restructure + add files

data/processed/RQ3/Violin/{write_a_book_with_flows-A2A_total_tokens_violin.pdf → BookWriter-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:921a84c06e55889f8a73ffc60fb74ba3b23aa10022596a1d1963232adc47c7de
+oid sha256:f2261cbfcbc6b44b06e5ab875757b738772d763bb3c19c9ad891cfa7e9acbc50
 size 30168

data/processed/RQ3/Violin/{write_a_book_with_flows-A2A_mix_total_tokens_violin.pdf → BookWriter-H-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8cf67830d8e85e6bcc0d7a057385758f45c7c9355fa978d7cc1e32d3b84931a1
+oid sha256:6ca372d3bc97df05dc167e83907182026fdd772c8c95dd80cfaef804c68debd2
 size 28331

data/processed/RQ3/Violin/{write_a_book_with_flows-MCP_total_tokens_violin.pdf → BookWriter-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa301b135f1f46b17e27dde1ebcc0c2d0fde9fef545d3eca3612217c0b6057d5
+oid sha256:236697ecf89bce49b9421aac8a9437d319d6e3c37d3c4445fc891cde7d285d5f
 size 30238

data/processed/RQ3/Violin/{email_auto_responder_flow-MCP_total_tokens_violin.pdf → EmailResponder-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e4040c3c585d42d668220510ed4ace950bcfb2f2028519d65c63a2a21076f35b
+oid sha256:e5c5823df3d8cd943cedd71101c0ddbe4c51254ccd87a64f345ae02a8c8959cb
 size 33833

data/processed/RQ3/Violin/{email_auto_responder_flow_total_tokens_violin.pdf → EmailResponder_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b0b3d95d2584aaabe75ce28a34f50e3937aaeb6ad6267aff6203c70d63d227be
+oid sha256:ba5c4fbde574472e01fdd3c4a1a6997d799cc9af7beb714ec53f2c882fd00477
 size 34591

data/processed/RQ3/Violin/{game_builder-MCP_total_tokens_violin.pdf → GameBuilder-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c10c5d39bda006b2a0dd747969e2dd25a3038ef6687b993b2d52fc26bffc302f
+oid sha256:9b697298ee3776722a0fdbe2631d0f57fc092419432aab1d6de57ec22f592f39
 size 27515

data/processed/RQ3/Violin/{game_builder_total_tokens_violin.pdf → GameBuilder_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a39458209d12f197f47d327b6715982900ed9a8d135f71e1628636ba14240268
+oid sha256:4e7aee638635029fc9509b4ed2a8210d14abc0fc8f41447b62a2178935074978
 size 29150

data/processed/RQ3/Violin/{landing_page_generator-A2A_total_tokens_violin.pdf → LandingPageGenerator-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:39e9bbd84eb049866e9fa48d940f20f65dd7a5f8217e87e313fc6a5bbacf7067
+oid sha256:e54288adb64b2c4b332370e2bc373041357d321b36257ceabf488252299a45c9
 size 35619

data/processed/RQ3/Violin/{landing_page_generator-A2A_mix_total_tokens_violin.pdf → LandingPageGenerator-H-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1bdac78f49c2e294084b8546746958989235bdc4a06c105654a618388cb2b6fa
+oid sha256:5cbcb03edbfb21b2e5ae885f20d45db15162e8f1290c34d82ec9fe415ca4af4d
 size 35800

data/processed/RQ3/Violin/{landing_page_generator-MCP_total_tokens_violin.pdf → LandingPageGenerator-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f998ba58a299ac49844c84fe663a988929e1b94428e426c49ebe41f8b6ec057
+oid sha256:839f0b11c30b5617356abe8b05e6fade6e5ddf41c353bfc365c3dab270a38b15
 size 35899

data/processed/RQ3/Violin/{markdown_validator-MCP_total_tokens_violin.pdf → MarkdownValidator-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7bead0f0de8074263cde20783bc25e5fa08bd17a1d5bb81d13fc1a38ed9e7c00
+oid sha256:9b6dba329e741ae4bdbf589e592b8736db0a18df699199adfcea621921e678b7
 size 30488

data/processed/RQ3/Violin/{markdown_validator_total_tokens_violin.pdf → MarkdownValidator_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad58e6a12fd6fa4f358df6a565f7e654e90eaa99baf737a663e8004c00a740a0
+oid sha256:a34c7426022464adf97a96b105f0ba212818fcd2f14a12856b759654765593ae
 size 30638

data/processed/RQ3/Violin/{intelligent_recruitment_platform-A2A_total_tokens_violin.pdf → RecruitmentAssistant-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8eb80d9de4acbd6409a20f207ce5ad2f082bd54b3c61cadc5faf7132b5f5f8e2
+oid sha256:8c2709b2d428d7392d7148fdb96058d47dd7bd60e396eb150683f27537d56b09
 size 32400

data/processed/RQ3/Violin/{intelligent_recruitment_platform-A2A_mix_total_tokens_violin.pdf → RecruitmentAssistant-H-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8ae22bfa81a2d30992a028abc9ebe5923c025e80282469e9af5ffeaa0d6997de
+oid sha256:7cc8f44dacb9827d8ac05176d46866e1a12476faadc942b1681096d0d30df7c8
 size 32196

data/processed/RQ3/Violin/{intelligent_recruitment_platform-MCP_total_tokens_violin.pdf → RecruitmentAssistant-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b5489bae6fb5dfc4a42087586b95deae8f7ba693d1dbb558c8ef06f1b619ff5
+oid sha256:2d761d9ab4ce7da8f073fc00c98d5699b3fab0575bb48eb5d4b409d720e3746e
 size 33330

data/processed/RQ3/Violin/{SQL_assistant-A2A_total_tokens_violin.pdf → SQLAssistant-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7368dfb27c83a38dd0f1e18178c5b0c397d0a37c1b0985173cc86f488b280afb
+oid sha256:703eadbc44de39db6efeabf7900db8103ac87f6f49208a8e962abd7844208bf7
 size 42192

data/processed/RQ3/Violin/{SQL_assistant-A2A_mix_total_tokens_violin.pdf → SQLAssistant-H-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e2ef4ee6fd871b76950bf50acbdb1d39ee8bce23486e4b313d5eb2dcc7f0d4b9
+oid sha256:fbad89c984c14e336258f3ebaee047f707e20555e2475b06a096e81ce103a2d1
 size 42992

data/processed/RQ3/Violin/{SQL_assistant-MCP_total_tokens_violin.pdf → SQLAssistant-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10c0be8f3528dfbfabf238b58c73bca8670dcc43255909a806b2a8d131ff51e9
+oid sha256:24c99a975197195f5bb47b53569a82653d862a749c5d102e55614375b97e0abd
 size 43109

data/processed/RQ3/Violin/{self_evaluation_loop_flow-A2A_total_tokens_violin.pdf → SocialMediaManager-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12daa50b440fd797def8f9634e252b308f7834f25ac54af0a9270a79db5e107e
+oid sha256:6b3a74d1c6cdb58c37689e33b2a1a442387d6ea48ed9c6d8f1c2623eed143490
 size 45169

data/processed/RQ3/Violin/{self_evaluation_loop_flow-A2A_mix_total_tokens_violin.pdf → SocialMediaManager-H-A2A_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0dd692f8db4a33e77f7d929efb0ff1a36cca550e7d12dd758923d2f9f2ad5eb
+oid sha256:657da0c544f836b46abcfd17e894a582b911764da0c5079ecf6c4f07a2708029
 size 41479

data/processed/RQ3/Violin/{self_evaluation_loop_flow-MCP_total_tokens_violin.pdf → SocialMediaManager-MCP_total_tokens_violin.pdf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2687f6480d63d6f4a9ed3a82cd9cf602e02e378b43f9201ef5e1f472f2a2e1ec
+oid sha256:4815259199a2e5179fe614b6d8f9cc1fb180ad0961a9b540cabeb4a74911d6ae
 size 44043

