Add extended public-safe evidence bundle v1

checksums/extended-evidence-v1.sha256

@@ -0,0 +1,309 @@
+2469e607e691cba884beda0fe55fe69bf1dbb14ca8f22765a81139f6bf60d18d  docs/extended-evidence-v1/EXCLUSIONS.md
+41d8ad21168e4958be51f2a6649c585c2151c807d95f6d05daaef0d837ee0404  docs/extended-evidence-v1/README.md
+70082850ca060f1a2b0a7762c85ba07d515b4c66d87382870dda0be5a5ceafd1  figures/data/qwen3-4b/v1/q3_4b_armA_traj.csv
+6f406c21eea7ee25bda9c00bec301d705a3427bf2de2e6ab839ef31cc701090a  figures/data/qwen3-4b/v1/q3_4b_armB_traj.csv
+5dd37b6c7d9a3d4c6ede536aec2a869aa26afd8c67c3096a04e66355827815e9  figures/data/qwen3-4b/v1/q3_4b_gen_traj.csv
+89992f95494655a1263b1f17c098c97e68bba6aa19279f29b7d22a049096ce50  figures/rendered/historical/v1/code250k_vs_gen.png
+66e0cfc000e9583110ce5d174ab1874210eee0b026d25fd863e7f69139f78ec3  figures/rendered/historical/v1/fig_bigdata_vs_gen.png
+c88b5a91d84cfbf30f5f9b1813ae470f808e348d27f224f57b8704490f839016  figures/rendered/historical/v1/fig_exp1_forgetting.png
+2b9633ec3524f23766a72e788206c114e02d7b0da5be8f3cf8a501848fe2fb52  figures/rendered/historical/v1/fig_exp1_three_recipes.png
+e8602bc3791560060deba3ad715eedf6fd9d1ef5d76c1774d3ef5a5fb0ccb2bd  figures/rendered/historical/v1/fig_exp4_inference.png
+bf243131b8c61f12c258760307178a8d7888e41c5ef1f5a8981f37b7f639b83a  figures/rendered/historical/v1/fig_exp5_method_vs_data.png
+58be85cb2527aa44c3e8c2579831138104bc70fa7e33794e1855df76988e9d6d  figures/rendered/historical/v1/fig_forgetting_matrix.png
+73c7ceff7bf8f01c0bc3e861c689532619dc2f358f51b37a0d9cedd8b07779bc  figures/rendered/historical/v1/fig_fusion_combined.png
+cc68719d712a0bc37321c46ec7d9b6368c08fc3f4523ef1d1c3d7771e5a99a37  figures/rendered/historical/v1/fig_gen800k_loss_al.png
+ca381bfb16abfde13f5e88c917b27698155fa073e99cba70ad50235c6565a155  figures/rendered/historical/v1/fig_merged_vs_specialist.png
+16825387c4de0251447e50ba54aa99585cf8fc2790b1d0f4c8438035c40e23f0  figures/rendered/historical/v1/fig_q3_4b_traj.png
+77ae4d2ff2f302212f1252b8bd1fbdc3c0d244dfd2a1be06d093a964575eff8e  figures/rendered/historical/v1/fig_serving.png
+1b8928ad66e0c86b58bc8d9088af11647e78f4f42180e95a0a5bd96017f623e3  figures/rendered/historical/v1/fig_smalldata_vs_gen.png
+0e3e6ae946db8f2922832f3b0b09673877c0cf68a55f9027fdc93d60399b9446  figures/rendered/historical/v1/fig_specialist_overall.png
+861c7f202252b6798391fb5bc98bc88d7379e38b3c7d77d2e4fbf0f74f144085  figures/rendered/historical/v1/fig_warm_saturation.png
+eb2084ef355799b98ab08000994e54aa8966196cbb3485235a942ed8966ac32c  figures/rendered/historical/v1/gen_per_domain_AL.png
+2f484b361bed4165c4aac24a52ad66b7a3f9ea7948d7f617c764784085ac39b5  figures/rendered/historical/v1/gen_vs_merge_attn.png
+221d21d5085f280ace83fe2532a3ae52858a6787af5f28addfd480095114984b  figures/rendered/historical/v1/joint_vs_pipeline_code.png
+7bb2f21888c8c79e731add5c61021aff4be2c58a9dfc499601d902700867802c  figures/rendered/historical/v1/overfit_attention_ladder.png
+d8b24897fdce78a542aa0f9387e13299ffe3a34c679b607470ac984bddbbf10c  figures/rendered/historical/v1/overfit_loss_al.png
+77b4ad8a26a01b67b052d7374de6b11350d89ba6eace1e8d135af159ffca5c7f  figures/rendered/historical/v1/reasonmix_5arm_per_domain.png
+7c314a567c8cadfcfb602bc0ad10c328b6dca571213b88ef656bf78404350af4  figures/rendered/historical/v1/recipe_budget.png
+0415d8aa007c14b576bb25e56837b95235cf9e68afe828a775be58f265b46cb9  figures/rendered/paper-current-candidate/v1/fig_architecture_v8.drawio
+3961c8a6ccc80992197b8746453cffd1d6f87ed158ce6103a72b4db361996f66  figures/rendered/paper-current-candidate/v1/fig_architecture_v8.png
+0a8c69fabf7ca2a57a13b4dd77edf2b94bfa1fd31fd9f28a259f7b014ec07048  figures/rendered/paper-current-candidate/v1/fig_architecture_v8.svg
+dc224065a687b7631cec8da84803ec6b8b1c5c34b737dfaa4e7077c9eb04318d  figures/rendered/paper-current-candidate/v1/fig_main_results.png
+c843a6bd2b0a20aad7d6b2816277b60af03b44bfd4a1a8a75bc7ac1c74bac746  figures/rendered/paper-current-candidate/v1/fig_qwen3_4b_matched.pdf
+1243e3db541c0be5dbdacdabf6046ad065911ec2080bacd7d6e69bc9100689db  figures/rendered/paper-current-candidate/v1/fig_qwen3_4b_matched.png
+da25818d68a3da676fed16a019beadd6b12f5e4d59ade85840c23a6070240323  licenses/extended-evidence-v1/LICENSE
+01369f893c8ca40e0b1bfefc3d0a006833bd1757dc287755d1dcd9bd6e5fd7bc  manifests/extended-evidence-v1/artifact-manifest.jsonl
+ea5920812d91ae4a6e6cc503936d1c87fb833b622ef2b5bb9fb24037a4e5195e  manifests/extended-evidence-v1/release-summary.json
+fe3d7a8074dc76533eefa49e627d779650c811fb0068698546330d811b8cb3c5  recipes/plotting/v1/build_mos_architecture_drawio.py
+c8e6297def7e8dfc361a437c798c8b8586f6f0ee99772cfa915ecc325ea9ba01  recipes/plotting/v1/fig_5arm_per_domain.py
+c6ec2787808a900478e9a92fb7f3092e04835603882d39a3ad0d24b2d486e193  recipes/plotting/v1/fig_code250k_vs_gen.py
+f49d9a5dfc3f7512af3cc86f0729774344357fa7de748e9917e90a5a4bf16d84  recipes/plotting/v1/fig_exp1_forgetting.py
+9fd98d1f1bb1a6e449bf503569fd8fbd82d89f7366a811f0fd9be4e3b9ac7e95  recipes/plotting/v1/fig_exp1_three_recipes.py
+ebe9bc1be8f0de7bdd83bcc74e71314fd9ca7758802bfa660128a265c5e4f9f1  recipes/plotting/v1/fig_exp4_inference.py
+44f959ce1232a19ea3dcee93bb41df5268171e47cf2d66ed4b5e7d84312497a5  recipes/plotting/v1/fig_exp5_method_vs_data.py
+39542922ee86c359fc3c6e0153daba478ada98e8d10da5e40b8b9994a440a87f  recipes/plotting/v1/fig_fusion_combined.py
+6967a36b0a4ce87c5650409a561bc90a186eb25633501593c87e6da9460c2bd2  recipes/plotting/v1/fig_matrix_and_arms.py
+6eb66c23518bd7e2b8edc74f23ddb53635cac18ae839c4862ec9bfbc1ca8bdc5  recipes/plotting/v1/fig_merged_vs_specialist.py
+a692b7e495c6d3ceaf46857436a115ffc299375936e37f820ddda02468a724b7  recipes/plotting/v1/fig_serving_specialist.py
+d28b0d29bda5f5ee10c4ea22fa24e7b1b07e5ba2d86c57f74f58285b06e9b58e  recipes/plotting/v1/plot_main_results.py
+663f813697a0026b093a0a6ef68d3b99be3a7c159eadbe2c5b3522e645d92eec  recipes/plotting/v1/plot_mos_5x5_gains.py
+6fdf9f53c384d1ec47ed182dcf2d44987ff30bb0255bef0061fa7c723b785c0e  recipes/plotting/v1/plot_qwen3_4b_matched.py
+a7beb62c3b84008553ff9051ecc76bbd859667bf28898b182317363c8128e404  recipes/plotting/v1/plot_recipe_budget.py
+b7ebb2c81dacb6ef06dcb87fb879ea663f2327b7417606520bb30cb299ef006a  recipes/requirements-v1.txt
+dae17f81b9fa4b5dba14545e732f6731e359e80e3845f937c09fe81ef8a6e068  results/historical-al-curves/v1/a2_code/al_curve.csv
+7d56cbca48fdec9cc37fc26e701ca5da6841884f8a49e51c929090c49cb5fe6b  results/historical-al-curves/v1/a2_creative_writing/al_curve.csv
+5da7a3f0e5d955881ce5903f4a7b7567212f89400f130a574fdc0cb783169551  results/historical-al-curves/v1/a2_factual_qa/al_curve.csv
+aee9cf5c10f54cadba4b30b513c18b2651f1fbc6b10d25da719eaea74f471682  results/historical-al-curves/v1/a2_general/al_curve.csv
+4ce95c7937d1362f0593ad93531d65620a484e138237ce793ea1b3ef03099c5b  results/historical-al-curves/v1/a2_math/al_curve.csv
+e88672fadfaaaeb1088504c919208f8b125f1aecb7a14942fc5888d572714857  results/historical-al-curves/v1/a3_code/al_curve.csv
+b5a492456138dc23d09264c26ca671082eb3fb36765ec5398aa31f56d77d2cc1  results/historical-al-curves/v1/a3_creative_writing/al_curve.csv
+b7b8776a83582db069ecd349784aa52d40dcd3bbd6c100c13780a873590c4b98  results/historical-al-curves/v1/a3_factual_qa/al_curve.csv
+4073a14f31f2d26f0c1974f3e571a45c24209bf2562b2c7725d207d134047a9e  results/historical-al-curves/v1/a3_general/al_curve.csv
+07ba1e5ce58ff013aab9a96738301295f32731facc3ec5b7678802d5ef8c996f  results/historical-al-curves/v1/a3_math/al_curve.csv
+cd427bc83af9d2f892924bd1993ead327acab61c06279e5c627abe2e18cc6809  results/historical-al-curves/v1/a4_code/al_curve.csv
+c1867fc99b121cfd613ce2a9d40f8f12a13491e146f85d3175d7bf8c3591c02f  results/historical-al-curves/v1/a4_creative_writing/al_curve.csv
+cf142156c17e74d5ac67967b93995dbaf47ee34d4ab28cf65d0044846cd52785  results/historical-al-curves/v1/a4_factual_qa/al_curve.csv
+81d3dcaacdd7405966d5eb5715dd04c434e714abf3a9620ae492430065c8dbc6  results/historical-al-curves/v1/a4_general/al_curve.csv
+8e717efab00e869e764262597696297a9bfd5cd7839e0d9ddd2ba623acb0aa05  results/historical-al-curves/v1/a4_math/al_curve.csv
+c105111e04a724b5b21af90fd5e1ecd2b10820db32c9fb041e3cfe153e862f6f  results/historical-al-curves/v1/a5_code/al_curve.csv
+454d8a6f5a03494e4d8aace991965482a858e04bbcdd4c16cb07773ba3895afa  results/historical-al-curves/v1/a5_creative_writing/al_curve.csv
+43df7e88a7ddd7bbbc9471c65682b4584af117b0d1f02dc0ade07bb29395c9cf  results/historical-al-curves/v1/a5_factual_qa/al_curve.csv
+e64d977591372dca44bb442ced4a6f5c1f39464a84471bf1ebc21abe5e29bdd3  results/historical-al-curves/v1/a5_general/al_curve.csv
+b85e52e687ba97220b8cad613cc209578df53767e40ab55fd9cf2fdf5808f3b1  results/historical-al-curves/v1/a5_math/al_curve.csv
+0784056349eac36c44dc1da700b5e051dd60ebfb832254fadf0e30049bdb6644  results/historical-al-curves/v1/a6_code/al_curve.csv
+6f0c6689cc637a4046f1cd36d32c97f5df4d6931d4bbd0df7f373929a7384258  results/historical-al-curves/v1/a6_creative_writing/al_curve.csv
+8ab6645dd36b4c7ed6e150591aa796d07fca2f453083a4ea7690b19d14f0522c  results/historical-al-curves/v1/a6_factual_qa/al_curve.csv
+dcc0ad60fb88da0a1c2b124b7e6daa94fa4bea76de8ac393eac3edefd900de8d  results/historical-al-curves/v1/a6_general/al_curve.csv
+743758c18a447dc760c512d1525a6adea57a0f8086d0d0b7b8acf0a88a9b16b7  results/historical-al-curves/v1/a6_math/al_curve.csv
+c3c3784e3a7046135fe0772b9e9acedbfe45473bf125a7c69766e08118a90b9f  results/historical-al-curves/v1/a7_code/al_curve.csv
+0eab7d6c41d11ccd0653115a7077d4b9e22943bbd9512867b71eaa2e0df01327  results/historical-al-curves/v1/a7_creative_writing/al_curve.csv
