty()}</span>
-			{:else if topIntegral}
-				<span>
-					<span class="font-semibold">{m.tools_takeaway_integral()}</span>
-					<b>{topIntegral.short}</b>
-					<code>{fmt(topIntegral.integral, 3)}</code>
-				</span>
-			{/if}
-			{#if selectedModels.length && topLatency}
-				<span class="text-muted-foreground">·</span>
-				<span>
-					<span class="font-semibold">{m.tools_takeaway_p95()}</span>
-					<b>{topLatency.short}</b>
-					<code>{Math.round(topLatency.p95)} {m.ranking_unit_ms()}</code>
-				</span>
-			{/if}
-			{#if selectedModels.length && topFnr}
-				<span class="text-muted-foreground">·</span>
-				<span>
-					<span class="font-semibold">{m.tools_takeaway_fnr()}</span>
-					<b>{topFnr.short}</b>
-					<code>{fmt(topFnr.fnr, 3)}</code>
-				</span>
-			{/if}
-		</div>
-
 		<section
-			class="mt-2 min-w-0 rounded-xl border bg-card p-4 shadow-sm"
+			class="mt-6 min-w-0 rounded-xl border bg-card p-4 shadow-sm"
 			aria-label={m.tools_controls_aria()}
 		>
 			<div class="grid min-w-0 gap-4 lg:grid-cols-[1.1fr_1.1fr_.7fr]">
@@ -863,8 +727,10 @@
 			</div>
 		</section>
 
-		<div class="mt-4 flex flex-wrap items-center gap-2 rounded-xl border bg-card px-4 py-3 text-xs">
-			<span class="font-semibold tracking-wider text-muted-foreground uppercase"
+		<div
+			class="mt-4 flex min-h-[46px] flex-wrap items-center gap-x-2 gap-y-3 rounded-xl border bg-card px-4 py-3 text-xs leading-none"
+		>
+			<span class="font-semibold tracking-wider text-muted-foreground uppercase leading-none"
 				>{m.tools_pinned_for_comparison()}</span
 			>
 			{#if pinnedModels.length}
@@ -882,13 +748,13 @@
 					</Badge>
 				{/each}
 			{:else}
-				<span class="text-muted-foreground">
+				<span class="leading-none text-muted-foreground">
 					{m.tools_no_pinned()}
 				</span>
 			{/if}
 		</div>
 
-		<div class="mt-6 grid min-w-0 gap-4 lg:grid-cols-2">
+		<div class="mt-4 grid min-w-0 gap-4 lg:grid-cols-2">
 			<section class="min-w-0 rounded-xl border bg-card p-4 shadow-sm">
 				<div class="mb-3 flex flex-wrap items-baseline gap-2">
 					<h2 class="font-mono text-sm font-semibold">{m.tools_radar_title()}</h2>
@@ -983,7 +849,7 @@
 											hoveredRadarModelKey = key;
 											updateTooltip(event, model.short, [
 												{ label: groupLabel(group), value: radarScale.toUpperCase() },
-												{ label: 'value', value: fmt(value, 3), tone: valueTone(value, true) }
+												{ label: 'value', value: fmt(value, 3) }
 											]);
 										}}
 										onpointermove={moveTooltip}
@@ -1115,17 +981,14 @@
 							fill-opacity="0.85"
 							stroke="currentColor"
 							stroke-width={isPinned ? 2 : 1}
-							class="cursor-pointer text-background transition-[r]"
-							role="button"
-							tabindex="0"
+							class="text-background transition-[r]"
+							role="img"
 							aria-label={`${row.model.short} FPR ${fmt(row.fpr, 3)} FNR ${fmt(row.fnr, 3)}`}
-							onclick={() => openScatterDrilldownSheet(row.model)}
-							onkeydown={(event) => handleScatterPointKeydown(event, row.model)}
 							onpointerenter={(event) => {
 								hoveredScatterModelKey = row.modelKey;
 								updateTooltip(event, row.model.short, [
-									{ label: 'FPR', value: fmt(row.fpr, 3), tone: valueTone(row.fpr, false) },
-									{ label: 'FNR', value: fmt(row.fnr, 3), tone: valueTone(row.fnr, false) }
+									{ label: 'FPR', value: fmt(row.fpr, 3) },
+									{ label: 'FNR', value: fmt(row.fnr, 3) }
 								]);
 							}}
 							onpointermove={moveTooltip}
@@ -1148,7 +1011,6 @@
 				</svg>
 				<div class="mt-3 flex flex-wrap gap-4 text-xs text-muted-foreground">
 					<span class="text-[#0dce97]">{m.tools_scatter_ideal()}</span>
-					<span>{m.tools_scatter_click()}</span>
 				</div>
 			</section>
 