data/processed/RQ3/Violin/architecture_token_deltas.md

@@ -1,166 +0,0 @@
-# Token shifts across architectures
-
-Two series are reported per task (when data is available):
-- **Pure CrewAI → MCP** (architecture labels: Unknown → MCP)
-- **MCP → A2A → A2A_mix**
-Each series shows the absolute change (Δ) and the relative percentage change of the mean total tokens.
-
-## SQL_assistant
-
-### MCP → A2A → A2A_mix
-
-| Model | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 13,881 | 11,189 | -2,692 (-19.4%) | 8,956 | -2,233 (-20.0%) |
-| GPT-4o-mini | 24,241 | 18,635 | -5,606 (-23.1%) | 13,734 | -4,901 (-26.3%) |
-| DeepSeek-V3.1 | 25,637 | 36,805 | +11,169 (43.6%) | 16,479 | -20,326 (-55.2%) |
-| DeepSeek-R1 | 80,522 | 96,469 | +15,947 (19.8%) | 69,364 | -27,105 (-28.1%) |
-| Gemini-2.5 | 19,110 | 32,531 | +13,422 (70.2%) | 24,448 | -8,083 (-24.8%) |
-| Gemini-2.5-NT | 15,014 | 25,782 | +10,769 (71.7%) | 18,476 | -7,306 (-28.3%) |
-| Qwen3-235b | 17,232 | 17,381 | +149 (0.9%) | 14,779 | -2,602 (-15.0%) |
-
-Project-level average (all models combined)
-
-| Metric | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| Avg tokens (all models) | 27,948 | 34,113 | +6,165 (22.1%) | 23,748 | -10,365 (-30.4%) |
-
-## email_auto_responder_flow
-
-### Pure CrewAI → MCP
-
-| Model | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| GPT-5 | 5,562 | 6,903 | +1,342 (24.1%) |
-| GPT-4o-mini | 8,044 | 9,437 | +1,393 (17.3%) |
-| DeepSeek-V3.1 | 7,431 | 9,580 | +2,149 (28.9%) |
-| DeepSeek-R1 | 16,982 | 15,403 | -1,579 (-9.3%) |
-| Gemini-2.5 | 35,189 | 38,066 | +2,877 (8.2%) |
-| Gemini-2.5-NT | 42,421 | 35,133 | -7,287 (-17.2%) |
-| Qwen3-235b | 5,900 | 7,434 | +1,533 (26.0%) |
-
-Project-level average (all models combined)
-
-| Metric | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| Avg tokens (all models) | 17,361 | 17,422 | +61 (0.4%) |
-
-## game_builder
-
-### Pure CrewAI → MCP
-
-| Model | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| GPT-5 | 22,904 | 22,425 | -479 (-2.1%) |
-| GPT-4o-mini | 17,823 | 47,996 | +30,173 (169.3%) |
-| DeepSeek-V3.1 | 41,341 | 55,289 | +13,947 (33.7%) |
-| DeepSeek-R1 | 44,536 | 44,851 | +316 (0.7%) |
-| Gemini-2.5 | 72,772 | 75,519 | +2,747 (3.8%) |
-| Gemini-2.5-NT | 67,964 | 50,782 | -17,181 (-25.3%) |
-| Qwen3-235b | 24,212 | 24,039 | -173 (-0.7%) |
-
-Project-level average (all models combined)
-
-| Metric | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| Avg tokens (all models) | 41,650 | 45,843 | +4,193 (10.1%) |
-
-## intelligent_recruitment_platform
-
-### MCP → A2A → A2A_mix
-
-| Model | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 32,411 | 31,991 | -419 (-1.3%) | 32,672 | +681 (2.1%) |
-| GPT-4o-mini | 99,933 | 74,951 | -24,982 (-25.0%) | 101,936 | +26,985 (36.0%) |
-| DeepSeek-V3.1 | 65,748 | 65,662 | -85 (-0.1%) | 65,048 | -614 (-0.9%) |
-| DeepSeek-R1 | 84,686 | 72,968 | -11,718 (-13.8%) | 79,028 | +6,060 (8.3%) |
-| Gemini-2.5 | 64,864 | 68,719 | +3,855 (5.9%) | 57,128 | -11,591 (-16.9%) |
-| Gemini-2.5-NT | 54,402 | 55,660 | +1,257 (2.3%) | 36,887 | -18,772 (-33.7%) |
-| Qwen3-235b | 87,098 | 97,717 | +10,620 (12.2%) | 73,259 | -24,458 (-25.0%) |
-
-Project-level average (all models combined)
-
-| Metric | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| Avg tokens (all models) | 69,877 | 66,810 | -3,068 (-4.4%) | 63,708 | -3,101 (-4.6%) |
-
-## landing_page_generator
-
-### MCP → A2A → A2A_mix
-
-| Model | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 12,653 | 14,965 | +2,312 (18.3%) | 25,296 | +10,331 (69.0%) |
-| GPT-4o-mini | 19,250 | 20,340 | +1,090 (5.7%) | 43,589 | +23,249 (114.3%) |
-| DeepSeek-V3.1 | 15,866 | 28,814 | +12,948 (81.6%) | 34,496 | +5,682 (19.7%) |
-| DeepSeek-R1 | 28,371 | 28,657 | +286 (1.0%) | 57,203 | +28,546 (99.6%) |
-| Gemini-2.5 | 26,527 | 21,256 | -5,271 (-19.9%) | 45,250 | +23,995 (112.9%) |
-| Gemini-2.5-NT | 33,428 | 22,826 | -10,602 (-31.7%) | 14,949 | -7,876 (-34.5%) |
-| Qwen3-235b | 14,478 | 18,142 | +3,664 (25.3%) | 17,628 | -514 (-2.8%) |
-
-Project-level average (all models combined)
-
-| Metric | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| Avg tokens (all models) | 21,510 | 22,143 | +632 (2.9%) | 34,059 | +11,916 (53.8%) |
-
-## markdown_validator
-
-### Pure CrewAI → MCP
-
-| Model | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| GPT-5 | 1,853 | 1,967 | +115 (6.2%) |
-| GPT-4o-mini | 1,853 | 1,942 | +88 (4.8%) |
-| DeepSeek-V3.1 | 1,948 | 2,063 | +115 (5.9%) |
-| DeepSeek-R1 | 5,912 | 5,885 | -26 (-0.4%) |
-| Gemini-2.5 | 135,224 | 95,503 | -39,722 (-29.4%) |
-| Gemini-2.5-NT | 31,330 | 61,899 | +30,569 (97.6%) |
-| Qwen3-235b | 1,854 | 1,994 | +140 (7.6%) |
-
-Project-level average (all models combined)
-
-| Metric | CrewAI (Unknown) | MCP | Δ MCP-CrewAI (Unknown) |
-| --- | --- | --- | --- |
-| Avg tokens (all models) | 25,711 | 24,465 | -1,246 (-4.8%) |
-
-## self_evaluation_loop_flow
-
-### MCP → A2A → A2A_mix
-
-| Model | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 38,940 | 40,570 | +1,630 (4.2%) | 16,796 | -23,773 (-58.6%) |
-| GPT-4o-mini | 108,089 | 97,814 | -10,275 (-9.5%) | 43,322 | -54,492 (-55.7%) |
-| DeepSeek-V3.1 | 95,721 | 87,367 | -8,355 (-8.7%) | 57,449 | -29,918 (-34.2%) |
-| DeepSeek-R1 | 109,146 | 124,960 | +15,814 (14.5%) | 98,225 | -26,735 (-21.4%) |
-| Gemini-2.5 | 12,749 | 12,955 | +207 (1.6%) | 20,423 | +7,468 (57.6%) |
-| Gemini-2.5-NT | 11,480 | 11,429 | -51 (-0.4%) | 9,642 | -1,787 (-15.6%) |
-| Qwen3-235b | 46,557 | 55,866 | +9,308 (20.0%) | 29,440 | -26,426 (-47.3%) |
-
-Project-level average (all models combined)
-
-| Metric | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| Avg tokens (all models) | 60,383 | 61,566 | +1,183 (2.0%) | 39,328 | -22,237 (-36.1%) |
-
-## write_a_book_with_flows
-
-### MCP → A2A → A2A_mix
-
-| Model | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 190,842 | 181,022 | -9,820 (-5.1%) | 117,358 | -63,664 (-35.2%) |
-| GPT-4o-mini | 133,339 | 139,665 | +6,326 (4.7%) | 102,564 | -37,101 (-26.6%) |
-| DeepSeek-V3.1 | 184,315 | 272,275 | +87,960 (47.7%) | 141,204 | -131,071 (-48.1%) |
-| DeepSeek-R1 | 190,705 | 186,678 | -4,027 (-2.1%) | 129,883 | -56,795 (-30.4%) |
-| Gemini-2.5 | 107,154 | 110,126 | +2,973 (2.8%) | 91,585 | -18,541 (-16.8%) |
-| Gemini-2.5-NT | 82,546 | 86,249 | +3,703 (4.5%) | 65,826 | -20,422 (-23.7%) |
-| Qwen3-235b | 110,716 | 118,406 | +7,690 (6.9%) | 76,512 | -41,893 (-35.4%) |
-
-Project-level average (all models combined)
-
-| Metric | MCP | A2A | Δ A2A-MCP | A2A_mix | Δ A2A_mix-A2A |
-| --- | --- | --- | --- | --- | --- |
-| Avg tokens (all models) | 142,802 | 156,346 | +13,543 (9.5%) | 103,562 | -52,784 (-33.8%) |

data/processed/RQ3/Violin/project_model_distribution.md

@@ -2,6 +2,580 @@
 
 Per-project, per-model total token usage with breakdown by execution outcome. Only the mean total tokens are reported. Baseline vs maximum is computed from the overall mean aggregated across all three statuses.
 