+d86f5a9a467df89b88deb48595e53915114626e692200afe2b1cceb68951f7ad  results/historical-al-curves/v1/a7_factual_qa/al_curve.csv
+c66d5d71f44d502be2f4319b57d99d7ab498e470e8b29808fc7943c5cde07423  results/historical-al-curves/v1/a7_general/al_curve.csv
+07fa6f01a556aff84978b938d7210e2c32cbcea0ff28b672895169ef2096b9fa  results/historical-al-curves/v1/a7_math/al_curve.csv
+63bc1738955954356859cdca199f48a3948003f6c6ceb523e7658dc6297c4fe5  results/historical-al-curves/v1/aw0_code/al_curve.csv
+58a7c10f203fdfca83c4445aab6b8db5bc93a5ebf6c5ee39a30adea955238048  results/historical-al-curves/v1/aw1_code/al_curve.csv
+766d1a00823d6db3f2cd6a3d96e11b935d31120f95f780b6d1f8b41a53a95f82  results/historical-al-curves/v1/aw2_code/al_curve.csv
+b3b6db437eb990291582e1fe8180fb77a5fc44394ae3a4bd595b58a1fa95116e  results/historical-al-curves/v1/aw3_code/al_curve.csv
+71a499635ea64fb0e45ea136aafe69f2466b508e0b66e6b8158a681ab0fb6ffb  results/historical-al-curves/v1/aw4_code/al_curve.csv
+e135392f04e9db93600e871c8a82b556b7f380c7622db82489bf158358fb0a29  results/historical-al-curves/v1/aw5_code/al_curve.csv
+23bdd93d5bfa5c24aed9089fb1f1903927a2de4bf26629ee42194d31994fe2c3  results/historical-al-curves/v1/aw6_code/al_curve.csv
+793994b7cbf0695dcc184bf095697988aef78be0ad6c8fabb615c2f16200d652  results/historical-al-curves/v1/aw7_code/al_curve.csv
+ed2bab825dda5d70ddb6a3e16a0a86633178f1c8e3a2149031d56f11be4b05b0  results/historical-al-curves/v1/b0_code/al_curve.csv
+f56a485034b29f7e1c65c8021efe18f5980d732c4e1556d6550e93a52d55f616  results/historical-al-curves/v1/b0_creative_writing/al_curve.csv
+f33c9d7fa3985e39ac50ff1d8dad02d72e7276a08c8d43492627be23ed67c579  results/historical-al-curves/v1/b0_factual_qa/al_curve.csv
+51ea0ab13244e62257a02239ca42aade659a535afff79faafc4f37c4d6ca5316  results/historical-al-curves/v1/b0_general/al_curve.csv
+405dd523d42728d330a5a5de80f12957a15f954c41f06474c0f898f841c7b4b9  results/historical-al-curves/v1/b0_math/al_curve.csv
+97932cc47a49befbf27f7957d827a37c3b93d691d8b5692562c9fe52d964d941  results/historical-al-curves/v1/b1_code/al_curve.csv
+f8185aae5e4724c5a86c6adfeb34c16e7767f929d3d633c0d60cab0f8127128b  results/historical-al-curves/v1/b1_creative_writing/al_curve.csv
+c2031c105435638bb133232dc1bd0c8b23454a368fa6bddc457426981dfbde56  results/historical-al-curves/v1/b1_factual_qa/al_curve.csv
+0967339a7d6711ef84490506221730589487c1d21586cd1e131bb4b5e0540a11  results/historical-al-curves/v1/b1_general/al_curve.csv
+1fb5f0f30ded6faad7a1f8a02050853cb500df707853dfbfe4239016e625a6bf  results/historical-al-curves/v1/b1_math/al_curve.csv
+784723990cdf7fb9cbae73598ade59da5c5e494666187d39ccab98e2a13e843c  results/historical-al-curves/v1/b2_code/al_curve.csv
+2d7c6e582ced15922331dbd31ca93ac17476daa07af395bba9f2c4e90793ada8  results/historical-al-curves/v1/b2_creative_writing/al_curve.csv
+8a7fc49463e125c3572f43300ae6f51229d3b7a9460f8a4cd7415f4029155e0d  results/historical-al-curves/v1/b2_factual_qa/al_curve.csv
+17a09873412bd568cc0e707da3502f5413a58fc10085a640d4e6433645ad0307  results/historical-al-curves/v1/b2_general/al_curve.csv
+8cfcdd3707b7474195e9c1d55361ebf1d7798eccc1bd5d2e77a5e354a05e0272  results/historical-al-curves/v1/b2_math/al_curve.csv
+e4f9520174ba9ba33873c3c667883034cb9112e0ed8fb1e9ec6f680142ded0c5  results/historical-al-curves/v1/b2_v3_cont_curve_8x5x1x6/al_curve.csv
+c36f8911c0042f6f686cd63a43330bb2dcb007709aa8b15807eb8a32dd259341  results/historical-al-curves/v1/b2_v3_curve_8x5x1x6/al_curve.csv
+0944a8604c0bd5a5218a9a0c6a4b3ed370657b300e8941d9a5265f5b21634b54  results/historical-al-curves/v1/b3_code/al_curve.csv
+7f35dd8bac7bec1c7fd38eee49ef144cf8cd1ea6577c28ebf7b02008a0e018c4  results/historical-al-curves/v1/b3_creative_writing/al_curve.csv
+393b855f4794d72ab7fd16d2dff264c975f3407f3b781831465fd79334e705f3  results/historical-al-curves/v1/b3_factual_qa/al_curve.csv
+76db113c70a264f7f3d96bb764b2befdcbd8ddd30593babbd13be47c7ccd36c2  results/historical-al-curves/v1/b3_general/al_curve.csv
+cb2c4c321389e1f978648ac9b634532606885924336ee9eff3e9ee3df0dd8806  results/historical-al-curves/v1/b3_math/al_curve.csv
+20dcba4e0f4a97d64d795ab5e97e377d14f083ce3ca8e9febf6ab1def6bd05d8  results/historical-al-curves/v1/b4_code/al_curve.csv
+cfb4e10bdb3f8896f321403b653c01f8edc7be518376a7508ee1fe16abf5133b  results/historical-al-curves/v1/b4_creative_writing/al_curve.csv
+6889a3c240b28f1708a8aa1844fbd9e016ede9cd06201d3f3444e852d6bd18a9  results/historical-al-curves/v1/b4_factual_qa/al_curve.csv
+93e2192320c39883bb6e536275fbc5af9992c285856a98c51f745a6c2cd322f1  results/historical-al-curves/v1/b4_general/al_curve.csv
+897845f483d8fbbc7f9cf587698bc4b7a96e126f7716172bc5547f7138ba7bfd  results/historical-al-curves/v1/b4_math/al_curve.csv
+092f185463e80a3f3870785173b7047d62662919946293db45b2e432318717e5  results/historical-al-curves/v1/b5_code/al_curve.csv
+0c193663d6bf65c6ffede829949cb80bc68801c4b52e7b3b254c410cb476e8d3  results/historical-al-curves/v1/b5_creative_writing/al_curve.csv
+bd5d632a8da9392f4ff22a73c1540e8990ba72407c09eb80e5f326abd31657d3  results/historical-al-curves/v1/b5_factual_qa/al_curve.csv
+7ab5552d57bee9349fa05b10f0ec20751db6d57745f192b0fff040df83c70db7  results/historical-al-curves/v1/b5_general/al_curve.csv
+601ddb84962c2e9a12c49338d3688b534b1b6e8efc3ea1f7c47dfd8274ac3e85  results/historical-al-curves/v1/b5_math/al_curve.csv
+d5858dc1897c4d64feadf087ea74740c3657b736e50b1561e442c7d7e66a0b3a  results/historical-al-curves/v1/b6_code/al_curve.csv
+752306586324989229b99b0eb4dc42d52faca211e44337bd279f063ee27c31c8  results/historical-al-curves/v1/b6_creative_writing/al_curve.csv
+97e815c6b08c33cd7c10581fed459dfe06d2a01328d1f433908f4ec0261b469f  results/historical-al-curves/v1/b6_factual_qa/al_curve.csv
+9da6e16cbbe001a4826ad54233041c396ce4dce90e4074639f6a74ce9c4dcb62  results/historical-al-curves/v1/b6_general/al_curve.csv
+66c4a1fc30f00ec6a172c79df7bc9bfd5eb20300b4b45961387b215a3459f223  results/historical-al-curves/v1/b6_math/al_curve.csv
+9caaf63ef828990870bcfbc53c48d849cfd1b9a7db2f33b028a82f32e7246ccd  results/historical-al-curves/v1/b7_code/al_curve.csv
+16228b3deb7f7e45f190ef229186df928daaa0a182b9a7f7b5a8fe0e0bf84748  results/historical-al-curves/v1/b7_creative_writing/al_curve.csv
+6d380638c474c93a128897ea0d5e38507c8fc639cd5afdac85cfbf55e9be72ae  results/historical-al-curves/v1/b7_factual_qa/al_curve.csv
+4c00e7142537032150da6be745d880fb00fc63e7722d3f016cc7801bc9180884  results/historical-al-curves/v1/b7_general/al_curve.csv
+ea0b445042ac886293657641a64bb039fd6e2cd26a6c2b73bdf3b6446b05e567  results/historical-al-curves/v1/b7_math/al_curve.csv
+4c7950c6ca3dbe06d5674d2568b5b183b56de92bf4fa76230af5c5d045e1eabd  results/historical-al-curves/v1/c0_code/al_curve.csv
+b74b5f158b2f61a23d899aec307d8615f8a33d158032b2a896804b0bf1450514  results/historical-al-curves/v1/c0_creative_writing/al_curve.csv
+b4e8a4ec59a51f14f79275cb1341280656a6bee50653fe497c88680af5acbcee  results/historical-al-curves/v1/c0_factual_qa/al_curve.csv
+0ee93e536ed44dbcfab2d08f8f8a76e6590b4f35cc8fc1a88cb58248700b2a78  results/historical-al-curves/v1/c0_general/al_curve.csv
+607c0206b4b6e6985297e22ee36d60fc8d315d5275b5b1d10546d2ec5de1b44e  results/historical-al-curves/v1/c0_math/al_curve.csv
+d7b3403ff3fe7e8cc6e119f046911c339bc297fb9d6484ea667326c5b4993f52  results/historical-al-curves/v1/c1_code/al_curve.csv
+954ff064cf4cec980f9439cd5c7434272ffc93ecd38f0eb7fba1c7c08c1fa518  results/historical-al-curves/v1/c1_creative_writing/al_curve.csv
+db592b66feeeb1b807c0cd192b474eb6beca4b3be00d1cf9bd1faafd6175b47a  results/historical-al-curves/v1/c1_factual_qa/al_curve.csv
+803576dbe5083e7b3a487a9780cb06bb58c63e82601156afe931fd3e04d7155d  results/historical-al-curves/v1/c1_general/al_curve.csv
+eca7f51ff784366375d5e29799fc5bf316a9ef803812d14e22c066f508c1cb35  results/historical-al-curves/v1/c1_math/al_curve.csv
+65685989c83c8cde6179b3baf8725a231c5d020e326accc2f4ccb65db4ebecb5  results/historical-al-curves/v1/c2_code/al_curve.csv
+868681f811ac8721811740d3c504f35bf37d1560d67d045c86e68023b7ea3dcd  results/historical-al-curves/v1/c2_creative_writing/al_curve.csv
+d5b2c7f25d9bcfb1ee072c5529e8cdfa8122de16cf8ae5b25601a27ea4b6b684  results/historical-al-curves/v1/c2_factual_qa/al_curve.csv
+211aca725b163fc8c6db2956352f254e03005204551d0bce4fe2c11fc24166ca  results/historical-al-curves/v1/c2_general/al_curve.csv
+ba923a05b754e4009f41f452da060bf128b4af1d97b5d60c69cb0adf970964b4  results/historical-al-curves/v1/c2_math/al_curve.csv
+7a7032f0fdb54aff1bb68456bcc4619107572d9e1c7be9f04653dffe903e2c31  results/historical-al-curves/v1/c3_code/al_curve.csv
+dd88bd7aa0c7d70467c88eba8ea31c26f4db5aa82c6c485f723f1bc6a8a47f7a  results/historical-al-curves/v1/c3_creative_writing/al_curve.csv
+8790230d20b4996547849d1fee887ccf70a036fb69c3377c93ef237b74e9576f  results/historical-al-curves/v1/c3_factual_qa/al_curve.csv
+d1586a3cf87eee78a83d6fdac95a70203c4b18fa56f371b1f3f1b801fe1c27b1  results/historical-al-curves/v1/c3_general/al_curve.csv
+4ad7ce036023a5644f713848b56b428fd78ba427e77ab2aea8039105c12cc5de  results/historical-al-curves/v1/c3_math/al_curve.csv
+7155ecbb5af764aff597a5071312c4fbe7ea3b9678d438e53a84db5994d544c3  results/historical-al-curves/v1/c4_code/al_curve.csv
+7a9502461a40264469134c1602dce11605d746fb362d7946826addfabe04ee9c  results/historical-al-curves/v1/c4_creative_writing/al_curve.csv
+deec1ade9fbb0ca6120706265aebad6ed2ac5dfd1303a4235276dfe4dd86af9a  results/historical-al-curves/v1/c4_factual_qa/al_curve.csv
+e92a4cdf7a876ca2100795d03735b95741af6acd8346c11fb2baca421f9b8b13  results/historical-al-curves/v1/c4_general/al_curve.csv