@@ -1157,12 +1019,11 @@
 					<h2 class="font-mono text-sm font-semibold">
 						Heatmap <span class="font-mono uppercase">{heatmapMetric}</span>
 					</h2>
-					<span class="text-xs text-muted-foreground">{m.tools_heatmap_hint()}</span>
 					<span class="ml-auto inline-flex items-center gap-2 text-xs text-muted-foreground">
-						0.0 <span
-							class="h-2 w-28 rounded-full bg-[linear-gradient(90deg,#0dce97,#ffca28,#ff4d4f)]"
+						{m.tools_heatmap_scale_min({ value: heatmapScaleLabel(heatmapColorScale.min) })}
+						<span class="h-2 w-28 rounded-full bg-[linear-gradient(90deg,#0dce97,#ffca28,#ff4d4f)]"
 						></span>
-						{m.tools_heatmap_scale()}
+						{m.tools_heatmap_scale_max({ value: heatmapScaleLabel(heatmapColorScale.max) })}
 					</span>
 				</div>
 				<div class="max-w-full overflow-auto">
@@ -1194,41 +1055,40 @@
 								{@const key = modelKey(model)}
 								{@const value = heatmapValue(row, key)}
 								{@const hasValue = isNumber(value)}
-								<button
-									type="button"
-									disabled={!hasValue}
+								<div
+									role="img"
 									class={cn(
-										'w-full rounded-md border-0 px-2 py-2 text-center font-mono text-sm font-semibold transition duration-150 disabled:cursor-default disabled:opacity-100',
-										hasValue &&
-											'cursor-pointer hover:brightness-110 hover:saturate-125 focus-visible:scale-[1.02] focus-visible:brightness-110 focus-visible:saturate-125 focus-visible:outline-none',
+										'w-full rounded-md px-2 py-2 text-center font-mono text-sm font-semibold',
+										hasValue && 'transition duration-150 hover:brightness-110 hover:saturate-125',
 										!isNumber(value) && 'bg-muted/70 text-muted-foreground',
 										isNumber(value) && heatTextColor(value)
 									)}
 									style={isNumber(value)
-										? `background:${heatColor(heatColorValue(value, heatmapMetric))}`
+										? `background:${heatColor(
+												value,
+												heatmapColorScale.min ?? value,
+												heatmapColorScale.max ?? value,
+												'sqrt'
+											)}`
 										: ''}
 									aria-label={`${model.short} ${heatmapRowLabel(row)} ${heatmapMetric.toUpperCase()} ${fmt(value, 3)}`}
-									onclick={() => openHeatmapSheet(model, row, value)}
-									onkeydown={(event) => handleHeatmapCellKeydown(event, model, row, value)}
 									onpointerenter={(event) =>
 										updateTooltip(event, model.short, [
 											{ label: groupLabel(row.groupKey), value: heatmapRowLabel(row) },
 											{
 												label: heatmapMetric.toUpperCase(),
-												value: fmt(value, 3),
-												tone: valueTone(value, heatmapMetric === 'f1' || heatmapMetric === 'score')
+												value: fmt(value, 3)
 											}
 										])}
 									onpointermove={moveTooltip}
 									onpointerleave={hideTooltip}
 								>
 									{isNumber(value) ? fmt(value, 2) : '-'}
-								</button>
+								</div>
 							{/each}
 						{/each}
 					</div>
 				</div>
-				<p class="mt-3 text-xs text-muted-foreground">{m.tools_heatmap_note()}</p>
 			</section>
 
 			<section class="min-w-0 rounded-xl border bg-card p-4 shadow-sm lg:col-span-2">
@@ -1292,8 +1152,7 @@
 											updateTooltip(event, groupLabel(group), [
 												{
 													label: series.key.toUpperCase(),
-													value: fmt(value, 3),
-													tone: valueTone(value, series.key === 'f1')
+													value: fmt(value, 3)
 												},
 												{
 													label: m.tools_mean(),
@@ -1437,13 +1296,9 @@
 								onpointerenter={(event) => {
 									hoveredParetoModelKey = row.modelKey;
 									updateTooltip(event, row.model.short, [
-										{
-											label: 'integral',
-											value: fmt(row.integral, 3),
-											tone: valueTone(row.integral, true)
-										},
+										{ label: 'integral', value: fmt(row.integral, 3) },
 										{ label: 'p95', value: `${fmtMs(row.p95)} ${m.ranking_unit_ms()}` },
-										{ label: 'FPR', value: fmt(row.fpr, 3), tone: valueTone(row.fpr, false) }
+										{ label: 'FPR', value: fmt(row.fpr, 3) }
 									]);
 								}}
 								onpointermove={moveTooltip}
@@ -1519,7 +1374,7 @@
 					columns={latencyColumns}
 					getRowKey={(row) => row.modelKey}
 					containerClass="max-w-full overflow-x-auto rounded-lg border"
-					tableClass="min-w-[42rem]"
+					tableClass="min-w-[50rem]"
 					headerClass="bg-muted/70 hover:bg-muted/70 [&_th]:text-muted-foreground"
 					rowClass="hover:bg-muted/30"
 				>
@@ -1545,6 +1400,8 @@
 								<span class="min-w-0 max-w-80 truncate">{row.model.short}</span>
 								<code class="text-muted-foreground">{row.model.params}</code>
 							</span>
+						{:else if column.key === 'errorRate'}
+							<span class="font-mono">{fmtPercent(row.errorRate)}</span>
 						{:else}
 							<span class="font-mono">{fmtMs(row[column.key])}</span>
 						{/if}
@@ -1560,6 +1417,7 @@
 					</h2>
 					<span class="text-xs text-muted-foreground">{m.tools_robustness_hint()}</span>
 				</div>
+				<p class="mb-3 text-xs leading-5 text-muted-foreground">{m.tools_robustness_note()}</p>
 				<div class="max-w-full overflow-x-auto rounded-lg border">
 					<table class="w-full min-w-[54rem] text-sm">
 						<thead class="bg-muted/70 text-xs text-muted-foreground">
@@ -1614,7 +1472,6 @@
 												deltaBadgeClass(sDelta, 'score')
 											)}
 										>
-											{deltaArrow(sDelta, 'score')}
 											{fmtDelta(sDelta)}
 										</span>
 									</td>
@@ -1627,7 +1484,6 @@
 												deltaBadgeClass(fDelta, 'fnr')
 											)}
 										>
-											{deltaArrow(fDelta, 'fnr')}
 											{fmtDelta(fDelta)}
 										</span>
 									</td>
@@ -1636,7 +1492,6 @@
 						</tbody>
 					</table>
 				</div>
-				<p class="mt-3 text-xs text-muted-foreground">{m.tools_robustness_note()}</p>
 			</section>
 		</div>
 