+## TOTAL AGGREGATION (All Projects)
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 1270 | 38,411 |
+| GPT-4o-mini | 1270 | 52,051 |
+| DeepSeek-V3.1 | 1270 | 59,788 |
+| DeepSeek-R1 | 1270 | 72,750 |
+| Gemini-2.5 | 1270 | 53,359 |
+| Gemini-2.5-NT | 1270 | 38,578 |
+| Qwen3-235b | 1270 | 39,751 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 1051 | 40,620 |
+| GPT-4o-mini | 1116 | 47,293 |
+| DeepSeek-V3.1 | 1098 | 59,360 |
+| DeepSeek-R1 | 826 | 60,570 |
+| Gemini-2.5 | 862 | 54,661 |
+| Gemini-2.5-NT | 896 | 36,901 |
+| Qwen3-235b | 1000 | 35,776 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 137 | 32,233 |
+| GPT-4o-mini | 144 | 81,392 |
+| DeepSeek-V3.1 | 165 | 64,056 |
+| DeepSeek-R1 | 308 | 114,919 |
+| Gemini-2.5 | 138 | 92,688 |
+| Gemini-2.5-NT | 91 | 111,927 |
+| Qwen3-235b | 196 | 66,003 |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 82 | 20,410 |
+| GPT-4o-mini | 10 | 160,542 |
+| DeepSeek-V3.1 | 7 | 26,352 |
+| DeepSeek-R1 | 136 | 51,229 |
+| Gemini-2.5 | 270 | 29,101 |
+| Gemini-2.5-NT | 283 | 20,300 |
+| Qwen3-235b | 74 | 23,934 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-5 = 38,411 tokens
+- Maximum (highest mean): DeepSeek-R1 = 72,750 tokens
+- Delta: 34,340 (89.4%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): Qwen3-235b = 35,776 tokens
+- Maximum (highest mean): DeepSeek-R1 = 60,570 tokens
+- Delta: 24,794 (69.3%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): GPT-5 = 32,233 tokens
+- Maximum (highest mean): DeepSeek-R1 = 114,919 tokens
+- Delta: 82,686 (256.5%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): Gemini-2.5-NT = 20,300 tokens
+- Maximum (highest mean): GPT-4o-mini = 160,542 tokens
+- Delta: 140,242 (690.8%)
+
+---
+# Series Aggregates
+
+## Series: BookWriter
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 163,074 |
+| GPT-4o-mini | 180 | 125,190 |
+| DeepSeek-V3.1 | 180 | 199,265 |
+| DeepSeek-R1 | 180 | 169,089 |
+| Gemini-2.5 | 180 | 102,955 |
+| Gemini-2.5-NT | 180 | 78,207 |
+| Qwen3-235b | 180 | 101,878 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 163,074 |
+| GPT-4o-mini | 180 | 125,190 |
+| DeepSeek-V3.1 | 179 | 198,154 |
+| DeepSeek-R1 | 163 | 156,596 |
+| Gemini-2.5 | 180 | 102,955 |
+| Gemini-2.5-NT | 179 | 78,117 |
+| Qwen3-235b | 179 | 101,880 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 1 | 398,139 |
+| DeepSeek-R1 | 17 | 288,873 |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 1 | 94,226 |
+| Qwen3-235b | 1 | 101,570 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): Gemini-2.5-NT = 78,207 tokens
+- Maximum (highest mean): DeepSeek-V3.1 = 199,265 tokens
+- Delta: 121,058 (154.8%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): Gemini-2.5-NT = 78,117 tokens
+- Maximum (highest mean): DeepSeek-V3.1 = 198,154 tokens
+- Delta: 120,036 (153.7%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): Gemini-2.5-NT = 94,226 tokens
+- Maximum (highest mean): DeepSeek-V3.1 = 398,139 tokens
+- Delta: 303,913 (322.5%)
+
+### Baseline vs maximum — Failure
+- No data to compare.
+
+## Series: EmailResponder
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 200 | 6,232 |
+| GPT-4o-mini | 200 | 8,741 |
+| DeepSeek-V3.1 | 200 | 8,506 |
+| DeepSeek-R1 | 200 | 16,192 |
+| Gemini-2.5 | 200 | 36,628 |
+| Gemini-2.5-NT | 200 | 38,777 |
+| Qwen3-235b | 200 | 6,667 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 200 | 6,232 |
+| GPT-4o-mini | 200 | 8,741 |
+| DeepSeek-V3.1 | 196 | 8,441 |
+| DeepSeek-R1 | 168 | 14,238 |
+| Gemini-2.5 | 181 | 22,482 |
+| Gemini-2.5-NT | 176 | 12,595 |
+| Qwen3-235b | 200 | 6,667 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 4 | 11,697 |
+| DeepSeek-R1 | 30 | 27,119 |
+| Gemini-2.5 | 18 | 158,396 |
+| Gemini-2.5-NT | 24 | 230,778 |
+| Qwen3-235b | 0 | - |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 2 | 16,502 |
+| Gemini-2.5 | 1 | 405,135 |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 0 | - |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-5 = 6,232 tokens
+- Maximum (highest mean): Gemini-2.5-NT = 38,777 tokens
+- Delta: 32,545 (522.2%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): GPT-5 = 6,232 tokens
+- Maximum (highest mean): Gemini-2.5 = 22,482 tokens
+- Delta: 16,250 (260.7%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): DeepSeek-V3.1 = 11,697 tokens
+- Maximum (highest mean): Gemini-2.5-NT = 230,778 tokens
+- Delta: 219,081 (1872.9%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): DeepSeek-R1 = 16,502 tokens
+- Maximum (highest mean): Gemini-2.5 = 405,135 tokens
+- Delta: 388,633 (2355.1%)
+
+## Series: GameBuilder
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 50 | 22,665 |
+| GPT-4o-mini | 50 | 32,909 |
+| DeepSeek-V3.1 | 50 | 48,315 |
+| DeepSeek-R1 | 50 | 44,694 |
+| Gemini-2.5 | 50 | 74,145 |
+| Gemini-2.5-NT | 50 | 59,373 |
+| Qwen3-235b | 50 | 24,125 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 50 | 22,665 |
+| GPT-4o-mini | 50 | 32,909 |
+| DeepSeek-V3.1 | 49 | 47,722 |
+| DeepSeek-R1 | 50 | 44,694 |
+| Gemini-2.5 | 50 | 74,145 |
+| Gemini-2.5-NT | 48 | 52,527 |
+| Qwen3-235b | 50 | 24,125 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 0 | - |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 2 | 223,682 |
+| Qwen3-235b | 0 | - |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 1 | 77,354 |
+| DeepSeek-R1 | 0 | - |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 0 | - |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-5 = 22,665 tokens
+- Maximum (highest mean): Gemini-2.5 = 74,145 tokens
+- Delta: 51,480 (227.1%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): GPT-5 = 22,665 tokens
+- Maximum (highest mean): Gemini-2.5 = 74,145 tokens
+- Delta: 51,480 (227.1%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): Gemini-2.5-NT = 223,682 tokens
+- Maximum (highest mean): Gemini-2.5-NT = 223,682 tokens
+- Delta: 0 (0.0%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): DeepSeek-V3.1 = 77,354 tokens
+- Maximum (highest mean): DeepSeek-V3.1 = 77,354 tokens
+- Delta: 0 (0.0%)
+
+## Series: LandingPageGenerator
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 17,638 |
+| GPT-4o-mini | 180 | 27,726 |
+| DeepSeek-V3.1 | 180 | 26,392 |
+| DeepSeek-R1 | 180 | 38,077 |
+| Gemini-2.5 | 180 | 31,011 |
+| Gemini-2.5-NT | 180 | 23,734 |
+| Qwen3-235b | 180 | 16,749 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 101 | 17,805 |
+| GPT-4o-mini | 177 | 27,799 |
+| DeepSeek-V3.1 | 172 | 26,666 |
+| DeepSeek-R1 | 49 | 35,431 |
+| Gemini-2.5 | 1 | 55,158 |
+| Gemini-2.5-NT | 2 | 188,992 |
+| Qwen3-235b | 109 | 17,466 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 2 | 28,430 |
+| DeepSeek-R1 | 8 | 61,223 |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 5 | 33,835 |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 79 | 17,425 |
+| GPT-4o-mini | 3 | 23,420 |
+| DeepSeek-V3.1 | 6 | 17,852 |
+| DeepSeek-R1 | 123 | 37,626 |
+| Gemini-2.5 | 179 | 30,876 |
+| Gemini-2.5-NT | 178 | 21,877 |
+| Qwen3-235b | 66 | 14,271 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): Qwen3-235b = 16,749 tokens
+- Maximum (highest mean): DeepSeek-R1 = 38,077 tokens
+- Delta: 21,328 (127.3%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): Qwen3-235b = 17,466 tokens
+- Maximum (highest mean): Gemini-2.5-NT = 188,992 tokens
+- Delta: 171,526 (982.0%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): DeepSeek-V3.1 = 28,430 tokens
+- Maximum (highest mean): DeepSeek-R1 = 61,223 tokens
+- Delta: 32,793 (115.3%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): Qwen3-235b = 14,271 tokens
+- Maximum (highest mean): DeepSeek-R1 = 37,626 tokens
+- Delta: 23,355 (163.7%)
+
+## Series: MarkdownValidator
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 120 | 1,910 |
+| GPT-4o-mini | 120 | 1,897 |
+| DeepSeek-V3.1 | 120 | 2,006 |
+| DeepSeek-R1 | 120 | 5,898 |
+| Gemini-2.5 | 120 | 115,364 |
+| Gemini-2.5-NT | 120 | 46,614 |
+| Qwen3-235b | 120 | 1,924 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 120 | 1,910 |
+| GPT-4o-mini | 120 | 1,897 |
+| DeepSeek-V3.1 | 120 | 2,006 |
+| DeepSeek-R1 | 120 | 5,898 |
+| Gemini-2.5 | 95 | 74,526 |
+| Gemini-2.5-NT | 113 | 27,744 |
+| Qwen3-235b | 120 | 1,924 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 0 | - |
+| Gemini-2.5 | 25 | 270,547 |
+| Gemini-2.5-NT | 7 | 351,232 |
+| Qwen3-235b | 0 | - |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-4o-mini = 1,897 tokens
+- Maximum (highest mean): Gemini-2.5 = 115,364 tokens
+- Delta: 113,466 (5980.0%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): GPT-4o-mini = 1,897 tokens
+- Maximum (highest mean): Gemini-2.5 = 74,526 tokens
+- Delta: 72,629 (3827.7%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): Gemini-2.5 = 270,547 tokens
+- Maximum (highest mean): Gemini-2.5-NT = 351,232 tokens
+- Delta: 80,686 (29.8%)
+
+### Baseline vs maximum — Failure
+- No data to compare.
+
+## Series: RecruitmentAssistant
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 32,358 |
+| GPT-4o-mini | 180 | 92,274 |
+| DeepSeek-V3.1 | 180 | 65,486 |
+| DeepSeek-R1 | 180 | 78,894 |
+| Gemini-2.5 | 180 | 63,570 |
+| Gemini-2.5-NT | 180 | 48,983 |
+| Qwen3-235b | 180 | 86,025 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 32,358 |
+| GPT-4o-mini | 180 | 92,274 |
+| DeepSeek-V3.1 | 180 | 65,486 |
+| DeepSeek-R1 | 171 | 77,784 |
+| Gemini-2.5 | 180 | 63,570 |
+| Gemini-2.5-NT | 180 | 48,983 |
+| Qwen3-235b | 133 | 69,503 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 8 | 108,468 |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 44 | 132,115 |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 1 | 32,043 |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 3 | 142,502 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-5 = 32,358 tokens
+- Maximum (highest mean): GPT-4o-mini = 92,274 tokens
+- Delta: 59,916 (185.2%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): GPT-5 = 32,358 tokens
+- Maximum (highest mean): GPT-4o-mini = 92,274 tokens
+- Delta: 59,916 (185.2%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): DeepSeek-R1 = 108,468 tokens
+- Maximum (highest mean): Qwen3-235b = 132,115 tokens
+- Delta: 23,647 (21.8%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): DeepSeek-R1 = 32,043 tokens
+- Maximum (highest mean): Qwen3-235b = 142,502 tokens
+- Delta: 110,459 (344.7%)
+
+## Series: SQLAssistant
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 11,342 |
+| GPT-4o-mini | 180 | 18,870 |
+| DeepSeek-V3.1 | 180 | 26,307 |
+| DeepSeek-R1 | 180 | 82,118 |
+| Gemini-2.5 | 180 | 25,363 |
+| Gemini-2.5-NT | 180 | 19,757 |
+| Qwen3-235b | 180 | 16,464 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 101 | 7,480 |
+| GPT-4o-mini | 114 | 13,910 |
+| DeepSeek-V3.1 | 70 | 13,760 |
+| DeepSeek-R1 | 36 | 33,567 |
+| Gemini-2.5 | 33 | 15,045 |
+| Gemini-2.5-NT | 45 | 11,328 |
+| Qwen3-235b | 112 | 12,176 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 79 | 16,279 |
+| GPT-4o-mini | 66 | 27,437 |
+| DeepSeek-V3.1 | 110 | 34,291 |
+| DeepSeek-R1 | 144 | 94,256 |
+| Gemini-2.5 | 57 | 37,607 |
+| Gemini-2.5-NT | 30 | 39,861 |
+| Qwen3-235b | 66 | 23,950 |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 0 | - |
+| GPT-4o-mini | 0 | - |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 0 | - |
+| Gemini-2.5 | 90 | 21,392 |
+| Gemini-2.5-NT | 105 | 17,626 |
+| Qwen3-235b | 2 | 9,587 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): GPT-5 = 11,342 tokens
+- Maximum (highest mean): DeepSeek-R1 = 82,118 tokens
+- Delta: 70,776 (624.0%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): GPT-5 = 7,480 tokens
+- Maximum (highest mean): DeepSeek-R1 = 33,567 tokens
+- Delta: 26,087 (348.8%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): GPT-5 = 16,279 tokens
+- Maximum (highest mean): DeepSeek-R1 = 94,256 tokens
+- Delta: 77,977 (479.0%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): Qwen3-235b = 9,587 tokens
+- Maximum (highest mean): Gemini-2.5 = 21,392 tokens
+- Delta: 11,805 (123.1%)
+
+## Series: SocialMediaManager
+
+### Overall per-model total tokens (all statuses)
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 180 | 32,102 |
+| GPT-4o-mini | 180 | 83,075 |
+| DeepSeek-V3.1 | 180 | 80,179 |
+| DeepSeek-R1 | 180 | 110,777 |
+| Gemini-2.5 | 180 | 15,376 |
+| Gemini-2.5-NT | 180 | 10,850 |
+| Qwen3-235b | 180 | 43,954 |
+
+### Direct Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 119 | 19,760 |
+| GPT-4o-mini | 95 | 36,934 |
+| DeepSeek-V3.1 | 132 | 61,643 |
+| DeepSeek-R1 | 69 | 42,398 |
+| Gemini-2.5 | 142 | 12,219 |
+| Gemini-2.5-NT | 153 | 9,820 |
+| Qwen3-235b | 97 | 23,270 |
+
+### Retry Success distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 58 | 53,963 |
+| GPT-4o-mini | 78 | 127,046 |
+| DeepSeek-V3.1 | 48 | 131,154 |
+| DeepSeek-R1 | 101 | 145,943 |
+| Gemini-2.5 | 38 | 27,172 |
+| Gemini-2.5-NT | 27 | 16,691 |
+| Qwen3-235b | 80 | 65,900 |
+
+### Failure distribution
+| Model | n | Mean |
+| --- | --- | --- |
+| GPT-5 | 3 | 99,016 |
+| GPT-4o-mini | 7 | 219,309 |
+| DeepSeek-V3.1 | 0 | - |
+| DeepSeek-R1 | 10 | 227,416 |
+| Gemini-2.5 | 0 | - |
+| Gemini-2.5-NT | 0 | - |
+| Qwen3-235b | 3 | 127,530 |
+
+### Baseline vs maximum (overall mean tokens)
+- Baseline (lowest mean): Gemini-2.5-NT = 10,850 tokens
+- Maximum (highest mean): DeepSeek-R1 = 110,777 tokens
+- Delta: 99,927 (920.9%)
+
+### Baseline vs maximum — Direct Success
+- Baseline (lowest mean): Gemini-2.5-NT = 9,820 tokens
+- Maximum (highest mean): DeepSeek-V3.1 = 61,643 tokens
+- Delta: 51,823 (527.7%)
+
+### Baseline vs maximum — Retry Success
+- Baseline (lowest mean): Gemini-2.5-NT = 16,691 tokens
+- Maximum (highest mean): DeepSeek-R1 = 145,943 tokens
+- Delta: 129,251 (774.4%)
+
+### Baseline vs maximum — Failure
+- Baseline (lowest mean): GPT-5 = 99,016 tokens
+- Maximum (highest mean): DeepSeek-R1 = 227,416 tokens
+- Delta: 128,400 (129.7%)
+
+---
+# Individual Project Details
+
 ## BookWriter-MCP
 