+e42b09abd0da17dadf2ec15a8f3a97ebd8abe052ed1cd28971fd35c4d82979b7  results/historical-al-curves/v1/c4_math/al_curve.csv
+3b5f8477eb60ba034b09b80033824d38bc061efb987392bba95ebf82f84085b2  results/historical-al-curves/v1/c5_code/al_curve.csv
+a7c05f9e0de662069c4c327511e5a44d8c5c2600a8a3d327d23a18e91f29fc75  results/historical-al-curves/v1/c5_creative_writing/al_curve.csv
+e38286d943199cec158517b54a799287c37c50d78e996d30ec13c8099c5cb080  results/historical-al-curves/v1/c5_factual_qa/al_curve.csv
+160423ad74f8f5bf5924b853e143138763795e3e7f0432729ebf07426d708c56  results/historical-al-curves/v1/c5_general/al_curve.csv
+093d79e141db17e3f2da8411b754057585377e77dbe2ca4b0c021c18721233ce  results/historical-al-curves/v1/c5_math/al_curve.csv
+efa094eefd03080e2d1612e2780cabc4b4fe82102bae3a70d685bf1fd9a5bdce  results/historical-al-curves/v1/c6_code/al_curve.csv
+990c56cc487367e9e418bbeff9132144433f082632b7416e2d22dd36f003c689  results/historical-al-curves/v1/c6_creative_writing/al_curve.csv
+e2ce13345d49e2960fa9e52649675e0a1c1af4b61f59e5d13a2375fb806263ee  results/historical-al-curves/v1/c6_factual_qa/al_curve.csv
+5d94f62ca87701b51af5ef930fe63be1df7958e85773ccd402edeacbf59a25b1  results/historical-al-curves/v1/c6_general/al_curve.csv
+4fc731227fb3cc66f90a5152759c216faada25a65fc332d8148541b9f1c4bac0  results/historical-al-curves/v1/c6_math/al_curve.csv
+4e8ea13685b72455b24ab172fa68d71743bc71a6d7ed7462fe30e36ee16a08d7  results/historical-al-curves/v1/c7_code/al_curve.csv
+307586723ae39dda90995aded367c77aa254e3820a2ded9f42f2c62ef83f463f  results/historical-al-curves/v1/c7_creative_writing/al_curve.csv
+ecff8fdc0258485e27b1f4ebf11449d4345b6f12cae0c7ef4ecbd4c0d3705167  results/historical-al-curves/v1/c7_factual_qa/al_curve.csv
+927daac26248cf56a8e7f09bd8c6b4a19a1f841cfb6b64d3d8643b5c0e017832  results/historical-al-curves/v1/c7_general/al_curve.csv
+e4af3170b25fc05d0ad872adb322836852e15d0d40fca7b7fd4f7f7b27b8dd62  results/historical-al-curves/v1/c7_math/al_curve.csv
+1789f2c326487f364f3c14bc9ef4773d2f5e6a3ddaee65cb8b1c2847f6116d47  results/historical-al-curves/v1/d0_code/al_curve.csv
+cc4024858d44b0252ff081901cb2d710e83257679835f6d40ff8c8744db33602  results/historical-al-curves/v1/d0_code_bench/al_curve.csv
+02c34be9c790947677503f07b3873849e14458c3cf18efed4717be917dadddab  results/historical-al-curves/v1/d0_creative_writing/al_curve.csv
+3eb6cfcd4fa4a755eb9d2f9a9c04ad4a82341a5eee734be5922d0a3530568c30  results/historical-al-curves/v1/d0_creative_writing_bench/al_curve.csv
+2140012b6c2b0f52b2aada13cfad0cdae99afeb8b7936281c65e197c2f8b0fac  results/historical-al-curves/v1/d0_factual_qa/al_curve.csv
+b88004349fe0af1749dcfd00d3d1719c63079daab247bd961224ef7d5355eb20  results/historical-al-curves/v1/d0_factual_qa_bench/al_curve.csv
+511471c17e2d15eedbd3481e7a9771170fccfdeeec09916974e3c1633793ee86  results/historical-al-curves/v1/d0_general/al_curve.csv
+898934747a054fa24b146c14a5f2aba09d61152cc5c38a902cf4f56ff5af1d42  results/historical-al-curves/v1/d0_general_bench/al_curve.csv
+5c4e2ae4caf4bef8303353cde38d98055123220e68f69c6896d65296edb81eb4  results/historical-al-curves/v1/d0_math/al_curve.csv
+1e361705ccdbb2fa0507cfc3beffebddd392d36038efa3e8dc2542d281777545  results/historical-al-curves/v1/d0_math_bench/al_curve.csv
+ebf1b0e2d4ec3da18de3011f029c029a81b8b14407c9b1d6a78333af3f07f433  results/historical-al-curves/v1/d1_code/al_curve.csv
+8912e2738a3460457ee2b1968dd3dacf93e7fdaa6b22cc064359777f5f5ccf01  results/historical-al-curves/v1/d1_creative_writing/al_curve.csv
+ed8eed33bff28b02c080d46b998217298a68e630abc23e732cc47251acadb2ab  results/historical-al-curves/v1/d1_factual_qa/al_curve.csv
+91824d146176f24973bfa6fbfe8a6d55249ac4881d1d5a23935c13f1d2c51618  results/historical-al-curves/v1/d1_general/al_curve.csv
+8ccaf221637cc8cf7631ad7d9dd260adcf9f09520eff899ee72c5867f44f1714  results/historical-al-curves/v1/d1_math/al_curve.csv
+2ba1138d82f05ed5fd7bc45186cff4351328324ad93608d2c25c20ddc8fe0b2d  results/historical-al-curves/v1/d2_code/al_curve.csv
+9c0dab3f0616fb762d43aeecab1e19d292d92ba3b72e74efc6ee3c37d2547dcf  results/historical-al-curves/v1/d2_creative_writing/al_curve.csv
+c63e517f6e5970d1b2208a455e50073b5fd531cfd9484b70890577a72baa74bf  results/historical-al-curves/v1/d2_factual_qa/al_curve.csv
+849b26bbb6d4cb5ec1b2ae207745d0c8cb5433b274404e116bb4a41bd9649f96  results/historical-al-curves/v1/d2_general/al_curve.csv
+9a6dd3a37a3daf3368a2262863b82ae34e47192e25d16ab763b4dbd39f7a5799  results/historical-al-curves/v1/d2_math/al_curve.csv
+48877d7e509598be3940beed6654af434efa1de2d9e4f46eb42352ddeb22c8ec  results/historical-al-curves/v1/d3_code/al_curve.csv
+2dc5e5d683438c79532dfabbe065e694a6b9603fa5ef7e73648eee7ad0ec7138  results/historical-al-curves/v1/d3_creative_writing/al_curve.csv
+e2e5ed61a2ed83cd6e09030ce52692106a8ae7ebdbe1c18bdf58fc0abe150845  results/historical-al-curves/v1/d3_factual_qa/al_curve.csv
+f080acc44e7d4ddc6a7cc253a722d9cfb73b531e235e0605101f951fed43e47c  results/historical-al-curves/v1/d3_general/al_curve.csv
+ba98a34a8360433870e3e12475fc275a7ab6d03bcc9b94029e7fa81a8dd97f2c  results/historical-al-curves/v1/d3_math/al_curve.csv
+06430bbb7a1dde00cb11eae287cf41212f0c4270fab85d6a7b4ea9cdfcc6376e  results/historical-al-curves/v1/d4_code/al_curve.csv
+138322d96b562a197c3dfce4be96877a63d5c60da9299baf9a4e1462e0457057  results/historical-al-curves/v1/d4_creative_writing/al_curve.csv
+f6a4e0d1aabcd786e184e3845c2d8f07afff7938c8faacc6a6e5831024300e55  results/historical-al-curves/v1/d5_code/al_curve.csv
+544e7047314513d17b726f10c4a47e3069d3d48eddb779cde9bef3f3aa9acb98  results/historical-al-curves/v1/d5_creative_writing/al_curve.csv
+2c4c0da28aec12ee646a01c36e54e226f6f2f7c2598daa589e24083b753178dc  results/historical-al-curves/v1/d5_factual_qa/al_curve.csv
+db3224fa3489c952487463cb049c202e60a01ef6665d2d5a2aecc01d44c2985b  results/historical-al-curves/v1/d5_general/al_curve.csv
+e492b9a6a7e2fa83288c02c0f450739f2d30d124d6e05aa8cce23225792fdafe  results/historical-al-curves/v1/d5_math/al_curve.csv
+d6fe8d0beec5432eab9c8fe8a9b4fcbb7b08effc638f07bc8d56dd73cf18b1bb  results/historical-al-curves/v1/d6_code/al_curve.csv
+dcc8681fc1864c9af927f1acb89ca6e18cf2b8bb16738eb500779cdd79ef1321  results/historical-al-curves/v1/d6_creative_writing/al_curve.csv
+ec1c8abcab0e73ceb72b7aeeaa56cacc483e032fe97bdb27b13096f659f933a4  results/historical-al-curves/v1/d6_factual_qa/al_curve.csv
+38416012f7b76ee350ed97bb6efe3f2b3454a430a1a588e3a1007035804f26fc  results/historical-al-curves/v1/d6_general/al_curve.csv
+a9ed8256345309365b7cd216a0c587ab31c6e66d76eb9790c4b66dafd2e96859  results/historical-al-curves/v1/d6_math/al_curve.csv
+cee9296d23831df5ee2411bb8a5f95c905796418fe20293c1a7c92981cdb99fe  results/historical-al-curves/v1/d7_code/al_curve.csv
+1dc7a8e5e0b04f920984b60ff2203bcf6dac7f48a5f331ffe0189bd893387d74  results/historical-al-curves/v1/d7_creative_writing/al_curve.csv
+701dbf445877534d1817c9a7132023654f10ea50e4878c3aebd98a79614e60f5  results/historical-al-curves/v1/d7_factual_qa/al_curve.csv
+bb45782350192168b00e368c3919d1a133096e2ba93b6c583f90ea903ab4a908  results/historical-al-curves/v1/d7_general/al_curve.csv
+ca8299b3e74ed2a8b1831f7ecaf2cfecff9c8931f1396da632354b4b6e93a4d9  results/historical-al-curves/v1/d7_math/al_curve.csv
+b5f3417a3a63c110dfa9d0bd7af497940873925f4f2ac969816034d64b54908c  results/historical-al-curves/v1/gendense_l0/al_curve.csv
+cc92b9d1cef97edf64e31bfabe6ef9b2dcfcc343d7f2c3fc492187adf72fce57  results/historical-al-curves/v1/gendense_l1/al_curve.csv
+76d88265db58b5a24a79096ed0dc6dc8ca2e5e24024a15c8777e38e6e0895eee  results/historical-al-curves/v1/gendense_l2/al_curve.csv
+6e29784a0f479940051139ec7bc3b281c5d747f8169770daf008a05b64e72414  results/historical-al-curves/v1/gendense_l3/al_curve.csv
+74378cdfcbbdb0342d963b8368350ca0a2dcaf7a9e331bbedd6b67bda9e313a5  results/historical-al-curves/v1/gendense_l4/al_curve.csv
+cb8e712bf404f41ce40f3329276dcef71a9e05852fe4c230f2a21a23a4cf348c  results/historical-al-curves/v1/gendense_l5/al_curve.csv
+7cd3d94ff8987950768e4722014f3956f2af4114d1bdb428361028603b6d1088  results/historical-al-curves/v1/gendense_l6/al_curve.csv
+e36b37efcd21b27b438bf1015266172c0804f369b48d64c243ab70cc9ab34867  results/historical-al-curves/v1/gendense_l7/al_curve.csv
+edf093e7b7be16957278d8c961197b9e913be5b5ac0678df48bcd54e8c963fbf  results/historical-al-curves/v1/genep4_code/al_curve.csv
+644c907eb90c2ed4790bf0805df65af178fa4348076e6b7a5258bbe7bb5eb7e4  results/historical-al-curves/v1/genep4_math/al_curve.csv
+f575b47ecd4d248d4f4c0f4a2e6e685ee60fc9aff335f0eae2e64d06d503ff17  results/historical-al-curves/v1/merge89/al_curve.csv
+ffa68e561efb0ff335945d7f5ce2c4ce884c1cea35fbecc0acfc8f0a3a7f5a1c  results/historical-al-curves/v1/mv2c89/al_curve.csv
+acfb67b4fa59a0eb7a175f2c110e8b609eb2ff530283f5ad770e949b3d15d595  results/historical-al-curves/v1/mv3x89/al_curve.csv
+62fe8604350ac825389c66aa3282b9c5076cfac2cbd33ce35f98222c8995edfd  results/historical-al-curves/v1/mx_code_code/al_curve.csv
+471ed1c8cb6266ce5462dd90f4fd767ff224b794b6431b51207b3db10e953d86  results/historical-al-curves/v1/mx_code_creative_writing/al_curve.csv
+a0ac4e1243a51d3a5ab2aaf79595f8e89836f1f43210073c32fa36530bef3638  results/historical-al-curves/v1/mx_code_factual_qa/al_curve.csv
+75ff837f1749d0f095ca7ba242cf6d00d7817b27049efa3c2cc4b51f9c5b4a9a  results/historical-al-curves/v1/mx_code_general/al_curve.csv
+27e2d44d0e1d64adbdc9dbb02f8fdfd5849a1acb65e886bca1b663dc947b1df2  results/historical-al-curves/v1/mx_code_math/al_curve.csv
+193bbf84b8b4802b7ca15e36ff505558842da29023f3f450441b771e3244bdc0  results/historical-al-curves/v1/mx_creative_writing_code/al_curve.csv