@@ -1644,82 +1499,6 @@
 			{m.tools_footer()}
 		</p>
 
-		<Sheet
-			bind:open={heatmapSheetOpen}
-			titleText={heatmapSelection ? heatmapSelection.rowLabel : ''}
-			contentClass="lg:data-[side=right]:w-[44rem]!"
-			titleClass="text-2xl"
-		>
-			{#if heatmapSelection}
-				{@const meta = benchmarkMeta[heatmapSelection.group]}
-				<div class="space-y-6 p-4 sm:p-5">
-					<div class="flex flex-wrap items-center gap-3">
-						<Badge variant="secondary" class="gap-2 rounded-full px-3 py-1">
-							<span
-								class="size-2.5 rounded-sm"
-								style={`background:${colorOf(heatmapSelection.model)}`}
-							></span>
-							{heatmapSelection.model.short}
-						</Badge>
-						{#if meta?.languages.length}
-							<Badge variant="secondary" class="rounded-full px-3 py-1 font-mono uppercase">
-								{meta.languages.map((lang) => lang.toUpperCase()).join(' · ')}
-							</Badge>
-						{/if}
-						<Badge variant="outline" class="rounded-full px-3 py-1 font-mono">
-							{groupLabel(heatmapSelection.group)}
-						</Badge>
-					</div>
-
-					<div class="grid gap-2 sm:grid-cols-2 xl:grid-cols-4">
-						{#each heatmapMetricItems(heatmapSelection) as item (item.label)}
-							<div class="rounded-lg border bg-card p-3 shadow-sm">
-								<div
-									class={cn(
-										'inline-flex items-center gap-1.5 rounded-md border px-2.5 py-1 font-mono text-xl font-semibold tabular-nums',
-										heatmapKpiClass(item.value, item.key)
-									)}
-								>
-									{heatmapMetricArrow(item.value, item.key)}
-									{fmt(item.value, 3)}
-								</div>
-								<div
-									class="mt-2 text-[11px] font-semibold tracking-wider text-muted-foreground uppercase"
-								>
-									{item.label}
-								</div>
-							</div>
-						{/each}
-					</div>
-
-					{#if meta?.whatItTests}
-						<p class="rounded-xl border bg-muted/30 p-4 text-sm leading-6 text-muted-foreground">
-							{localized(meta.whatItTests, '')}
-						</p>
-					{/if}
-
-					<section class="border-t pt-5">
-						<h3 class="font-mono text-base font-semibold">
-							{m.tools_heatmap_sheet_prompts_title()}
-						</h3>
-						<p class="mt-3 text-sm leading-6 text-muted-foreground">
-							{m.tools_heatmap_sheet_prompts_description({
-								metric: heatmapSelection.metric.toUpperCase(),
-								value: fmt(
-									heatmapMetricValue(heatmapSelection.metrics, heatmapSelection.metric) ??
-										heatmapSelection.value,
-									3
-								)
-							})}
-						</p>
-						<ul class="mt-4 list-disc space-y-1 pl-6 font-mono text-sm text-muted-foreground">
-							<li>{heatmapSelection.rowLabel}</li>
-						</ul>
-					</section>
-				</div>
-			{/if}
-		</Sheet>
-
 		{#if chartTooltip}
 			<div
 				bind:this={chartTooltipElement}
@@ -1731,7 +1510,7 @@
 					{#each chartTooltip.rows as row (`${row.label}-${row.value}`)}
 						<div class="grid grid-cols-[auto_1fr] gap-5">
 							<span class="break-words text-muted-foreground">{row.label}</span>
-							<span class={cn('text-right', tooltipToneClass(row.tone))}>{row.value}</span>
+							<span class="text-right">{row.value}</span>
 						</div>
 					{/each}
 				</div>