 ### Overall per-model total tokens (all statuses)

data/processed/RQ3/Violin/project_token_stats.md

@@ -1,1329 +0,0 @@
-# Project token statistics
-
-## Overall status token summary
-
-| Status | n | Mean |
-| --- | --- | --- |
-| Pass (no retry) | 6849 | 47,691 |
-| Pass (with retry) | 1179 | 83,133 |
-| Failure | 862 | 29,935 |
-
-## SQL_assistant-A2A
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 36 | 8,736 |
-| GPT-4o-mini | 50 | 17,092 |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 3 | 55,457 |
-| Gemini-2.5 | 2 | 23,118 |
-| Gemini-2.5-NT | 2 | 36,636 |
-| Qwen3-235b | 35 | 13,848 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 24 | 14,869 |
-| GPT-4o-mini | 10 | 26,350 |
-| DeepSeek-V3.1 | 60 | 36,805 |
-| DeepSeek-R1 | 57 | 98,627 |
-| Gemini-2.5 | 20 | 44,005 |
-| Gemini-2.5-NT | 5 | 71,593 |
-| Qwen3-235b | 23 | 23,436 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 38 | 26,988 |
-| Gemini-2.5-NT | 53 | 21,051 |
-| Qwen3-235b | 2 | 9,587 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 8,736 | 14,869 | +6,133 (70.2%) | - | - |
-| GPT-4o-mini | 17,092 | 26,350 | +9,258 (54.2%) | - | - |
-| DeepSeek-V3.1 | - | 36,805 | - | - | - |
-| DeepSeek-R1 | 55,457 | 98,627 | +43,170 (77.8%) | - | - |
-| Gemini-2.5 | 23,118 | 44,005 | +20,887 (90.4%) | 26,988 | +3,870 (16.7%) |
-| Gemini-2.5-NT | 36,636 | 71,593 | +34,956 (95.4%) | 21,051 | -15,585 (-42.5%) |
-| Qwen3-235b | 13,848 | 23,436 | +9,588 (69.2%) | 9,587 | -4,261 (-30.8%) |
-
-## SQL_assistant-A2A_mix
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 46 | 6,953 |
-| GPT-4o-mini | 43 | 10,004 |
-| DeepSeek-V3.1 | 44 | 12,571 |
-| DeepSeek-R1 | 20 | 29,145 |
-| Gemini-2.5 | 25 | 14,332 |
-| Gemini-2.5-NT | 33 | 9,561 |
-| Qwen3-235b | 40 | 10,594 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 14 | 15,537 |
-| GPT-4o-mini | 17 | 23,168 |
-| DeepSeek-V3.1 | 16 | 27,226 |
-| DeepSeek-R1 | 40 | 89,473 |
-| Gemini-2.5 | 26 | 30,481 |
-| Gemini-2.5-NT | 19 | 33,544 |
-| Qwen3-235b | 20 | 23,149 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 9 | 35,120 |
-| Gemini-2.5-NT | 8 | 19,462 |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 6,953 | 15,537 | +8,584 (123.5%) | - | - |
-| GPT-4o-mini | 10,004 | 23,168 | +13,165 (131.6%) | - | - |
-| DeepSeek-V3.1 | 12,571 | 27,226 | +14,655 (116.6%) | - | - |
-| DeepSeek-R1 | 29,145 | 89,473 | +60,327 (207.0%) | - | - |
-| Gemini-2.5 | 14,332 | 30,481 | +16,149 (112.7%) | 35,120 | +20,788 (145.0%) |
-| Gemini-2.5-NT | 9,561 | 33,544 | +23,983 (250.8%) | 19,462 | +9,902 (103.6%) |
-| Qwen3-235b | 10,594 | 23,149 | +12,555 (118.5%) | - | - |
-
-## SQL_assistant-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 19 | 6,378 |
-| GPT-4o-mini | 21 | 14,334 |
-| DeepSeek-V3.1 | 26 | 15,771 |
-| DeepSeek-R1 | 13 | 35,319 |
-| Gemini-2.5 | 6 | 15,321 |
-| Gemini-2.5-NT | 10 | 12,100 |
-| Qwen3-235b | 37 | 12,304 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 41 | 17,358 |
-| GPT-4o-mini | 39 | 29,576 |
-| DeepSeek-V3.1 | 34 | 33,181 |
-| DeepSeek-R1 | 47 | 93,025 |
-| Gemini-2.5 | 11 | 42,818 |
-| Gemini-2.5-NT | 6 | 33,419 |
-| Qwen3-235b | 23 | 25,159 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 43 | 13,573 |
-| Gemini-2.5-NT | 44 | 13,166 |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 6,378 | 17,358 | +10,981 (172.2%) | - | - |
-| GPT-4o-mini | 14,334 | 29,576 | +15,242 (106.3%) | - | - |
-| DeepSeek-V3.1 | 15,771 | 33,181 | +17,409 (110.4%) | - | - |
-| DeepSeek-R1 | 35,319 | 93,025 | +57,706 (163.4%) | - | - |
-| Gemini-2.5 | 15,321 | 42,818 | +27,497 (179.5%) | 13,573 | -1,748 (-11.4%) |
-| Gemini-2.5-NT | 12,100 | 33,419 | +21,319 (176.2%) | 13,166 | +1,066 (8.8%) |
-| Qwen3-235b | 12,304 | 25,159 | +12,855 (104.5%) | - | - |
-
-## email_auto_responder_flow-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 100 | 6,903 |
-| GPT-4o-mini | 100 | 9,437 |
-| DeepSeek-V3.1 | 98 | 9,531 |
-| DeepSeek-R1 | 96 | 14,798 |
-| Gemini-2.5 | 91 | 33,629 |
-| Gemini-2.5-NT | 88 | 12,696 |
-| Qwen3-235b | 100 | 7,434 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 2 | 11,981 |
-| DeepSeek-R1 | 4 | 29,922 |
-| Gemini-2.5 | 9 | 82,929 |
-| Gemini-2.5-NT | 12 | 199,673 |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 6,903 | - | - | - | - |
-| GPT-4o-mini | 9,437 | - | - | - | - |
-| DeepSeek-V3.1 | 9,531 | 11,981 | +2,450 (25.7%) | - | - |
-| DeepSeek-R1 | 14,798 | 29,922 | +15,123 (102.2%) | - | - |
-| Gemini-2.5 | 33,629 | 82,929 | +49,300 (146.6%) | - | - |
-| Gemini-2.5-NT | 12,696 | 199,673 | +186,977 (1472.7%) | - | - |
-| Qwen3-235b | 7,434 | - | - | - | - |
-
-## email_auto_responder_flow
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 100 | 5,562 |
-| GPT-4o-mini | 100 | 8,044 |
-| DeepSeek-V3.1 | 98 | 7,350 |
-| DeepSeek-R1 | 72 | 13,490 |
-| Gemini-2.5 | 90 | 11,211 |
-| Gemini-2.5-NT | 88 | 12,494 |
-| Qwen3-235b | 100 | 5,900 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 2 | 11,414 |
-| DeepSeek-R1 | 26 | 26,688 |
-| Gemini-2.5 | 9 | 233,862 |
-| Gemini-2.5-NT | 12 | 261,883 |
-| Qwen3-235b | 0 | - |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 2 | 16,502 |
-| Gemini-2.5 | 1 | 405,135 |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 5,562 | - | - | - | - |
-| GPT-4o-mini | 8,044 | - | - | - | - |
-| DeepSeek-V3.1 | 7,350 | 11,414 | +4,064 (55.3%) | - | - |
-| DeepSeek-R1 | 13,490 | 26,688 | +13,198 (97.8%) | 16,502 | +3,012 (22.3%) |
-| Gemini-2.5 | 11,211 | 233,862 | +222,651 (1986.0%) | 405,135 | +393,924 (3513.7%) |
-| Gemini-2.5-NT | 12,494 | 261,883 | +249,388 (1996.0%) | - | - |
-| Qwen3-235b | 5,900 | - | - | - | - |
-
-## game_builder-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 25 | 22,425 |
-| GPT-4o-mini | 25 | 47,996 |
-| DeepSeek-V3.1 | 25 | 55,289 |
-| DeepSeek-R1 | 25 | 44,851 |
-| Gemini-2.5 | 25 | 75,519 |
-| Gemini-2.5-NT | 25 | 50,782 |
-| Qwen3-235b | 25 | 24,039 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 22,425 | - | - | - | - |
-| GPT-4o-mini | 47,996 | - | - | - | - |
-| DeepSeek-V3.1 | 55,289 | - | - | - | - |
-| DeepSeek-R1 | 44,851 | - | - | - | - |
-| Gemini-2.5 | 75,519 | - | - | - | - |
-| Gemini-2.5-NT | 50,782 | - | - | - | - |
-| Qwen3-235b | 24,039 | - | - | - | - |
-
-## game_builder
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 25 | 22,904 |
-| GPT-4o-mini | 25 | 17,823 |
-| DeepSeek-V3.1 | 24 | 39,841 |
-| DeepSeek-R1 | 25 | 44,536 |
-| Gemini-2.5 | 25 | 72,772 |
-| Gemini-2.5-NT | 23 | 54,423 |
-| Qwen3-235b | 25 | 24,212 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 2 | 223,682 |
-| Qwen3-235b | 0 | - |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 1 | 77,354 |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 22,904 | - | - | - | - |
-| GPT-4o-mini | 17,823 | - | - | - | - |
-| DeepSeek-V3.1 | 39,841 | - | - | 77,354 | +37,513 (94.2%) |
-| DeepSeek-R1 | 44,536 | - | - | - | - |
-| Gemini-2.5 | 72,772 | - | - | - | - |
-| Gemini-2.5-NT | 54,423 | 223,682 | +169,258 (311.0%) | - | - |
-| Qwen3-235b | 24,212 | - | - | - | - |
-
-## intelligent_recruitment_platform-A2A
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 31,991 |
-| GPT-4o-mini | 60 | 74,951 |
-| DeepSeek-V3.1 | 60 | 65,662 |
-| DeepSeek-R1 | 56 | 73,964 |
-| Gemini-2.5 | 60 | 68,719 |
-| Gemini-2.5-NT | 60 | 55,660 |
