The Daily Markdown

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'objective_reasoning': 0.5, 'human_preference': 0.3, 'terminal_work': 0.2}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'objective_reasoning': 0.5, 'human_preference': 0.3, 'terminal_work': 0.2}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'executed_edits': 0.35, 'terminal_work': 0.35, 'objective_coding': 0.2, 'human_preference': 0.1}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'routine_tasks': 0.6, 'objective_coding': 0.25, 'human_preference': 0.15}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'executed_edits': 0.35, 'terminal_work': 0.35, 'objective_coding': 0.2, 'human_preference': 0.1}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'planner': {'objective_reasoning': 0.5, 'human_preference': 0.3, 'terminal_work': 0.2}, 'mid': {'executed_edits': 0.35, 'terminal_work': 0.35, 'objective_coding': 0.2, 'human_preference': 0.1}, 'rote': {'routine_tasks': 0.6, 'objective_coding': 0.25, 'human_preference': 0.15}}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'planner': {'objective_reasoning': 0.5, 'human_preference': 0.3, 'terminal_work': 0.2}, 'mid': {'executed_edits': 0.35, 'terminal_work': 0.35, 'objective_coding': 0.2, 'human_preference': 0.1}, 'rote': {'routine_tasks': 0.6, 'objective_coding': 0.25, 'human_preference': 0.15}}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only.

R = weighted mean over present families: sum(preset family weight × mean calibrated score within that family) / sum(present family weights). Role weights: {'objective_reasoning': 0.5, 'human_preference': 0.3, 'terminal_work': 0.2}. Best published configuration: use the highest raw score per source and model, never average efforts. Confirmed evaluator spelling aliases fold through an explicit table (Gemma 4 31B IT/instruct); arbitrary instruct and checkpoint suffixes stay distinct. Explicit effort suffixes fold into the base model. Full date aliases fold only when that source has one dated release for the base; checkpoint IDs, versions, tiers and parameter sizes stay distinct. Non-Arena sources use z-scores and one positive linear fit on at least 10 model-key overlaps with Arena text, frozen in this snapshot. Failed fits remain visible but contribute nothing. Models with just one contributing independent evaluator appear under 'New, not enough ratings yet', without a rank. Other models are ranked by their blended score; models with fewer than 2 rating families, 2 independent evaluators or 50% of the available weight are tagged partial. Available families have at least one calibrated, non-excluded source in this snapshot. Coverage weight remains the share of preset weight present. Boards warn when less than 50% of preset weight is available and name missing families. Terminal results retain the named agent configuration and serve as workflow proxies. Scores are Arena-scaled proxies; no pairwise win probability. Price only filters eligibility. PirateBench is our build check only. Fit: total checkpoint parameters × 0.5 × 1.25 GB, <= 80% of RAM/VRAM; all experts count.