+0c193663d6bf65c6ffede829949cb80bc68801c4b52e7b3b254c410cb476e8d3  results/historical-al-curves/v1/mx_creative_writing_creative_writing/al_curve.csv
+169dcde892e3f47236c2e5a440b18478aac40b6d578920e865124e3ce1ce729e  results/historical-al-curves/v1/mx_creative_writing_factual_qa/al_curve.csv
+79ffe500ee72208d201cde35dada407cd3120dd3761ebb53cc89107efac2420b  results/historical-al-curves/v1/mx_creative_writing_general/al_curve.csv
+4cb4168637b1d2ecfc71fe0709a3cabc8a59dbdf0d4a8894af7473afb3fe5458  results/historical-al-curves/v1/mx_creative_writing_math/al_curve.csv
+0da911fc412c8831a08db40720644b2d07d1878e8b2d9505d2c344c58ed18f32  results/historical-al-curves/v1/mx_factual_qa_code/al_curve.csv
+c22215e9aac307a3d72ec1dc4054e8c65bbfaae5e81c501d2e628b89e0ccac6a  results/historical-al-curves/v1/mx_factual_qa_creative_writing/al_curve.csv
+ebc0bf6bf21573185e2a4e494f4d2ba3a3fdefc0ab5bdcf7dd6cbb77bdb63e14  results/historical-al-curves/v1/mx_factual_qa_factual_qa/al_curve.csv
+42edb75d5a4adece22f31fc5895568f5a0195373097fd356b1f463afaf03fca7  results/historical-al-curves/v1/mx_factual_qa_general/al_curve.csv
+29dfae8c16ec355149bec331dab105867ec8741f0af6dc411d15d9074f2927ce  results/historical-al-curves/v1/mx_factual_qa_math/al_curve.csv
+ea26ed3625839230a90b0b5fb91ad8808856c13e16e1e303cf799c736c5b570d  results/historical-al-curves/v1/mx_general_code/al_curve.csv
+8e0fb5a0a6730b5f416190e30515cf294b4646c09d285f070d8d3eb15e0ce6bf  results/historical-al-curves/v1/mx_general_creative_writing/al_curve.csv
+d8767ce37119ecb19671fe82e28c340b6323d8c5fcce7e336ee004cc30472499  results/historical-al-curves/v1/mx_general_factual_qa/al_curve.csv
+127bbd05b322c4e702ca8637c1754d299986f0ef8d08a9a37f8d734bb22e7ead  results/historical-al-curves/v1/mx_general_general/al_curve.csv
+3701966065391c5c89b1dca082fd9ab93ebc4ee444b51db0ae8d05268a5a50f1  results/historical-al-curves/v1/mx_general_math/al_curve.csv
+a6b1afa6900cb541c4979a6714d42aed473565d9d646088961e2f5c4ccb6736b  results/historical-al-curves/v1/mx_math_code/al_curve.csv
+5eba966fd12aaf38b508164c325bb480d95aceec7a8fd3b359ff12084c8b8635  results/historical-al-curves/v1/mx_math_creative_writing/al_curve.csv
+42c6c251cfe5c3559195b53a9860f00e545286f24aacf56e700f5d8c24c29aa8  results/historical-al-curves/v1/mx_math_factual_qa/al_curve.csv
+0ef053311d4cae8ba0d339bf066e396bbc9624fccb00ee79a9ccf6215190f772  results/historical-al-curves/v1/mx_math_general/al_curve.csv
+601ddb84962c2e9a12c49338d3688b534b1b6e8efc3ea1f7c47dfd8274ac3e85  results/historical-al-curves/v1/mx_math_math/al_curve.csv
+ae5e4086fd51850db624e71aad1f6390564afea4c01909138b8ec518174a6ddd  results/historical-al-curves/v1/rc89/al_curve.csv
+1e682fc087dfa045ab6cd0604b6a6d60ccc9779188c383b3cce5d98922cb3348  results/historical-al-curves/v1/route89/al_curve.csv
+226d571094d6326d5c77156c8b6f9b6f23565eb0a558589f23bc047a89b59fb9  results/historical-al-curves/v1/rw89x/al_curve.csv
+d3786c58dd808b47532eadae27684b67369fd668725e3c58ab5b6d95fbb8eb0f  results/historical-al-curves/v1/rwr2_0_code/al_curve.csv
+98daad9455c1beef6de932d4c515c3f211ecce2b6db33d4195d2882f2a4ec8a0  results/historical-al-curves/v1/rwr2_1_code/al_curve.csv
+ee1dc53951abdc88505e4eb04a160a2a0c069529ae4c7921f4a161c087bc7ed7  results/historical-al-curves/v1/rwr2_2_code/al_curve.csv
+d66a15fbbacb240934dc919517d50be294ee76085c831d44adfc4131b9f9a324  results/historical-al-curves/v1/scnAg0_code/al_curve.csv
+c0d060695f2f0c87fcd74cd065b720d47548eaf06d91dbc942fd1a8aa5d179ca  results/historical-al-curves/v1/scnAg0_math/al_curve.csv
+071f37849a80671bebc47c9c9bd9020b0a150dc39185575f323621cf5808e6a2  results/historical-al-curves/v1/scnAg1_code/al_curve.csv
+f74a58358baddb9ad7038e3be19273d990e93fd1f6814d2c14b2fe9fe2a840ab  results/historical-al-curves/v1/scnAg1_math/al_curve.csv
+e81f887a2721c5e0334c956f7a4de7f31102b5a76b453e8510c34883172148d3  results/historical-al-curves/v1/scnAg2_code/al_curve.csv
+b8d5de6fafb9437dd34f8dd0a92b861e3ff4303e21d37fdd82ea0ba426563025  results/historical-al-curves/v1/scnAg2_math/al_curve.csv
+d177d4bd11c69bb5a669851dd02451a604ebeceb9d9763594d5f9f083bf75c85  results/historical-al-curves/v1/scnAg3_code/al_curve.csv
+81a5d982e68efcd144f0353c18f97d1eb5372018d1e7bcdd0ebff652b15e16f7  results/historical-al-curves/v1/scnAg3_math/al_curve.csv
+7de24f2594949bfe5ee6faa5ca99d976d72e97741ea961315aebfe1e447bd0d8  results/historical-al-curves/v1/scnAm0_math/al_curve.csv
+3cbab716c23dd154accda87f4bcf794b2fc0651752aad17d823dc39b062b1660  results/historical-al-curves/v1/scnAm1_math/al_curve.csv
+bf5646e0311035d6a58c3aa05f4c16c5a03ec9cad2a5e3f238d0870a468c932f  results/historical-al-curves/v1/scnAm2_math/al_curve.csv
+d1d8c2c472413573c1649ed6ece2cf45351f0f2801b90d3c1c3c1765de330ad0  results/historical-al-curves/v1/scnAm3_math/al_curve.csv
+a27520a897853df3e676180297494c22b731734ac6b01dd7128ab0e28744dcae  results/historical-al-curves/v1/scnAm_code/al_curve.csv
+df9636a07e3c7c635e4c25e7c4c2ec39a062c3b691ed01d7630c902de36665cf  results/historical-al-curves/v1/sd0_code/al_curve.csv
+50f86a08650e19fdec4c1705af72c2667a25ad12c21fd228f540dfc17807fd91  results/historical-al-curves/v1/sd1_code/al_curve.csv
+b966d37e18be328c18ac4488c53ca7ad314680d6878ba721b1f0b126bfde4351  results/historical-al-curves/v1/sd2_code/al_curve.csv
+595ca5c6d0f9c564c9c4d3c1858259f14ba43a6c59a2c3a980017c8dddea6d5c  results/historical-al-curves/v1/sd3_code/al_curve.csv
+49f06e20a64540e35a1882f0d6f77c2ef97b0de40221adbdd4ce5ebb2e4fa6cc  results/historical-al-curves/v1/sd4_code/al_curve.csv
+203fe4ec71adef91a5e2c932d23b23a55f16b4db72cf8c8c26c10391807247f3  results/historical-al-curves/v1/sd5_code/al_curve.csv
+93e9532b8afa55c7d04ebf21a097f5050719ae7faaf79c514a0b49fdb76433dd  results/historical-al-curves/v1/sd5pk_creative_writing/al_curve.csv
+6b48f906b7b3fc0ec47d1b4cb5a758c6779bae518405ab96773a31add496885e  results/historical-al-curves/v1/sd5pk_factual_qa/al_curve.csv
+b024f11ac64c058c715cb5652556058e912dc5b2b27e6a45f1c2a98ff9839b20  results/historical-al-curves/v1/sd5pk_general/al_curve.csv
+86a5dc6aa38a11753c1f7c543bac0be7e9508412fd9d3621b8a1dccf675a52b7  results/historical-al-curves/v1/sd5pk_math/al_curve.csv
+bd8c6ca5ee8bbc993cceb689528c6324f0c06175684ade3addd753d4e30bd124  results/historical-al-curves/v1/sd6_code/al_curve.csv
+60e6b8260e89c9efab753e867c2bde1b4a8ee54b1715a7215216849297f33fb8  results/historical-al-curves/v1/sd7_code/al_curve.csv
+e803f1bd631885a062f8cf78cfc78fd6010d5a75dd7a6575fc71256d386a5d91  results/historical-al-curves/v1/sg0_code/al_curve.csv
+99e0a6ef3217f4d18ddc84092678467918f548c52e2fac892f779dade359cad5  results/historical-al-curves/v1/sg1_code/al_curve.csv
+2312de5495b1bb06fd039a8ca49124b2ea19ae82174889c996aba3af5890cba7  results/historical-al-curves/v1/sg2_code/al_curve.csv
+3abbfc13e520c00535e6cd2ecff9aa5aa91b8e3e4f1feca133a8c4557c580dbb  results/historical-al-curves/v1/sg3_code/al_curve.csv
+86a4d3bdeca1c89716a4afc24dc1f9a56a14e8df479fc8071fca91f8b0b3da05  results/historical-al-curves/v1/sg4_code/al_curve.csv
+e5b0bec024da1370030d6c903dc14a353aa22d92a6c1a1efe2116374307265a8  results/historical-al-curves/v1/sg5_code/al_curve.csv
+182709212d475a24728d84a642a0f589d0ad3807183a40a5375a95998f9ab05b  results/historical-al-curves/v1/sg6_code/al_curve.csv
+171d83662652e87ba1819c8e3ed696d30ae974994d486f50904a8673155c15a4  results/historical-al-curves/v1/sg7_code/al_curve.csv
+763cb159181b98cffc15de07014e37bf2c378c84953a02147d860bfb76ff9f64  results/historical-al-curves/v1/sgpk_creative_writing/al_curve.csv
+94d277ef87c5eb724a44e1e4e6e4bd310c2e86415ec10069c4cd65d111baf549  results/historical-al-curves/v1/sgpk_factual_qa/al_curve.csv
+d9d8404db125c2cf0594ec8f30cc2f95fdf144eb2768dd1518dad1ad251bd08f  results/historical-al-curves/v1/sgpk_general/al_curve.csv
+566138ead9cd5d33e7434ac677ef7064c85c90587017164f9b12c89942b9f2d5  results/historical-al-curves/v1/sgpk_math/al_curve.csv
+b21b54dcc70720f4b52af10e8d85653072580c197b8020c128871e208a462f9f  results/historical-al-curves/v1/wr2be0_code/al_curve.csv
+79a52b7a1a9ad1aa7412ec94606f1e762c1b5ede13fa3ea5c2a6d6da6509200f  results/historical-al-curves/v1/wr2be1_code/al_curve.csv
+b8c3d0f74241fd61dc13cb394b7780c546160e26730a0956d1938285f096a65e  results/historical-al-curves/v1/wr2be2_code/al_curve.csv
+fe39d6da277ea0f5fcaa73778bb46c2ef01a8309487153963f92764d1dc8d780  results/historical-al-curves/v1/wr3e0_code/al_curve.csv
+dec31dbef3209b9019a41fe89f9c6a586f4a3dc0878287df741953101288a583  results/historical-al-curves/v1/wr3e1_code/al_curve.csv
+6cfca51dd583996e8e68b1f2659aa472c2445ab4e0a847ef4d6a796276ae16fd  results/historical-al-curves/v1/wr3e2_code/al_curve.csv
+7e858ff1246214c81cf70976c920436d3271f9c77bd4e13a659ec9f4d3768864  results/main-figure-r1/v1/frozen-aggregate-cells.json
+b2b335f41aa9a2381726c9d2c90aa8eb12a1a3b4f31ebfa60c31045ec168855d  verification/v1/bootstrap_b1_exact_3p2m_routed.py
+aee31b4fac56dba2507403e989daf7010a50a01fde3edb7ebca0bad57a90f2ae  verification/v1/bootstrap_b2_epoch5_router.py
+a20de477a77313a08619b1977b80e7fb7ffe83bd08bd923e0aa5c8132dd6901d  verification/v1/bootstrap_paired_al.py
+9734fc9b2b24a74ffa63708ae0ae7442962cef40f9147c14c3528c416c78d0ca  verification/v1/r1_mainfig_verify_exports.py
+f52bef510022d3ca470104bb184cdffe83ead49648dd712e64b0ea9149d901e3  verification/v1/verify_b1_exact_3p2m_routed_artifacts.py
+7427838713af5748fbfd600af2b8a8d4d50d58a7403f072c352c025b12a0e3d1  verification/v1/verify_b2_epoch5_artifacts.py
+67e2df392b3f69b5e9e8d4c0c2dddc20f7e5b79a9dcbb7f8cd3fffd278cf5a33  verification/v1/verify_b5_qwen3_4b_evidence.py