-| Qwen3-235b | 44 | 73,316 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 3 | 68,009 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 15 | 164,476 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 1 | 32,043 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 1 | 169,975 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 31,991 | - | - | - | - |
-| GPT-4o-mini | 74,951 | - | - | - | - |
-| DeepSeek-V3.1 | 65,662 | - | - | - | - |
-| DeepSeek-R1 | 73,964 | 68,009 | -5,956 (-8.1%) | 32,043 | -41,921 (-56.7%) |
-| Gemini-2.5 | 68,719 | - | - | - | - |
-| Gemini-2.5-NT | 55,660 | - | - | - | - |
-| Qwen3-235b | 73,316 | 164,476 | +91,160 (124.3%) | 169,975 | +96,659 (131.8%) |
-
-## intelligent_recruitment_platform-A2A_mix
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 32,672 |
-| GPT-4o-mini | 60 | 101,936 |
-| DeepSeek-V3.1 | 60 | 65,048 |
-| DeepSeek-R1 | 57 | 76,741 |
-| Gemini-2.5 | 60 | 57,128 |
-| Gemini-2.5-NT | 60 | 36,887 |
-| Qwen3-235b | 49 | 67,089 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 3 | 122,480 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 11 | 100,743 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 32,672 | - | - | - | - |
-| GPT-4o-mini | 101,936 | - | - | - | - |
-| DeepSeek-V3.1 | 65,048 | - | - | - | - |
-| DeepSeek-R1 | 76,741 | 122,480 | +45,739 (59.6%) | - | - |
-| Gemini-2.5 | 57,128 | - | - | - | - |
-| Gemini-2.5-NT | 36,887 | - | - | - | - |
-| Qwen3-235b | 67,089 | 100,743 | +33,653 (50.2%) | - | - |
-
-## intelligent_recruitment_platform-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 32,411 |
-| GPT-4o-mini | 60 | 99,933 |
-| DeepSeek-V3.1 | 60 | 65,748 |
-| DeepSeek-R1 | 58 | 82,498 |
-| Gemini-2.5 | 60 | 64,864 |
-| Gemini-2.5-NT | 60 | 54,402 |
-| Qwen3-235b | 40 | 68,264 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 2 | 148,137 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 18 | 124,320 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 2 | 128,765 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 32,411 | - | - | - | - |
-| GPT-4o-mini | 99,933 | - | - | - | - |
-| DeepSeek-V3.1 | 65,748 | - | - | - | - |
-| DeepSeek-R1 | 82,498 | 148,137 | +65,639 (79.6%) | - | - |
-| Gemini-2.5 | 64,864 | - | - | - | - |
-| Gemini-2.5-NT | 54,402 | - | - | - | - |
-| Qwen3-235b | 68,264 | 124,320 | +56,055 (82.1%) | 128,765 | +60,501 (88.6%) |
-
-## landing_page_generator-A2A
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 16 | 18,863 |
-| GPT-4o-mini | 58 | 20,076 |
-| DeepSeek-V3.1 | 60 | 28,814 |
-| DeepSeek-R1 | 7 | 46,577 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 27 | 20,066 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 3 | 34,203 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 44 | 13,548 |
-| GPT-4o-mini | 2 | 27,976 |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 53 | 26,291 |
-| Gemini-2.5 | 60 | 21,256 |
-| Gemini-2.5-NT | 60 | 22,826 |
-| Qwen3-235b | 30 | 14,804 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 18,863 | - | - | 13,548 | -5,315 (-28.2%) |
-| GPT-4o-mini | 20,076 | - | - | 27,976 | +7,900 (39.4%) |
-| DeepSeek-V3.1 | 28,814 | - | - | - | - |
-| DeepSeek-R1 | 46,577 | - | - | 26,291 | -20,286 (-43.6%) |
-| Gemini-2.5 | - | - | - | 21,256 | - |
-| Gemini-2.5-NT | - | - | - | 22,826 | - |
-| Qwen3-235b | 20,066 | 34,203 | +14,137 (70.5%) | 14,804 | -5,263 (-26.2%) |
-
-## landing_page_generator-A2A_mix
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 25 | 29,492 |
-| GPT-4o-mini | 60 | 43,589 |
-| DeepSeek-V3.1 | 57 | 34,783 |
-| DeepSeek-R1 | 9 | 68,378 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 29 | 19,428 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 2 | 28,430 |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 2 | 33,284 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 35 | 22,299 |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 1 | 30,301 |
-| DeepSeek-R1 | 51 | 55,231 |
-| Gemini-2.5 | 60 | 45,250 |
-| Gemini-2.5-NT | 60 | 14,949 |
-| Qwen3-235b | 29 | 14,748 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 29,492 | - | - | 22,299 | -7,194 (-24.4%) |
-| GPT-4o-mini | 43,589 | - | - | - | - |
-| DeepSeek-V3.1 | 34,783 | 28,430 | -6,353 (-18.3%) | 30,301 | -4,482 (-12.9%) |
-| DeepSeek-R1 | 68,378 | - | - | 55,231 | -13,147 (-19.2%) |
-| Gemini-2.5 | - | - | - | 45,250 | - |
-| Gemini-2.5-NT | - | - | - | 14,949 | - |
-| Qwen3-235b | 19,428 | 33,284 | +13,856 (71.3%) | 14,748 | -4,680 (-24.1%) |
-
-## landing_page_generator-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 12,653 |
-| GPT-4o-mini | 59 | 19,333 |
-| DeepSeek-V3.1 | 55 | 15,912 |
-| DeepSeek-R1 | 33 | 24,081 |
-| Gemini-2.5 | 1 | 55,158 |
-| Gemini-2.5-NT | 2 | 188,992 |
-| Qwen3-235b | 53 | 15,068 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 8 | 61,223 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 1 | 14,308 |
-| DeepSeek-V3.1 | 5 | 15,362 |
-| DeepSeek-R1 | 19 | 21,990 |
-| Gemini-2.5 | 59 | 26,041 |
-| Gemini-2.5-NT | 58 | 28,064 |
-| Qwen3-235b | 7 | 10,009 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 12,653 | - | - | - | - |
-| GPT-4o-mini | 19,333 | - | - | 14,308 | -5,025 (-26.0%) |
-| DeepSeek-V3.1 | 15,912 | - | - | 15,362 | -550 (-3.5%) |
-| DeepSeek-R1 | 24,081 | 61,223 | +37,142 (154.2%) | 21,990 | -2,091 (-8.7%) |
-| Gemini-2.5 | 55,158 | - | - | 26,041 | -29,117 (-52.8%) |
-| Gemini-2.5-NT | 188,992 | - | - | 28,064 | -160,929 (-85.2%) |
-| Qwen3-235b | 15,068 | - | - | 10,009 | -5,059 (-33.6%) |
-
-## markdown_validator-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 1,967 |
-| GPT-4o-mini | 60 | 1,942 |
-| DeepSeek-V3.1 | 60 | 2,063 |
-| DeepSeek-R1 | 60 | 5,885 |
-| Gemini-2.5 | 48 | 67,759 |
-| Gemini-2.5-NT | 56 | 39,448 |
-| Qwen3-235b | 60 | 1,994 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 12 | 206,480 |
-| Gemini-2.5-NT | 4 | 376,206 |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 1,967 | - | - | - | - |
-| GPT-4o-mini | 1,942 | - | - | - | - |
-| DeepSeek-V3.1 | 2,063 | - | - | - | - |
-| DeepSeek-R1 | 5,885 | - | - | - | - |
-| Gemini-2.5 | 67,759 | 206,480 | +138,721 (204.7%) | - | - |
-| Gemini-2.5-NT | 39,448 | 376,206 | +336,758 (853.7%) | - | - |
-| Qwen3-235b | 1,994 | - | - | - | - |
-
-## markdown_validator
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 1,853 |
-| GPT-4o-mini | 60 | 1,853 |
-| DeepSeek-V3.1 | 60 | 1,948 |
-| DeepSeek-R1 | 60 | 5,912 |
-| Gemini-2.5 | 47 | 81,437 |
-| Gemini-2.5-NT | 57 | 16,246 |
-| Qwen3-235b | 60 | 1,854 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 13 | 329,685 |
-| Gemini-2.5-NT | 3 | 317,934 |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 1,853 | - | - | - | - |
-| GPT-4o-mini | 1,853 | - | - | - | - |
-| DeepSeek-V3.1 | 1,948 | - | - | - | - |
-| DeepSeek-R1 | 5,912 | - | - | - | - |
-| Gemini-2.5 | 81,437 | 329,685 | +248,248 (304.8%) | - | - |
-| Gemini-2.5-NT | 16,246 | 317,934 | +301,688 (1857.0%) | - | - |
-| Qwen3-235b | 1,854 | - | - | - | - |
-
-## self_evaluation_loop_flow-A2A
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 31 | 23,996 |
-| GPT-4o-mini | 24 | 42,258 |
-| DeepSeek-V3.1 | 46 | 70,407 |
-| DeepSeek-R1 | 21 | 45,776 |
-| Gemini-2.5 | 56 | 12,176 |
-| Gemini-2.5-NT | 53 | 10,675 |