docs/extended-evidence-v1/EXCLUSIONS.md

@@ -0,0 +1,18 @@
+# Deliberate exclusions
+
+This bundle excludes:
+
+- `REPORT.md`, `EVIDENCE_LEDGER.md`, manuscript sources, and private planning;
+- raw prompt datasets, prompt text, per-prompt generations, and router groups;
+- raw benchmark/training logs and JSONL result sidecars;
+- model weights, optimizer state, activations, prepared datasets, and caches;
+- private internal training code or data;
+- files containing credentials, usernames, or absolute internal paths;
+- path-heavy R1/B1/C1/C3 manifests except the minimized aggregate R1 projection;
+- `gen_curves.py`, `r1_mainfig_preflight.py`,
+  `r1_mainfig_validate_cells.py`, and `verify_c3_qwen3_4b_intake.py`, whose
+  current forms retain internal execution topology and need a separate
+  parameterization review.
+
+Historical aggregate curves and figures are retained for reuse, but their
+presence does not promote them to current-paper evidence.

docs/extended-evidence-v1/README.md

@@ -0,0 +1,61 @@
+# MoS-DFlash extended evidence bundle v1
+
+Frozen: `2026-07-23T20:26:58Z`
+
+This public-safe bundle contains reusable aggregate acceptance-length curves,
+figure data, author-generated rendered figures, plotting recipes, and
+verification utilities from the MoS-DFlash research workspace.
+
+## Evidence boundary
+
+- `results/historical-al-curves/v1/` contains aggregate AL values only. The
+  legacy `ckpt` column is a logical run/checkpoint identifier, not a filesystem
+  path. These curves are historical experiment records and are **not all
+  current-paper claims**.
+- `results/main-figure-r1/v1/frozen-aggregate-cells.json` is a deliberately
+  minimized public projection. It retains only the complete 52/52 status and
+  the numeric panels consumed by the plotting code. All cell-level file paths,
+  prompt references, sidecars, and raw records were omitted.
+- `figures/rendered/historical/v1/` preserves historical author-generated
+  figures. `figures/rendered/paper-current-candidate/v1/` records the selected
+  paper assets at this freeze; the manuscript remains the authority for which
+  figures are finally submitted.
+- `recipes/plotting/v1/` and `verification/v1/` contain code, not raw inputs.
+  Callers must supply their own licensed datasets, model paths, and aggregate
+  sidecars where required.
+
+## Reproducible examples
+
+```bash
+python recipes/plotting/v1/plot_main_results.py \
+  --evidence results/main-figure-r1/v1/frozen-aggregate-cells.json \
+  --output /tmp/fig_main_results.png
+
+python recipes/plotting/v1/plot_mos_5x5_gains.py \
+  --evidence results/main-figure-r1/v1/frozen-aggregate-cells.json \
+  --output-dir /tmp/mos-matrices
+
+python recipes/plotting/v1/build_mos_architecture_drawio.py \
+  --output /tmp/fig_architecture_v8.drawio
+```
+
+The Qwen3-4B matched-volume plotting recipe expects the aggregate trajectory
+from the separately frozen `releases/b5-qwen3-4b-fixed-budget/` evidence lane.
+
+## Audit
+
+- `manifests/extended-evidence-v1/artifact-manifest.jsonl` maps every published
+  content object to its repository-relative source path and source SHA-256.
+- `checksums/extended-evidence-v1.sha256` freezes staged bytes.
+- Files with hard-coded Weka or `/tmp` output locations were not copied as-is.
+  Only explicit sanitized copies with relative output paths are present, and
+  their source and staged hashes differ in the manifest.
+
+No raw prompts, per-prompt generations, benchmark/training logs, private model
+or activation material, credentials, absolute internal paths, `REPORT.md`, or
+`EVIDENCE_LEDGER.md` are included.
+
+## License
+
+Code and author-generated artifacts in this bundle are released under the MIT
+license in `licenses/extended-evidence-v1/LICENSE`.

figures/data/qwen3-4b/v1/q3_4b_armA_traj.csv

@@ -0,0 +1,33 @@
+point,step,overall_al,code,math,factual_qa,creative_writing,general
+step_10000,10000,3.1315,3.1009,4.4019,2.6961,2.6091,2.8495
+step_12500,12500,3.1969,3.1276,4.5913,2.7403,2.6474,2.8777
+step_15000,15000,3.2012,3.1329,4.5902,2.7227,2.6764,2.8840
+step_17500,17500,3.2312,3.1796,4.5761,2.7718,2.6783,2.9503
+step_20000,20000,3.2405,3.1978,4.6114,2.7855,2.6783,2.9295
+step_22500,22500,3.2702,3.2066,4.6825,2.7797,2.7210,2.9610
+step_2500,2500,2.9335,2.8528,4.1414,2.5529,2.3996,2.7209
+step_25000,25000,3.2897,3.2198,4.7660,2.8254,2.6797,2.9575
+step_27500,27500,3.2950,3.2186,4.7407,2.7984,2.7394,2.9779
+step_30000,30000,3.3085,3.2391,4.7723,2.8213,2.7316,2.9781
+step_32500,32500,3.3256,3.2592,4.8172,2.8294,2.7085,3.0135
+step_35000,35000,3.3261,3.2914,4.8069,2.8414,2.6962,2.9948
+step_37500,37500,3.3239,3.2841,4.8004,2.8284,2.7176,2.9890
+step_40000,40000,3.3506,3.2863,4.8107,2.8388,2.7968,3.0202
+step_42500,42500,3.3679,3.2620,4.9353,2.8437,2.7576,3.0407
+step_45000,45000,3.3627,3.2946,4.8828,2.8456,2.7726,3.0178
+step_47500,47500,3.3761,3.3020,4.8511,2.8643,2.8334,3.0299
+step_5000,5000,3.0275,2.9729,4.2464,2.6362,2.5029,2.7791
+step_50000,50000,3.3793,3.3113,4.9204,2.8589,2.7821,3.0235
+step_52500,52500,3.3732,3.3012,4.9109,2.8472,2.7816,3.0251
+step_55000,55000,3.3787,3.3068,4.8855,2.8820,2.7892,3.0299
+step_57500,57500,3.3867,3.3318,4.8669,2.8990,2.8080,3.0275
+step_60000,60000,3.3851,3.3215,4.9434,2.8583,2.7836,3.0184
+step_62500,62500,3.3869,3.2986,4.9339,2.8783,2.8065,3.0170
+step_65000,65000,3.3844,3.3203,4.8788,2.8800,2.7953,3.0476
+step_67500,67500,3.3854,3.3001,4.8815,2.8600,2.8460,3.0393
+step_70000,70000,3.3894,3.3288,4.9029,2.8620,2.8070,3.0465
+step_72500,72500,3.3713,3.3143,4.8882,2.8801,2.7253,3.0484
+step_7500,7500,3.0940,3.0365,4.3622,2.6655,2.5707,2.8351
+epoch_0_step_24970,24970,3.2848,3.2287,4.6862,2.8052,2.7273,2.9769
+epoch_1_step_49940,49940,3.3640,3.2700,4.8682,2.8645,2.8029,3.0143
+epoch_2_step_74910,74910,3.3845,3.3085,4.9530,2.8821,2.7355,3.0432

figures/data/qwen3-4b/v1/q3_4b_armB_traj.csv

@@ -0,0 +1,33 @@
+point,step,overall_al,code,math,factual_qa,creative_writing,general
+step_10000,10000,3.3373,3.2303,4.7736,2.8494,2.7661,3.0674
+step_12500,12500,3.3604,3.2559,4.8735,2.8494,2.7591,3.0638
+step_15000,15000,3.3424,3.2609,4.7799,2.8671,2.7336,3.0708
+step_17500,17500,3.3587,3.2884,4.7953,2.8823,2.7616,3.0662
+step_20000,20000,3.3828,3.2908,4.8577,2.8816,2.7912,3.0929
+step_22500,22500,3.4117,3.3280,4.8962,2.8947,2.8319,3.1077
+step_2500,2500,3.2904,3.2040,4.6521,2.8480,2.7205,3.0272
+step_25000,25000,3.4146,3.3412,4.9544,2.8993,2.7676,3.1106
+step_27500,27500,3.3808,3.3244,4.8682,2.8768,2.7483,3.0864
+step_30000,30000,3.4306,3.3264,4.9819,2.9117,2.8277,3.1051
+step_32500,32500,3.4111,3.3200,4.9407,2.8812,2.8235,3.0901
+step_35000,35000,3.4409,3.3530,4.9654,2.9262,2.8673,3.0927
+step_37500,37500,3.4415,3.3442,4.9805,2.9096,2.8555,3.1174
+step_40000,40000,3.4502,3.3550,5.0112,2.9184,2.8487,3.1178
+step_42500,42500,3.4495,3.3484,5.0098,2.9469,2.8272,3.1151
+step_45000,45000,3.4678,3.3501,5.0866,2.9109,2.8646,3.1270
+step_47500,47500,3.4720,3.3631,5.0507,2.9503,2.8834,3.1122
+step_5000,5000,3.3201,3.2545,4.7445,2.8371,2.7114,3.0531
+step_50000,50000,3.4360,3.3625,4.9667,2.9192,2.8004,3.1313
+step_52500,52500,3.4759,3.3722,5.0564,2.9192,2.9047,3.1270
+step_55000,55000,3.4906,3.3837,5.0837,2.9562,2.8976,3.1320
+step_57500,57500,3.4824,3.3605,5.1011,2.9401,2.8743,3.1363
+step_60000,60000,3.4746,3.3761,5.0779,2.9522,2.8455,3.1212
+step_62500,62500,3.4826,3.3732,5.0982,2.9484,2.8856,3.1076
+step_65000,65000,3.4856,3.3630,5.1509,2.9388,2.8593,3.1162
+step_67500,67500,3.4704,3.3592,5.1142,2.9123,2.8272,3.1390
+step_70000,70000,3.4722,3.3882,5.0450,2.9536,2.8582,3.1158
+step_72500,72500,3.4776,3.3580,5.1156,2.9522,2.8587,3.1035
+step_7500,7500,3.3310,3.2471,4.7096,2.8579,2.7861,3.0542
+epoch_0_step_24970,24970,3.3997,3.3121,4.9177,2.9156,2.7651,3.0880
+epoch_1_step_49940,49940,3.4776,3.3692,5.0779,2.9344,2.8970,3.1094
+epoch_2_step_74910,74910,3.4775,3.3702,5.0536,2.9517,2.8796,3.1326

figures/data/qwen3-4b/v1/q3_4b_gen_traj.csv

@@ -0,0 +1,30 @@
+step,overall_al,code,math,factual_qa,creative_writing,general
+10000,3.1686,3.0954,4.4225,2.7484,2.6225,2.9544
+11250,3.1772,3.1039,4.4655,2.7516,2.6328,2.9324
+1250,2.8705,2.8119,3.8965,2.5357,2.4068,2.7016
+12500,3.2067,3.1265,4.5161,2.7495,2.6620,2.9796
+13750,3.2080,3.1213,4.5298,2.7881,2.6315,2.9692
+15000,3.2329,3.1571,4.5298,2.7828,2.6881,3.0066
+16250,3.2303,3.1638,4.5333,2.8110,2.6524,2.9912
+17500,3.2244,3.1572,4.5287,2.8225,2.6216,2.9920
+18750,3.2505,3.1740,4.5890,2.8148,2.6895,2.9849
+20000,3.2580,3.1773,4.6245,2.8278,2.6648,2.9958
+21250,3.2841,3.2001,4.6594,2.8040,2.7143,3.0429
+22500,3.2748,3.2091,4.6305,2.8277,2.6620,3.0446
+23750,3.2820,3.2021,4.6091,2.8431,2.7181,3.0378
+2500,2.9964,2.9148,4.1491,2.6310,2.5090,2.7783
+25000,3.3024,3.2186,4.6353,2.8531,2.7463,3.0587
+26250,3.2875,3.2030,4.6233,2.8247,2.7414,3.0452
+27500,3.2927,3.2635,4.6114,2.8239,2.7157,3.0490
+28750,3.2976,3.2152,4.6911,2.8471,2.7057,3.0287
+30000,3.2872,3.2221,4.6485,2.8589,2.6593,3.0472
+31250,3.2953,3.2416,4.6257,2.8402,2.6999,3.0691
+32500,3.2755,3.1940,4.6257,2.8419,2.6853,3.0304
+33750,3.3071,3.2128,4.6789,2.8465,2.7220,3.0751
+35000,3.2808,3.2014,4.6642,2.8179,2.7018,3.0183
+36250,3.2694,3.1896,4.6233,2.8381,2.6616,3.0345
+3750,3.0412,2.9734,4.2085,2.6458,2.5334,2.8447
+5000,3.0631,3.0204,4.2495,2.6940,2.5024,2.8490
+6250,3.1019,3.0239,4.3264,2.6841,2.5902,2.8851
+7500,3.1331,3.0864,4.3506,2.7112,2.5946,2.9230
+8750,3.1352,3.0659,4.3782,2.7174,2.5937,2.9209

figures/rendered/paper-current-candidate/v1/fig_architecture_v8.drawio

@@ -0,0 +1,728 @@
+<?xml version='1.0' encoding='utf-8'?>
+<mxfile host="Electron" modified="2026-07-21T00:00:00.000Z" version="30.3.14">
+  <diagram id="mos-v8" name="MoS architecture">
+    <mxGraphModel dx="0" dy="0" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="2400" pageHeight="1420" math="0" shadow="0" background="#FFFFFF">
+      <root>
+        <mxCell id="0" />
+        <mxCell id="1" parent="0" />
+        <mxCell id="v_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="700" y="20" width="2" height="650" as="geometry" />
+        </mxCell>
+        <mxCell id="v_3" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="30" y="700" width="2240" height="2" as="geometry" />
+        </mxCell>
+        <mxCell id="v_4" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2330" y="700" width="40" height="2" as="geometry" />
+        </mxCell>
+        <mxCell id="panel_A" value="A" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#263746;strokeColor=#263746;strokeWidth=1.8;fontColor=#FFFFFF;fontFamily=Helvetica;fontSize=30;fontStyle=1;align=center;verticalAlign=middle;spacing=4;ellipse;" vertex="1" parent="1">
+          <mxGeometry x="40" y="28" width="46" height="46" as="geometry" />
+        </mxCell>
+        <mxCell id="title_A" value="Construct training signals" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=32;fontColor=#263746;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="102" y="26" width="540" height="50" as="geometry" />
+        </mxCell>
+        <mxCell id="v_5" value="TRAINING GROUPS" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="40" y="88" width="400" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="domain_1" value="Domain 1" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;verticalAlign=top;spacingTop=22;" vertex="1" parent="1">
+          <mxGeometry x="40" y="130" width="120" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="v_6" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#5F879C;strokeColor=#5F879C;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_1">
+          <mxGeometry x="0" y="0" width="120" height="9" as="geometry" />
+        </mxCell>
+        <mxCell id="v_7" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_1">
+          <mxGeometry x="28" y="60" width="64" height="3" as="geometry" />
+        </mxCell>
+        <mxCell id="domain_2" value="Domain 2" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;verticalAlign=top;spacingTop=22;" vertex="1" parent="1">
+          <mxGeometry x="200" y="130" width="120" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="v_8" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#5F879C;strokeColor=#5F879C;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_2">
+          <mxGeometry x="0" y="0" width="120" height="9" as="geometry" />
+        </mxCell>
+        <mxCell id="v_9" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_2">
+          <mxGeometry x="28" y="60" width="64" height="3" as="geometry" />
+        </mxCell>
+        <mxCell id="domain_mid" value="⋯" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;verticalAlign=top;spacingTop=22;" vertex="1" parent="1">
+          <mxGeometry x="360" y="130" width="120" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="v_10" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#5F879C;strokeColor=#5F879C;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_mid">
+          <mxGeometry x="0" y="0" width="120" height="9" as="geometry" />
+        </mxCell>
+        <mxCell id="v_11" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_mid">
+          <mxGeometry x="28" y="60" width="64" height="3" as="geometry" />
+        </mxCell>
+        <mxCell id="domain_n" value="Domain N" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;verticalAlign=top;spacingTop=22;" vertex="1" parent="1">
+          <mxGeometry x="520" y="130" width="120" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="v_12" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#5F879C;strokeColor=#5F879C;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_n">
+          <mxGeometry x="0" y="0" width="120" height="9" as="geometry" />
+        </mxCell>
+        <mxCell id="v_13" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#D7E0E5;strokeColor=#D7E0E5;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="domain_n">
+          <mxGeometry x="28" y="60" width="64" height="3" as="geometry" />
+        </mxCell>
+        <mxCell id="xy_pair" value="&lt;b&gt;Request–response&lt;/b&gt;&lt;br&gt;pair (x, y)" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="90" y="250" width="280" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="offline_assignment" value="&lt;b&gt;Prompt-only&lt;/b&gt;&lt;br&gt;training label d" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=2.0;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="400" y="250" width="250" height="78" as="geometry" />
+        </mxCell>
+        <mxCell id="e_14" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.5;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.2;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="domain_1" target="xy_pair">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_15" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.5;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.4;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="domain_2" target="xy_pair">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_16" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.5;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.6;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="domain_mid" target="xy_pair">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_17" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.5;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.8;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="domain_n" target="xy_pair">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_18" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="xy_pair" target="offline_assignment">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="v_19" value="Response y" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="40" y="360" width="130" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="response_0" value="y&lt;sub&gt;1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="210" y="350" width="60" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="response_1" value="y&lt;sub&gt;2&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="280" y="350" width="60" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="response_2" value="y&lt;sub&gt;a&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFF1DD;strokeColor=#C77E26;strokeWidth=1.8;fontColor=#C77E26;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="350" y="350" width="60" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="response_3" value="y&lt;sub&gt;4&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="420" y="350" width="60" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="response_4" value="⋯" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="490" y="350" width="60" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="e_20" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="xy_pair" target="response_0">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="230" y="340" />
+              <mxPoint x="240" y="340" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="v_21" value="sampled anchor" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#C77E26;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="340" y="405" width="150" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="masked_candidate_block" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="250" y="440" width="390" height="112" as="geometry" />
+        </mxCell>
+        <mxCell id="v_22" value="Masked candidate block" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#4F9169;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="masked_candidate_block">
+          <mxGeometry x="15" y="10" width="360" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="masked_0" value="y&lt;sub&gt;a&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFF1DD;strokeColor=#C77E26;strokeWidth=1.8;fontColor=#C77E26;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="masked_candidate_block">
+          <mxGeometry x="28" y="52" width="68" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="masked_1" value="[M]" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="masked_candidate_block">
+          <mxGeometry x="108" y="52" width="68" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="masked_2" value="[M]" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="masked_candidate_block">
+          <mxGeometry x="188" y="52" width="68" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="masked_3" value="[M]" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=24;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="masked_candidate_block">
+          <mxGeometry x="268" y="52" width="68" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="e_23" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#C77E26;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.33;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="response_2" target="masked_candidate_block">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="frozen_target_train" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="40" y="560" width="180" height="110" as="geometry" />
+        </mxCell>
+        <mxCell id="v_24" value="Frozen target" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#6D7B84;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="frozen_target_train">
+          <mxGeometry x="12" y="10" width="156" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="v_25" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_target_train">
+          <mxGeometry x="28" y="52" width="124" height="8" as="geometry" />
+        </mxCell>
+        <mxCell id="v_26" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_target_train">
+          <mxGeometry x="28" y="68" width="124" height="8" as="geometry" />
+        </mxCell>
+        <mxCell id="v_27" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_target_train">
+          <mxGeometry x="28" y="84" width="124" height="8" as="geometry" />
+        </mxCell>
+        <mxCell id="frozen_feature_stack" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="250" y="570" width="190" height="92" as="geometry" />
+        </mxCell>
+        <mxCell id="v_28" value="Frozen features" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#6D7B84;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="frozen_feature_stack">
+          <mxGeometry x="10" y="8" width="170" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="v_29" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_feature_stack">
+          <mxGeometry x="20" y="46" width="150" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="v_30" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_feature_stack">
+          <mxGeometry x="20" y="59" width="136" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="v_31" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="frozen_feature_stack">
+          <mxGeometry x="20" y="72" width="122" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="train_zt_port" value="Target context&lt;br&gt;&lt;b&gt;z&lt;sub&gt;t&lt;/sub&gt;&lt;/b&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="480" y="555" width="170" height="54" as="geometry" />
+        </mxCell>
+        <mxCell id="train_q_port" value="Prompt features&lt;br&gt;&lt;b&gt;q(x)&lt;/b&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.8;fontColor=#6D7B84;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="480" y="620" width="170" height="54" as="geometry" />
+        </mxCell>
+        <mxCell id="e_32" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7B84;strokeWidth=1.7;endArrow=blockThin;endFill=1;endSize=9;exitX=0;exitY=0.55;exitDx=0;exitDy=0;entryX=0.25;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="xy_pair" target="frozen_target_train">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="20" y="295" />
+              <mxPoint x="20" y="540" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="e_33" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7B84;strokeWidth=1.7;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="frozen_target_train" target="frozen_feature_stack">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_34" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#5F879C;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.35;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="frozen_feature_stack" target="train_zt_port">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_35" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7B84;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;dashed=1;dashPattern=8 6;exitX=1;exitY=0.75;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="frozen_feature_stack" target="train_q_port">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="panel_B" value="B" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#263746;strokeColor=#263746;strokeWidth=1.8;fontColor=#FFFFFF;fontFamily=Helvetica;fontSize=30;fontStyle=1;align=center;verticalAlign=middle;spacing=4;ellipse;" vertex="1" parent="1">
+          <mxGeometry x="750" y="28" width="46" height="46" as="geometry" />
+        </mxCell>
+        <mxCell id="title_B" value="Jointly train MoS" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=32;fontColor=#263746;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="812" y="26" width="520" height="50" as="geometry" />
+        </mxCell>
+        <mxCell id="v_36" value="ONE EXPANDED DRAFT LAYER" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#355F74;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="750" y="88" width="500" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="train_h_in" value="h&lt;sub&gt;ℓ&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=30;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="760" y="250" width="90" height="66" as="geometry" />
+        </mxCell>
+        <mxCell id="train_norm_1" value="Norm" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=25;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="890" y="250" width="110" height="66" as="geometry" />
+        </mxCell>
+        <mxCell id="train_attention" value="Shared&lt;br&gt;attention" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=2.2;fontColor=#355F74;fontFamily=Helvetica;fontSize=27;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1040" y="215" width="210" height="136" as="geometry" />
+        </mxCell>
+        <mxCell id="train_plus_1" value="+" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#263746;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=32;fontStyle=1;align=center;verticalAlign=middle;spacing=4;ellipse;" vertex="1" parent="1">
+          <mxGeometry x="1280" y="257" width="52" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="train_norm_2" value="Norm" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=25;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1370" y="250" width="100" height="66" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_bank" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="1500" y="135" width="300" height="260" as="geometry" />
+        </mxCell>
+        <mxCell id="v_37" value="MLP bank" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=25;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="15" y="10" width="270" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_1" value="MLP&lt;sub&gt;ℓ,1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.1;fontColor=#6D7C87;fontFamily=Helvetica;fontSize=22;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="18" y="54" width="264" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_2" value="MLP&lt;sub&gt;ℓ,2&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.1;fontColor=#6D7C87;fontFamily=Helvetica;fontSize=22;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="18" y="92" width="264" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_3" value="MLP&lt;sub&gt;ℓ,d&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.2;fontColor=#5C477B;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="18" y="130" width="264" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_4" value="⋯" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.1;fontColor=#6D7C87;fontFamily=Helvetica;fontSize=22;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="18" y="168" width="264" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="train_mlp_5" value="MLP&lt;sub&gt;ℓ,N&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.1;fontColor=#6D7C87;fontFamily=Helvetica;fontSize=22;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="train_mlp_bank">
+          <mxGeometry x="18" y="206" width="264" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="train_plus_2" value="+" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#263746;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=32;fontStyle=1;align=center;verticalAlign=middle;spacing=4;ellipse;" vertex="1" parent="1">
+          <mxGeometry x="1840" y="257" width="52" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="train_h_out" value="h&lt;sub&gt;ℓ+1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=28;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1930" y="250" width="100" height="66" as="geometry" />
+        </mxCell>
+        <mxCell id="zt_train_0" value="z&lt;sub&gt;t&lt;/sub&gt;&lt;sup&gt;1&lt;/sup&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="790" y="145" width="56" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="zt_train_1" value="z&lt;sub&gt;t&lt;/sub&gt;&lt;sup&gt;2&lt;/sup&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="854" y="145" width="56" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="zt_train_2" value="⋯" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="918" y="145" width="56" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="train_projection" value="Target-context&lt;br&gt;projection" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=23;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1040" y="130" width="210" height="66" as="geometry" />
+        </mxCell>
+        <mxCell id="training_selector" value="Training label d" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=2.0;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1280" y="120" width="180" height="44" as="geometry" />
+        </mxCell>
+        <mxCell id="e_38" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_h_in" target="train_norm_1">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_39" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_norm_1" target="train_attention">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_40" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_attention" target="train_plus_1">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_41" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_plus_1" target="train_norm_2">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_42" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.6;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_norm_2" target="train_mlp_3">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_43" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.6;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_mlp_3" target="train_plus_2">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_44" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_plus_2" target="train_h_out">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_45" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#5F879C;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="zt_train_2" target="train_projection">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_46" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#5F879C;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_projection" target="train_attention">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_47" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;dashed=1;dashPattern=8 6;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="training_selector" target="train_mlp_3">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="1480" y="142" />
+              <mxPoint x="1480" y="280" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="e_48" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.7;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_h_in" target="train_plus_1">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="805" y="380" />
+              <mxPoint x="1305" y="380" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="v_49" value="residual" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#6D7C87;fontStyle=0;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="940" y="390" width="120" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="e_50" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=1.7;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_plus_1" target="train_plus_2">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="1305" y="420" />
+              <mxPoint x="1865" y="420" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="v_51" value="residual" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#6D7C87;fontStyle=0;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1540" y="388" width="120" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="v_52" value="Same d across layers" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#5C477B;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="760" y="445" width="240" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="same_group_0" value="MLP&lt;sub&gt;1,d&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.0;fontColor=#5C477B;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1060" y="438" width="160" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="same_group_1" value="MLP&lt;sub&gt;2,d&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.0;fontColor=#5C477B;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1270" y="438" width="160" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="same_group_2" value="MLP&lt;sub&gt;L,d&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.0;fontColor=#5C477B;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1550" y="438" width="160" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="v_53" value="⋯" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=28;fontColor=#5C477B;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1455" y="445" width="60" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="e_54" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="same_group_0" target="same_group_1">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_55" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.2;endArrow=none;endFill=0;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="same_group_1" target="v_53">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_56" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.2;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="v_53" target="same_group_2">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="initialization_inset" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FBFCFD;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="2050" y="80" width="310" height="160" as="geometry" />
+        </mxCell>
+        <mxCell id="v_57" value="Initialization" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=24;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="initialization_inset">
+          <mxGeometry x="15" y="10" width="280" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="v_58" value="Public DFlash or trained generalist" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=19;fontColor=#6D7C87;fontStyle=0;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="initialization_inset">
+          <mxGeometry x="15" y="50" width="280" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="v_59" value="↓" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#6D7C87;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="initialization_inset">
+          <mxGeometry x="135" y="80" width="40" height="20" as="geometry" />
+        </mxCell>
+        <mxCell id="v_60" value="Shared modules +&lt;br&gt;copied MLP groups" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#5C477B;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="initialization_inset">
+          <mxGeometry x="15" y="108" width="280" height="42" as="geometry" />
+        </mxCell>
+        <mxCell id="block_logits" value="Parallel predictions" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#7A62A0;strokeWidth=1.8;fontColor=#5C477B;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2070" y="280" width="270" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="block_loss" value="Block-prediction loss" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#7A62A0;strokeWidth=2.0;fontColor=#5C477B;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2070" y="360" width="270" height="52" as="geometry" />
+        </mxCell>
+        <mxCell id="block_update" value="Updates shared modules&lt;br&gt;and selected MLP" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=1.8;fontColor=#5C477B;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2070" y="440" width="270" height="62" as="geometry" />
+        </mxCell>
+        <mxCell id="e_61" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="train_h_out" target="block_logits">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_62" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="block_logits" target="block_loss">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_63" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.5;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="block_loss" target="block_update">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="v_64" value="ROUTER SUPERVISION" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#2B6D64;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="760" y="535" width="300" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="router_detached_features" value="Prompt features&lt;br&gt;(no gradient)" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.8;fontColor=#6D7B84;fontFamily=Helvetica;fontSize=19;fontStyle=1;align=center;verticalAlign=middle;spacing=4;dashed=1;dashPattern=8 6;" vertex="1" parent="1">