-| Qwen3-235b | 27 | 30,147 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 27 | 55,285 |
-| GPT-4o-mini | 32 | 125,026 |
-| DeepSeek-V3.1 | 14 | 143,090 |
-| DeepSeek-R1 | 34 | 165,493 |
-| Gemini-2.5 | 4 | 23,860 |
-| Gemini-2.5-NT | 7 | 17,136 |
-| Qwen3-235b | 32 | 72,094 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 2 | 98,808 |
-| GPT-4o-mini | 4 | 213,454 |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 5 | 181,907 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 1 | 230,991 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 23,996 | 55,285 | +31,289 (130.4%) | 98,808 | +74,812 (311.8%) |
-| GPT-4o-mini | 42,258 | 125,026 | +82,768 (195.9%) | 213,454 | +171,196 (405.1%) |
-| DeepSeek-V3.1 | 70,407 | 143,090 | +72,683 (103.2%) | - | - |
-| DeepSeek-R1 | 45,776 | 165,493 | +119,718 (261.5%) | 181,907 | +136,131 (297.4%) |
-| Gemini-2.5 | 12,176 | 23,860 | +11,684 (96.0%) | - | - |
-| Gemini-2.5-NT | 10,675 | 17,136 | +6,461 (60.5%) | - | - |
-| Qwen3-235b | 30,147 | 72,094 | +41,947 (139.1%) | 230,991 | +200,844 (666.2%) |
-
-## self_evaluation_loop_flow-A2A_mix
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 55 | 15,857 |
-| GPT-4o-mini | 51 | 33,008 |
-| DeepSeek-V3.1 | 49 | 51,074 |
-| DeepSeek-R1 | 27 | 40,199 |
-| Gemini-2.5 | 35 | 13,495 |
-| Gemini-2.5-NT | 45 | 7,487 |
-| Qwen3-235b | 35 | 14,019 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 5 | 27,133 |
-| GPT-4o-mini | 9 | 101,766 |
-| DeepSeek-V3.1 | 11 | 85,846 |
-| DeepSeek-R1 | 31 | 133,368 |
-| Gemini-2.5 | 25 | 30,123 |
-| Gemini-2.5-NT | 15 | 16,107 |
-| Qwen3-235b | 24 | 51,432 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 2 | 336,866 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 1 | 41,361 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 15,857 | 27,133 | +11,276 (71.1%) | - | - |
-| GPT-4o-mini | 33,008 | 101,766 | +68,758 (208.3%) | - | - |
-| DeepSeek-V3.1 | 51,074 | 85,846 | +34,772 (68.1%) | - | - |
-| DeepSeek-R1 | 40,199 | 133,368 | +93,169 (231.8%) | 336,866 | +296,667 (738.0%) |
-| Gemini-2.5 | 13,495 | 30,123 | +16,628 (123.2%) | - | - |
-| Gemini-2.5-NT | 7,487 | 16,107 | +8,620 (115.1%) | - | - |
-| Qwen3-235b | 14,019 | 51,432 | +37,413 (266.9%) | 41,361 | +27,342 (195.0%) |
-
-## self_evaluation_loop_flow-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 33 | 22,287 |
-| GPT-4o-mini | 20 | 40,555 |
-| DeepSeek-V3.1 | 37 | 64,742 |
-| DeepSeek-R1 | 21 | 41,849 |
-| Gemini-2.5 | 51 | 11,390 |
-| Gemini-2.5-NT | 55 | 10,904 |
-| Qwen3-235b | 35 | 27,216 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 26 | 57,749 |
-| GPT-4o-mini | 37 | 134,943 |
-| DeepSeek-V3.1 | 23 | 145,557 |
-| DeepSeek-R1 | 36 | 138,306 |
-| Gemini-2.5 | 9 | 20,447 |
-| Gemini-2.5-NT | 5 | 17,821 |
-| Qwen3-235b | 24 | 72,110 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 1 | 99,433 |
-| GPT-4o-mini | 3 | 227,117 |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 3 | 230,298 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 1 | 110,239 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 22,287 | 57,749 | +35,462 (159.1%) | 99,433 | +77,146 (346.1%) |
-| GPT-4o-mini | 40,555 | 134,943 | +94,388 (232.7%) | 227,117 | +186,562 (460.0%) |
-| DeepSeek-V3.1 | 64,742 | 145,557 | +80,815 (124.8%) | - | - |
-| DeepSeek-R1 | 41,849 | 138,306 | +96,458 (230.5%) | 230,298 | +188,449 (450.3%) |
-| Gemini-2.5 | 11,390 | 20,447 | +9,057 (79.5%) | - | - |
-| Gemini-2.5-NT | 10,904 | 17,821 | +6,917 (63.4%) | - | - |
-| Qwen3-235b | 27,216 | 72,110 | +44,894 (165.0%) | 110,239 | +83,023 (305.1%) |
-
-## write_a_book_with_flows-A2A
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 181,022 |
-| GPT-4o-mini | 60 | 139,665 |
-| DeepSeek-V3.1 | 59 | 270,142 |
-| DeepSeek-R1 | 52 | 173,528 |
-| Gemini-2.5 | 60 | 110,126 |
-| Gemini-2.5-NT | 60 | 86,249 |
-| Qwen3-235b | 60 | 118,406 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 1 | 398,139 |
-| DeepSeek-R1 | 8 | 272,152 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 181,022 | - | - | - | - |
-| GPT-4o-mini | 139,665 | - | - | - | - |
-| DeepSeek-V3.1 | 270,142 | 398,139 | +127,997 (47.4%) | - | - |
-| DeepSeek-R1 | 173,528 | 272,152 | +98,623 (56.8%) | - | - |
-| Gemini-2.5 | 110,126 | - | - | - | - |
-| Gemini-2.5-NT | 86,249 | - | - | - | - |
-| Qwen3-235b | 118,406 | - | - | - | - |
-
-## write_a_book_with_flows-A2A_mix
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 117,358 |
-| GPT-4o-mini | 60 | 102,564 |
-| DeepSeek-V3.1 | 60 | 141,204 |
-| DeepSeek-R1 | 59 | 128,948 |
-| Gemini-2.5 | 60 | 91,585 |
-| Gemini-2.5-NT | 60 | 65,826 |
-| Qwen3-235b | 60 | 76,512 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 1 | 185,051 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 117,358 | - | - | - | - |
-| GPT-4o-mini | 102,564 | - | - | - | - |
-| DeepSeek-V3.1 | 141,204 | - | - | - | - |
-| DeepSeek-R1 | 128,948 | 185,051 | +56,103 (43.5%) | - | - |
-| Gemini-2.5 | 91,585 | - | - | - | - |
-| Gemini-2.5-NT | 65,826 | - | - | - | - |
-| Qwen3-235b | 76,512 | - | - | - | - |
-
-## write_a_book_with_flows-MCP
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 60 | 190,842 |
-| GPT-4o-mini | 60 | 133,339 |
-| DeepSeek-V3.1 | 60 | 184,315 |
-| DeepSeek-R1 | 52 | 171,033 |
-| Gemini-2.5 | 60 | 107,154 |
-| Gemini-2.5-NT | 59 | 82,348 |
-| Qwen3-235b | 59 | 110,871 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 8 | 318,572 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 1 | 94,226 |
-| Qwen3-235b | 1 | 101,570 |
-
-### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 190,842 | - | - | - | - |
-| GPT-4o-mini | 133,339 | - | - | - | - |
-| DeepSeek-V3.1 | 184,315 | - | - | - | - |
-| DeepSeek-R1 | 171,033 | 318,572 | +147,539 (86.3%) | - | - |
-| Gemini-2.5 | 107,154 | - | - | - | - |
-| Gemini-2.5-NT | 82,348 | 94,226 | +11,878 (14.4%) | - | - |
-| Qwen3-235b | 110,871 | 101,570 | -9,301 (-8.4%) | - | - |
-
-## Series aggregates (by task prefix)
-
-### SQL_assistant (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 101 | 7,480 |
-| GPT-4o-mini | 114 | 13,910 |
-| DeepSeek-V3.1 | 70 | 13,760 |
-| DeepSeek-R1 | 36 | 33,567 |
-| Gemini-2.5 | 33 | 15,045 |
-| Gemini-2.5-NT | 45 | 11,328 |
-| Qwen3-235b | 112 | 12,176 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 79 | 16,279 |
-| GPT-4o-mini | 66 | 27,437 |
-| DeepSeek-V3.1 | 110 | 34,291 |
-| DeepSeek-R1 | 144 | 94,256 |
-| Gemini-2.5 | 57 | 37,607 |
-| Gemini-2.5-NT | 30 | 39,861 |
-| Qwen3-235b | 66 | 23,950 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 90 | 21,392 |
-| Gemini-2.5-NT | 105 | 17,626 |
-| Qwen3-235b | 2 | 9,587 |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 7,480 | 16,279 | +8,799 (117.6%) | - | - |
-| GPT-4o-mini | 13,910 | 27,437 | +13,527 (97.2%) | - | - |
-| DeepSeek-V3.1 | 13,760 | 34,291 | +20,532 (149.2%) | - | - |
-| DeepSeek-R1 | 33,567 | 94,256 | +60,689 (180.8%) | - | - |
-| Gemini-2.5 | 15,045 | 37,607 | +22,563 (150.0%) | 21,392 | +6,347 (42.2%) |
-| Gemini-2.5-NT | 11,328 | 39,861 | +28,532 (251.9%) | 17,626 | +6,297 (55.6%) |
-| Qwen3-235b | 12,176 | 23,950 | +11,774 (96.7%) | 9,587 | -2,589 (-21.3%) |
-
-### email_auto_responder_flow (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 200 | 6,232 |
-| GPT-4o-mini | 200 | 8,741 |
-| DeepSeek-V3.1 | 196 | 8,441 |
-| DeepSeek-R1 | 168 | 14,238 |
-| Gemini-2.5 | 181 | 22,482 |
-| Gemini-2.5-NT | 176 | 12,595 |
-| Qwen3-235b | 200 | 6,667 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 4 | 11,697 |
-| DeepSeek-R1 | 30 | 27,119 |
-| Gemini-2.5 | 18 | 158,396 |
-| Gemini-2.5-NT | 24 | 230,778 |
-| Qwen3-235b | 0 | - |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 2 | 16,502 |
-| Gemini-2.5 | 1 | 405,135 |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 6,232 | - | - | - | - |
-| GPT-4o-mini | 8,741 | - | - | - | - |
-| DeepSeek-V3.1 | 8,441 | 11,697 | +3,257 (38.6%) | - | - |
-| DeepSeek-R1 | 14,238 | 27,119 | +12,881 (90.5%) | 16,502 | +2,264 (15.9%) |
-| Gemini-2.5 | 22,482 | 158,396 | +135,913 (604.5%) | 405,135 | +382,653 (1702.0%) |
-| Gemini-2.5-NT | 12,595 | 230,778 | +218,183 (1732.3%) | - | - |
-| Qwen3-235b | 6,667 | - | - | - | - |
-
-### game_builder (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 50 | 22,665 |
-| GPT-4o-mini | 50 | 32,909 |
-| DeepSeek-V3.1 | 49 | 47,722 |
-| DeepSeek-R1 | 50 | 44,694 |
-| Gemini-2.5 | 50 | 74,145 |
-| Gemini-2.5-NT | 48 | 52,527 |
-| Qwen3-235b | 50 | 24,125 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 2 | 223,682 |
-| Qwen3-235b | 0 | - |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 1 | 77,354 |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 0 | - |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 22,665 | - | - | - | - |
-| GPT-4o-mini | 32,909 | - | - | - | - |
-| DeepSeek-V3.1 | 47,722 | - | - | 77,354 | +29,632 (62.1%) |
-| DeepSeek-R1 | 44,694 | - | - | - | - |
-| Gemini-2.5 | 74,145 | - | - | - | - |
-| Gemini-2.5-NT | 52,527 | 223,682 | +171,154 (325.8%) | - | - |
-| Qwen3-235b | 24,125 | - | - | - | - |
-
-### intelligent_recruitment_platform (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 180 | 32,358 |
-| GPT-4o-mini | 180 | 92,274 |
-| DeepSeek-V3.1 | 180 | 65,486 |
-| DeepSeek-R1 | 171 | 77,784 |
-| Gemini-2.5 | 180 | 63,570 |
-| Gemini-2.5-NT | 180 | 48,983 |
-| Qwen3-235b | 133 | 69,503 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 8 | 108,468 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 44 | 132,115 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 1 | 32,043 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 3 | 142,502 |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 32,358 | - | - | - | - |
-| GPT-4o-mini | 92,274 | - | - | - | - |
-| DeepSeek-V3.1 | 65,486 | - | - | - | - |
-| DeepSeek-R1 | 77,784 | 108,468 | +30,683 (39.4%) | 32,043 | -45,741 (-58.8%) |
-| Gemini-2.5 | 63,570 | - | - | - | - |
-| Gemini-2.5-NT | 48,983 | - | - | - | - |
-| Qwen3-235b | 69,503 | 132,115 | +62,612 (90.1%) | 142,502 | +72,999 (105.0%) |
-
-### landing_page_generator (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 101 | 17,805 |
-| GPT-4o-mini | 177 | 27,799 |
-| DeepSeek-V3.1 | 172 | 26,666 |
-| DeepSeek-R1 | 49 | 35,431 |
-| Gemini-2.5 | 1 | 55,158 |
-| Gemini-2.5-NT | 2 | 188,992 |
-| Qwen3-235b | 109 | 17,466 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 2 | 28,430 |
-| DeepSeek-R1 | 8 | 61,223 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 5 | 33,835 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 79 | 17,425 |
-| GPT-4o-mini | 3 | 23,420 |
-| DeepSeek-V3.1 | 6 | 17,852 |
-| DeepSeek-R1 | 123 | 37,626 |
-| Gemini-2.5 | 179 | 30,876 |
-| Gemini-2.5-NT | 178 | 21,877 |
-| Qwen3-235b | 66 | 14,271 |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 17,805 | - | - | 17,425 | -381 (-2.1%) |
-| GPT-4o-mini | 27,799 | - | - | 23,420 | -4,379 (-15.8%) |
-| DeepSeek-V3.1 | 26,666 | 28,430 | +1,763 (6.6%) | 17,852 | -8,815 (-33.1%) |
-| DeepSeek-R1 | 35,431 | 61,223 | +25,792 (72.8%) | 37,626 | +2,195 (6.2%) |
-| Gemini-2.5 | 55,158 | - | - | 30,876 | -24,282 (-44.0%) |
-| Gemini-2.5-NT | 188,992 | - | - | 21,877 | -167,115 (-88.4%) |
-| Qwen3-235b | 17,466 | 33,835 | +16,369 (93.7%) | 14,271 | -3,195 (-18.3%) |
-
-### markdown_validator (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 120 | 1,910 |
-| GPT-4o-mini | 120 | 1,897 |
-| DeepSeek-V3.1 | 120 | 2,006 |
-| DeepSeek-R1 | 120 | 5,898 |
-| Gemini-2.5 | 95 | 74,526 |
-| Gemini-2.5-NT | 113 | 27,744 |
-| Qwen3-235b | 120 | 1,924 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 0 | - |
-| Gemini-2.5 | 25 | 270,547 |
-| Gemini-2.5-NT | 7 | 351,232 |
-| Qwen3-235b | 0 | - |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 1,910 | - | - | - | - |
-| GPT-4o-mini | 1,897 | - | - | - | - |
-| DeepSeek-V3.1 | 2,006 | - | - | - | - |
-| DeepSeek-R1 | 5,898 | - | - | - | - |
-| Gemini-2.5 | 74,526 | 270,547 | +196,021 (263.0%) | - | - |
-| Gemini-2.5-NT | 27,744 | 351,232 | +323,488 (1166.0%) | - | - |
-| Qwen3-235b | 1,924 | - | - | - | - |
-
-### self_evaluation_loop_flow (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 119 | 19,760 |
-| GPT-4o-mini | 95 | 36,934 |
-| DeepSeek-V3.1 | 132 | 61,643 |
-| DeepSeek-R1 | 69 | 42,398 |
-| Gemini-2.5 | 142 | 12,219 |
-| Gemini-2.5-NT | 153 | 9,820 |
-| Qwen3-235b | 97 | 23,270 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 58 | 53,963 |
-| GPT-4o-mini | 78 | 127,046 |
-| DeepSeek-V3.1 | 48 | 131,154 |
-| DeepSeek-R1 | 101 | 145,943 |
-| Gemini-2.5 | 38 | 27,172 |
-| Gemini-2.5-NT | 27 | 16,691 |
-| Qwen3-235b | 80 | 65,900 |
-
-### Failure
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 3 | 99,016 |
-| GPT-4o-mini | 7 | 219,309 |
-| DeepSeek-V3.1 | 0 | - |
-| DeepSeek-R1 | 10 | 227,416 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 0 | - |
-| Qwen3-235b | 3 | 127,530 |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 19,760 | 53,963 | +34,202 (173.1%) | 99,016 | +79,256 (401.1%) |
-| GPT-4o-mini | 36,934 | 127,046 | +90,112 (244.0%) | 219,309 | +182,375 (493.8%) |
-| DeepSeek-V3.1 | 61,643 | 131,154 | +69,511 (112.8%) | - | - |
-| DeepSeek-R1 | 42,398 | 145,943 | +103,544 (244.2%) | 227,416 | +185,018 (436.4%) |
-| Gemini-2.5 | 12,219 | 27,172 | +14,953 (122.4%) | - | - |
-| Gemini-2.5-NT | 9,820 | 16,691 | +6,872 (70.0%) | - | - |
-| Qwen3-235b | 23,270 | 65,900 | +42,630 (183.2%) | 127,530 | +104,260 (448.0%) |
-
-### write_a_book_with_flows (aggregated across architectures)
-
-### Pass (no retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 180 | 163,074 |
-| GPT-4o-mini | 180 | 125,190 |
-| DeepSeek-V3.1 | 179 | 198,154 |
-| DeepSeek-R1 | 163 | 156,596 |
-| Gemini-2.5 | 180 | 102,955 |
-| Gemini-2.5-NT | 179 | 78,117 |
-| Qwen3-235b | 179 | 101,880 |
-
-### Pass (with retry)
-
-| Model | n | Mean |
-| --- | --- | --- |
-| GPT-5 | 0 | - |
-| GPT-4o-mini | 0 | - |
-| DeepSeek-V3.1 | 1 | 398,139 |
-| DeepSeek-R1 | 17 | 288,873 |
-| Gemini-2.5 | 0 | - |
-| Gemini-2.5-NT | 1 | 94,226 |
-| Qwen3-235b | 1 | 101,570 |
-
-#### Status comparison (mean, abs & % vs Pass no retry)
-
-| Model | Pass (no retry) mean | Pass (with retry) mean | Δ retry vs no | Failure mean | Δ failure vs no |
-| --- | --- | --- | --- | --- | --- |
-| GPT-5 | 163,074 | - | - | - | - |
-| GPT-4o-mini | 125,190 | - | - | - | - |
-| DeepSeek-V3.1 | 198,154 | 398,139 | +199,985 (100.9%) | - | - |
-| DeepSeek-R1 | 156,596 | 288,873 | +132,277 (84.5%) | - | - |
-| Gemini-2.5 | 102,955 | - | - | - | - |
-| Gemini-2.5-NT | 78,117 | 94,226 | +16,109 (20.6%) | - | - |
-| Qwen3-235b | 101,880 | 101,570 | -310 (-0.3%) | - | - |