+          <mxGeometry x="760" y="575" width="230" height="64" as="geometry" />
+        </mxCell>
+        <mxCell id="router_pool" value="Mean pool" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.8;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1020" y="580" width="130" height="58" as="geometry" />
+        </mxCell>
+        <mxCell id="request_router_train" value="Request router" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.8;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1190" y="580" width="170" height="58" as="geometry" />
+        </mxCell>
+        <mxCell id="route_loss" value="Routing loss" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#4F8E83;strokeWidth=2.0;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1390" y="580" width="150" height="58" as="geometry" />
+        </mxCell>
+        <mxCell id="router_update" value="Router update" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.8;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1570" y="580" width="180" height="58" as="geometry" />
+        </mxCell>
+        <mxCell id="e_65" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="router_detached_features" target="router_pool">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_66" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="router_pool" target="request_router_train">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_67" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="request_router_train" target="route_loss">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_68" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="route_loss" target="router_update">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="mos_checkpoint" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#263746;strokeWidth=2.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="2070" y="535" width="270" height="125" as="geometry" />
+        </mxCell>
+        <mxCell id="v_69" value="MoS checkpoint" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="mos_checkpoint">
+          <mxGeometry x="15" y="12" width="240" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="v_70" value="Shared modules · MLP bank&lt;br&gt;Request router" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=19;fontColor=#6D7C87;fontStyle=0;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="mos_checkpoint">
+          <mxGeometry x="15" y="52" width="240" height="54" as="geometry" />
+        </mxCell>
+        <mxCell id="e_71" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0.65;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="block_update" target="mos_checkpoint">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_72" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.75;entryDx=0;entryDy=0;" edge="1" parent="1" source="router_update" target="mos_checkpoint">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="1870" y="609" />
+              <mxPoint x="1870" y="630" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="panel_C" value="C" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#263746;strokeColor=#263746;strokeWidth=1.8;fontColor=#FFFFFF;fontFamily=Helvetica;fontSize=30;fontStyle=1;align=center;verticalAlign=middle;spacing=4;ellipse;" vertex="1" parent="1">
+          <mxGeometry x="40" y="730" width="46" height="46" as="geometry" />
+        </mxCell>
+        <mxCell id="title_C" value="Request-routed speculative decoding" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=32;fontColor=#263746;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="102" y="728" width="760" height="50" as="geometry" />
+        </mxCell>
+        <mxCell id="v_73" value="ROUTE ONCE" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#2B6D64;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="40" y="792" width="220" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_prompt" value="Prompt x" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=26;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="40" y="845" width="160" height="68" as="geometry" />
+        </mxCell>
+        <mxCell id="target_prefill" value="Frozen target&lt;br&gt;prompt pass" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.8;fontColor=#6D7B84;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="240" y="835" width="230" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="prompt_features" value="Prompt features" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.8;fontColor=#6D7B84;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;verticalAlign=top;spacingTop=10;" vertex="1" parent="1">
+          <mxGeometry x="510" y="835" width="180" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="prompt_feature_strip_0" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="prompt_features">
+          <mxGeometry x="25" y="48" width="130" height="5" as="geometry" />
+        </mxCell>
+        <mxCell id="prompt_feature_strip_1" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="prompt_features">
+          <mxGeometry x="25" y="58" width="110" height="5" as="geometry" />
+        </mxCell>
+        <mxCell id="prompt_feature_strip_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="prompt_features">
+          <mxGeometry x="25" y="68" width="90" height="5" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_mean_pool" value="Mean pool" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.8;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="730" y="845" width="150" height="68" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_request_router" value="Request router" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.8;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="920" y="845" width="190" height="68" as="geometry" />
+        </mxCell>
+        <mxCell id="route_choice" value="Predicted&lt;br&gt;group d&lt;sub&gt;pred&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=2.2;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1150" y="835" width="130" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="route_lock" value="LOCK" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#4F8E83;strokeWidth=2.2;fontColor=#2B6D64;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="1320" y="845" width="90" height="68" as="geometry" />
+        </mxCell>
+        <mxCell id="e_74" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.1;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_prompt" target="target_prefill">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_75" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7B84;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="target_prefill" target="prompt_features">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_76" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="prompt_features" target="infer_mean_pool">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_77" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_mean_pool" target="infer_request_router">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_78" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_request_router" target="route_choice">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_79" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F8E83;strokeWidth=1.9;endArrow=none;endFill=0;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="route_choice" target="route_lock">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="v_80" value="One decision per request" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=25;fontColor=#2B6D64;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1460" y="825" width="360" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="v_81" value="cycle 1   —   cycle 2   —   ⋯   —   cycle T" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#2B6D64;fontStyle=0;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1460" y="866" width="520" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="v_82" value="Fixed for all layers and cycles" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#2B6D64;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1460" y="904" width="420" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="loaded_checkpoint" value="Trained MoS checkpoint" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FBFCFD;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#6D7C87;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2110" y="835" width="250" height="72" as="geometry" />
+        </mxCell>
+        <mxCell id="e_83" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7C87;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;dashed=1;dashPattern=8 6;exitX=0.75;exitY=1;exitDx=0;exitDy=0;entryX=0.75;entryY=0;entryDx=0;entryDy=0;" edge="1" parent="1" source="mos_checkpoint" target="loaded_checkpoint">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="2300" y="690" />
+              <mxPoint x="2300" y="810" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="v_84" value="DATA PATH" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=23;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="40" y="950" width="180" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_context" value="Target context&lt;br&gt;&lt;b&gt;z&lt;sub&gt;t&lt;/sub&gt;&lt;/b&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.8;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="40" y="1000" width="190" height="70" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_masked" value="Masked block&lt;br&gt;&lt;b&gt;[M] [M] [M]&lt;/b&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="40" y="1110" width="190" height="70" as="geometry" />
+        </mxCell>
+        <mxCell id="e_85" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#5F879C;strokeWidth=1.8;endArrow=blockThin;endFill=1;endSize=9;exitX=0.05;exitY=1;exitDx=0;exitDy=0;entryX=1;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="target_prefill" target="infer_context">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="250" y="975" />
+              <mxPoint x="250" y="1035" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="selected_path" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FBFCFD;strokeColor=#263746;strokeWidth=2.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="290" y="950" width="1220" height="280" as="geometry" />
+        </mxCell>
+        <mxCell id="v_86" value="Selected MoS path" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=27;fontColor=#263746;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="20" y="12" width="300" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="v_87" value="fixed group d&lt;sub&gt;pred&lt;/sub&gt; for the whole request" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#5C477B;fontStyle=1;align=right;verticalAlign=middle;spacing=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="670" y="14" width="520" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_1_layer" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="20" y="62" width="340" height="170" as="geometry" />
+        </mxCell>
+        <mxCell id="v_88" value="Layer 1" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=25;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="infer_1_layer">
+          <mxGeometry x="15" y="10" width="310" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_1_shared" value="Shared&lt;br&gt;modules" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=2.0;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_1_layer">
+          <mxGeometry x="18" y="60" width="165" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_1_selected" value="MLP&lt;sub&gt;1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.2;fontColor=#5C477B;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_1_layer">
+          <mxGeometry x="205" y="64" width="115" height="56" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_1_inactive_1" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_1_layer">
+          <mxGeometry x="215" y="132" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_1_inactive_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_1_layer">
+          <mxGeometry x="215" y="145" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="e_89" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.3;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_1_shared" target="infer_1_selected">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_2_layer" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="420" y="62" width="340" height="170" as="geometry" />
+        </mxCell>
+        <mxCell id="v_90" value="Layer 2" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=25;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="infer_2_layer">
+          <mxGeometry x="15" y="10" width="310" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_2_shared" value="Shared&lt;br&gt;modules" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=2.0;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_2_layer">
+          <mxGeometry x="18" y="60" width="165" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_2_selected" value="MLP&lt;sub&gt;2&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.2;fontColor=#5C477B;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_2_layer">
+          <mxGeometry x="205" y="64" width="115" height="56" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_2_inactive_1" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_2_layer">
+          <mxGeometry x="215" y="132" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_2_inactive_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_2_layer">
+          <mxGeometry x="215" y="145" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="e_91" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.3;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_2_shared" target="infer_2_selected">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_l_layer" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="820" y="62" width="340" height="170" as="geometry" />
+        </mxCell>
+        <mxCell id="v_92" value="Layer L" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=25;fontColor=#263746;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="infer_l_layer">
+          <mxGeometry x="15" y="10" width="310" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_l_shared" value="Shared&lt;br&gt;modules" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=2.0;fontColor=#355F74;fontFamily=Helvetica;fontSize=22;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_l_layer">
+          <mxGeometry x="18" y="60" width="165" height="88" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_l_selected" value="MLP&lt;sub&gt;L&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=2.2;fontColor=#5C477B;fontFamily=Helvetica;fontSize=20;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_l_layer">
+          <mxGeometry x="205" y="64" width="115" height="56" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_l_inactive_1" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_l_layer">
+          <mxGeometry x="215" y="132" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="infer_l_inactive_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F7F5FA;strokeColor=#D7E0E5;strokeWidth=1.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="infer_l_layer">
+          <mxGeometry x="215" y="145" width="95" height="7" as="geometry" />
+        </mxCell>
+        <mxCell id="e_93" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.3;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_l_shared" target="infer_l_selected">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_94" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#5F879C;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.3;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_context" target="infer_1_shared">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="260" y="1035" />
+              <mxPoint x="290" y="1035" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="e_95" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#4F9169;strokeWidth=2.0;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.75;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_masked" target="infer_1_shared">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="260" y="1145" />
+              <mxPoint x="290" y="1145" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="e_96" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.6;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_1_selected" target="infer_2_shared">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_97" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.6;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_2_selected" target="infer_l_shared">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="v_98" value="Other MLP groups are not evaluated or merged" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#6D7C87;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="selected_path">
+          <mxGeometry x="250" y="242" width="720" height="26" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_block" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#7A62A0;strokeWidth=2.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="1550" y="970" width="340" height="210" as="geometry" />
+        </mxCell>
+        <mxCell id="v_99" value="Candidate block (parallel)" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#5C477B;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="candidate_block">
+          <mxGeometry x="15" y="12" width="310" height="34" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_tokens_0" value="t&lt;sub&gt;1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="18" y="70" width="48" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_tokens_1" value="t&lt;sub&gt;2&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="76" y="70" width="48" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_tokens_2" value="t&lt;sub&gt;3&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="134" y="70" width="48" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_tokens_3" value="⋯" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="192" y="70" width="48" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_tokens_4" value="t&lt;sub&gt;16&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="250" y="70" width="48" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="candidate_parallel_bus" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#7A62A0;strokeColor=#7A62A0;strokeWidth=0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="candidate_block">
+          <mxGeometry x="35" y="166" width="270" height="3" as="geometry" />
+        </mxCell>
+        <mxCell id="e_100" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;exitX=0.16666666666666666;exitY=0;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="candidate_parallel_bus" target="candidate_tokens_0">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_101" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=0;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="candidate_parallel_bus" target="candidate_tokens_2">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_102" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;exitX=0.8333333333333334;exitY=0;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="candidate_parallel_bus" target="candidate_tokens_4">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="e_103" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#7A62A0;strokeWidth=2.6;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="infer_l_selected" target="candidate_parallel_bus">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="1530" y="1110" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="target_verification" value="" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=2.0;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;container=1;pointerEvents=0;" vertex="1" parent="1">
+          <mxGeometry x="1910" y="950" width="320" height="230" as="geometry" />
+        </mxCell>
+        <mxCell id="v_104" value="Target verification" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=24;fontColor=#6D7B84;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="target_verification">
+          <mxGeometry x="15" y="12" width="290" height="36" as="geometry" />
+        </mxCell>
+        <mxCell id="verify_tokens_0" value="t&lt;sub&gt;1&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="target_verification">
+          <mxGeometry x="32" y="66" width="54" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="verify_tokens_1" value="t&lt;sub&gt;2&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="target_verification">
+          <mxGeometry x="96" y="66" width="54" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="verify_tokens_2" value="t&lt;sub&gt;3&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF4ED;strokeColor=#4F9169;strokeWidth=1.8;fontColor=#4F9169;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="target_verification">
+          <mxGeometry x="160" y="66" width="54" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="verify_tokens_3" value="t&lt;sub&gt;4&lt;/sub&gt;" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFF1DD;strokeColor=#C77E26;strokeWidth=1.8;fontColor=#C77E26;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="target_verification">
+          <mxGeometry x="224" y="66" width="54" height="48" as="geometry" />
+        </mxCell>
+        <mxCell id="v_105" value="✓       ✓       ✓" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=24;fontColor=#4F9169;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="target_verification">
+          <mxGeometry x="40" y="122" width="190" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="v_106" value="fix" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=19;fontColor=#C77E26;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="target_verification">
+          <mxGeometry x="245" y="122" width="45" height="30" as="geometry" />
+        </mxCell>
+        <mxCell id="v_107" value="Accept matching prefix" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#4F9169;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="target_verification">
+          <mxGeometry x="25" y="164" width="270" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="v_108" value="Correct first mismatch" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=20;fontColor=#C77E26;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="target_verification">
+          <mxGeometry x="25" y="198" width="270" height="28" as="geometry" />
+        </mxCell>
+        <mxCell id="e_109" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.1;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="candidate_block" target="target_verification">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="response_output" value="Response" style="rounded=1;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#FFFFFF;strokeColor=#D7E0E5;strokeWidth=1.8;fontColor=#263746;fontFamily=Helvetica;fontSize=21;fontStyle=1;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="2250" y="1025" width="140" height="72" as="geometry" />
+        </mxCell>
+        <mxCell id="e_110" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#263746;strokeWidth=2.1;endArrow=blockThin;endFill=1;endSize=9;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="target_verification" target="response_output">
+          <mxGeometry relative="1" as="geometry" />
+        </mxCell>
+        <mxCell id="cycle_note" value="Accepted or corrected tokens start the next verification cycle" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#6D7C87;fontStyle=1;align=center;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="560" y="1260" width="900" height="32" as="geometry" />
+        </mxCell>
+        <mxCell id="e_111" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;convertToSvg=1;strokeColor=#6D7B84;strokeWidth=1.6;endArrow=blockThin;endFill=1;endSize=9;exitX=0.5;exitY=1;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="target_verification" target="infer_context">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="2070" y="1310" />
+              <mxPoint x="20" y="1310" />
+              <mxPoint x="20" y="1035" />
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="legend_swatch_0" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EAF2F6;strokeColor=#5F879C;strokeWidth=1.6;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="40" y="1360" width="34" height="24" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_label_0" value="shared" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="86" y="1354" width="200" height="36" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_swatch_1" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#EEEAF5;strokeColor=#7A62A0;strokeWidth=1.6;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="300" y="1360" width="34" height="24" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_label_1" value="selected MLP" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="346" y="1354" width="200" height="36" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_swatch_2" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#E9F3F1;strokeColor=#4F8E83;strokeWidth=1.6;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="560" y="1360" width="34" height="24" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_label_2" value="routing" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="606" y="1354" width="200" height="36" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_swatch_3" value="" style="rounded=0;arcSize=10;whiteSpace=wrap;html=1;convertToSvg=1;fillColor=#F1F4F5;strokeColor=#9AA7AF;strokeWidth=1.6;fontColor=#263746;fontFamily=Helvetica;fontSize=24;fontStyle=0;align=center;verticalAlign=middle;spacing=4;" vertex="1" parent="1">
+          <mxGeometry x="820" y="1360" width="34" height="24" as="geometry" />
+        </mxCell>
+        <mxCell id="legend_label_3" value="frozen / inactive" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=21;fontColor=#6D7C87;fontStyle=1;align=left;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="866" y="1354" width="200" height="36" as="geometry" />
+        </mxCell>
+        <mxCell id="exactness_note" value="Exact verification preserves the target distribution" style="text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;fontFamily=Helvetica;fontSize=22;fontColor=#6D7C87;fontStyle=1;align=right;verticalAlign=middle;spacing=0;" vertex="1" parent="1">
+          <mxGeometry x="1500" y="1354" width="860" height="36" as="geometry" />
+        </mxCell>
+      </root>
+    </mxGraphModel>
+  </diagram>
+</mxfile>