data/processed/RQ3/plot_total_tokens_violin.py

@@ -814,14 +814,22 @@ def generate_project_model_distribution_md(
         "Only the mean total tokens are reported. Baseline vs maximum is computed from the overall mean aggregated across all three statuses."
     )
 
-    for task, arch, name in projects:
-        pdata = project_data.get(name, {})
+    def _base_task_name(task: str) -> str:
+        suffixes = ("-H_A2A", "-H-A2A", "-MCP", "-A2A")
+        for suffix in suffixes:
+            if task.endswith(suffix):
+                return task[: -len(suffix)]
+        return task
+
+    def render_distribution_block(
+        title: str, pdata: Dict[str, Dict[str, List[float]]]
+    ) -> None:
         lines.append("")
-        lines.append(f"## {name}")
+        lines.append(f"## {title}")
         if not pdata:
             lines.append("")
             lines.append("> No token data found.")
-            continue
+            return
 
         # overall per-model summary
         lines.append("")
@@ -914,6 +922,44 @@ def generate_project_model_distribution_md(
             else:
                 lines.append("- No data to compare.")
 
+    # 1. Total aggregation across all projects
+    total_data: Dict[str, Dict[str, List[float]]] = defaultdict(
+        lambda: defaultdict(list)
+    )
+    for _, _, name in projects:
+        pdata = project_data.get(name, {})
+        for status in STATUS_ORDER:
+            by_model = pdata.get(status, {})
+            for model, vals in by_model.items():
+                total_data[status][model].extend(vals)
+    render_distribution_block("TOTAL AGGREGATION (All Projects)", total_data)
+
+    # 2. Series aggregation
+    series_data: Dict[str, Dict[str, Dict[str, List[float]]]] = defaultdict(
+        lambda: defaultdict(lambda: defaultdict(list))
+    )
+    for task, arch, name in projects:
+        base_task = _base_task_name(task)
+        pdata = project_data.get(name, {})
+        for status in STATUS_ORDER:
+            by_model = pdata.get(status, {})
+            for model, vals in by_model.items():
+                series_data[base_task][status][model].extend(vals)
+
+    lines.append("")
+    lines.append("---")
+    lines.append("# Series Aggregates")
+    for task in sorted(series_data.keys()):
+        render_distribution_block(f"Series: {task}", series_data[task])
+
+    # 3. Individual projects
+    lines.append("")
+    lines.append("---")
+    lines.append("# Individual Project Details")
+    for task, arch, name in projects:
+        pdata = project_data.get(name, {})
+        render_distribution_block(name, pdata)
+
     out_path = out_dir / "project_model_distribution.md"
     out_dir.mkdir(parents=True, exist_ok=True)
     out_path.write_text("\n".join(lines), encoding="utf-8")