licenses/extended-evidence-v1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 sgl-project
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

manifests/extended-evidence-v1/release-summary.json

@@ -0,0 +1,34 @@
+{
+  "artifact_type_counts": {
+    "aggregate-al-curve": 249,
+    "aggregate-main-figure-evidence": 1,
+    "documentation": 2,
+    "environment-recipe": 1,
+    "figure-data": 3,
+    "license": 1,
+    "plotting-recipe": 15,
+    "rendered-figure": 22,
+    "rendered-figure-or-source": 6,
+    "verification-recipe": 7
+  },
+  "bytes": 5833027,
+  "files": 307,
+  "freeze_id": "extended-evidence-v1",
+  "freeze_time": "2026-07-23T20:26:58Z",
+  "license": "mit",
+  "safety_scan": {
+    "absolute_internal_paths": "pass",
+    "credentials": "pass",
+    "private_internal_material": "absent",
+    "raw_logs": "absent",
+    "raw_prompts_or_generations": "absent"
+  },
+  "transformations": {
+    "direct-copy": 293,
+    "generated-from-selected-script-imports": 1,
+    "generated-public-documentation": 2,
+    "sanitized-copy; replaced 1 hard-coded output path(s) with relative output path(s)": 9,
+    "sanitized-copy; replaced 2 hard-coded output path(s) with relative output path(s)": 1,
+    "sanitized-json-projection; retained aggregate panels and completion metadata only; removed cells, run_root, sidecars, prompt references, source checkpoints, and file paths": 1
+  }
+}

recipes/plotting/v1/build_mos_architecture_drawio.py

@@ -0,0 +1,856 @@
+#!/usr/bin/env python3
+"""Build a clean, editable draw.io source for the MoS architecture figure.
+
+The script only writes draw.io XML. Export and visual verification are run on
+the server so the local desktop environment never invokes the draw.io CLI.
+"""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+import xml.etree.ElementTree as ET
+
+
+W, H = 2400, 1420
+
+INK = "#263746"
+MUTED = "#6D7C87"
+RULE = "#D7E0E5"
+SURFACE = "#FBFCFD"
+WHITE = "#FFFFFF"
+
+SHARED = "#5F879C"
+SHARED_DARK = "#355F74"
+SHARED_FILL = "#EAF2F6"
+
+MLP = "#7A62A0"
+MLP_DARK = "#5C477B"
+MLP_FILL = "#EEEAF5"
+MLP_FAINT = "#F7F5FA"
+
+ROUTE = "#4F8E83"
+ROUTE_DARK = "#2B6D64"
+ROUTE_FILL = "#E9F3F1"
+
+FROZEN = "#9AA7AF"
+FROZEN_DARK = "#6D7B84"
+FROZEN_FILL = "#F1F4F5"
+
+GOOD = "#4F9169"
+GOOD_FILL = "#EAF4ED"
+AMBER = "#C77E26"
+AMBER_FILL = "#FFF1DD"
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--output", type=Path, required=True)
+    return parser.parse_args()
+
+
+class Diagram:
+    def __init__(self) -> None:
+        self.mxfile = ET.Element(
+            "mxfile",
+            {"host": "Electron", "modified": "2026-07-21T00:00:00.000Z", "version": "30.3.14"},
+        )
+        page = ET.SubElement(self.mxfile, "diagram", {"id": "mos-v8", "name": "MoS architecture"})
+        self.model = ET.SubElement(
+            page,
+            "mxGraphModel",
+            {
+                "dx": "0",
+                "dy": "0",
+                "grid": "1",
+                "gridSize": "10",
+                "guides": "1",
+                "tooltips": "1",
+                "connect": "1",
+                "arrows": "1",
+                "fold": "1",
+                "page": "1",
+                "pageScale": "1",
+                "pageWidth": str(W),
+                "pageHeight": str(H),
+                "math": "0",
+                "shadow": "0",
+                "background": WHITE,
+            },
+        )
+        self.root = ET.SubElement(self.model, "root")
+        ET.SubElement(self.root, "mxCell", {"id": "0"})
+        ET.SubElement(self.root, "mxCell", {"id": "1", "parent": "0"})
+        self.counter = 2
+
+    def new_id(self, prefix: str) -> str:
+        cell_id = f"{prefix}_{self.counter}"
+        self.counter += 1
+        return cell_id
+
+    def vertex(
+        self,
+        value: str,
+        x: int,
+        y: int,
+        w: int,
+        h: int,
+        style: str,
+        *,
+        cell_id: str | None = None,
+        parent: str = "1",
+    ) -> str:
+        cell_id = cell_id or self.new_id("v")
+        cell = ET.SubElement(
+            self.root,
+            "mxCell",
+            {"id": cell_id, "value": value, "style": style, "vertex": "1", "parent": parent},
+        )
+        ET.SubElement(
+            cell,
+            "mxGeometry",
+            {"x": str(x), "y": str(y), "width": str(w), "height": str(h), "as": "geometry"},
+        )
+        return cell_id
+
+    def edge(
+        self,
+        source: str,
+        target: str,
+        *,
+        color: str = INK,
+        width: float = 2.0,
+        dashed: bool = False,
+        exit_xy: tuple[float, float] | None = None,
+        entry_xy: tuple[float, float] | None = None,
+        points: list[tuple[int, int]] | None = None,
+        end_arrow: str = "blockThin",
+        cell_id: str | None = None,
+    ) -> str:
+        cell_id = cell_id or self.new_id("e")
+        style = [
+            "edgeStyle=orthogonalEdgeStyle",
+            "rounded=1",
+            "orthogonalLoop=1",
+            "jettySize=auto",
+            "html=1",
+            "convertToSvg=1",
+            f"strokeColor={color}",
+            f"strokeWidth={width}",
+            f"endArrow={end_arrow}",
+            "endFill=1" if end_arrow != "none" else "endFill=0",
+            "endSize=9",
+        ]
+        if dashed:
+            style.extend(["dashed=1", "dashPattern=8 6"])
+        if exit_xy is not None:
+            style.extend([f"exitX={exit_xy[0]}", f"exitY={exit_xy[1]}", "exitDx=0", "exitDy=0"])
+        if entry_xy is not None:
+            style.extend([f"entryX={entry_xy[0]}", f"entryY={entry_xy[1]}", "entryDx=0", "entryDy=0"])
+        cell = ET.SubElement(
+            self.root,
+            "mxCell",
+            {
+                "id": cell_id,
+                "value": "",
+                "style": ";".join(style) + ";",
+                "edge": "1",
+                "parent": "1",
+                "source": source,
+                "target": target,
+            },
+        )
+        geom = ET.SubElement(cell, "mxGeometry", {"relative": "1", "as": "geometry"})
+        if points:
+            array = ET.SubElement(geom, "Array", {"as": "points"})
+            for px, py in points:
+                ET.SubElement(array, "mxPoint", {"x": str(px), "y": str(py)})
+        return cell_id
+
+    def save(self, output: Path) -> None:
+        ET.indent(self.mxfile, space="  ")
+        output.parent.mkdir(parents=True, exist_ok=True)
+        ET.ElementTree(self.mxfile).write(output, encoding="utf-8", xml_declaration=True)
+
+
+def rect_style(
+    fill: str = WHITE,
+    stroke: str = RULE,
+    *,
+    font: str = INK,
+    size: int = 24,
+    bold: bool = False,
+    rounded: int = 1,
+    stroke_width: float = 1.8,
+    align: str = "center",
+    dashed: bool = False,
+    extra: str = "",
+) -> str:
+    items = [
+        f"rounded={rounded}",
+        "arcSize=10",
+        "whiteSpace=wrap",
+        "html=1",
+        "convertToSvg=1",
+        f"fillColor={fill}",
+        f"strokeColor={stroke}",
+        f"strokeWidth={stroke_width}",
+        f"fontColor={font}",
+        "fontFamily=Helvetica",
+        f"fontSize={size}",
+        f"fontStyle={1 if bold else 0}",
+        f"align={align}",
+        "verticalAlign=middle",
+        "spacing=4",
+    ]
+    if dashed:
+        items.extend(["dashed=1", "dashPattern=8 6"])
+    if extra:
+        items.append(extra.rstrip(";"))
+    return ";".join(items) + ";"
+
+
+def text_style(size: int = 24, *, font: str = INK, bold: bool = False, align: str = "left") -> str:
+    return (
+        "text;html=1;convertToSvg=1;strokeColor=none;fillColor=none;whiteSpace=wrap;"
+        f"fontFamily=Helvetica;fontSize={size};fontColor={font};"
+        f"fontStyle={1 if bold else 0};align={align};verticalAlign=middle;spacing=0;"
+    )
+
+
+def panel_title(d: Diagram, letter: str, title: str, x: int, y: int, width: int) -> None:
+    d.vertex(
+        letter,
+        x,
+        y,
+        46,
+        46,
+        rect_style(INK, INK, font=WHITE, size=30, bold=True, extra="ellipse"),
+        cell_id=f"panel_{letter}",
+    )
+    d.vertex(title, x + 62, y - 2, width, 50, text_style(32, bold=True), cell_id=f"title_{letter}")
+
+
+def add_separator(d: Diagram, x: int, y: int, w: int, h: int = 2) -> None:
+    d.vertex("", x, y, w, h, rect_style(RULE, RULE, rounded=0, stroke_width=0))
+
+
+def token_row(
+    d: Diagram,
+    prefix: str,
+    x: int,
+    y: int,
+    labels: list[str],
+    kinds: list[str],
+    *,
+    cell_w: int = 56,
+    cell_h: int = 48,
+    gap: int = 8,
+    parent: str = "1",
+    font_size: int = 24,
+) -> list[str]:
+    palette = {
+        "plain": (WHITE, RULE, INK),
+        "shared": (SHARED_FILL, SHARED, SHARED_DARK),
+        "anchor": (AMBER_FILL, AMBER, AMBER),
+        "mask": (GOOD_FILL, GOOD, GOOD),
+        "frozen": (FROZEN_FILL, FROZEN, FROZEN_DARK),
+        "good": (GOOD_FILL, GOOD, GOOD),
+        "correct": (AMBER_FILL, AMBER, AMBER),
+    }
+    ids: list[str] = []
+    for idx, (label, kind) in enumerate(zip(labels, kinds)):
+        fill, stroke, font = palette[kind]
+        ids.append(
+            d.vertex(
+                label,
+                x + idx * (cell_w + gap),
+                y,
+                cell_w,
+                cell_h,
+                rect_style(fill, stroke, font=font, size=font_size, bold=True),
+                cell_id=f"{prefix}_{idx}",
+                parent=parent,
+            )
+        )
+    return ids
+
+
+def document_card(d: Diagram, label: str, x: int, y: int, cell_id: str) -> str:
+    card = d.vertex(
+        label,
+        x,
+        y,
+        120,
+        78,
+        rect_style(
+            WHITE,
+            RULE,
+            size=19,
+            bold=True,
+            extra="container=1;pointerEvents=0;verticalAlign=top;spacingTop=22",
+        ),
+        cell_id=cell_id,
+    )
+    d.vertex("", 0, 0, 120, 9, rect_style(SHARED, SHARED, rounded=0, stroke_width=0), parent=card)
+    d.vertex("", 28, 60, 64, 3, rect_style(RULE, RULE, rounded=0, stroke_width=0), parent=card)
+    return card
+
+
+def frozen_target(d: Diagram, x: int, y: int, w: int, h: int, cell_id: str) -> str:
+    box = d.vertex(
+        "",
+        x,
+        y,
+        w,
+        h,
+        rect_style(FROZEN_FILL, FROZEN, extra="container=1;pointerEvents=0"),
+        cell_id=cell_id,
+    )
+    d.vertex("Frozen target", 12, 10, w - 24, 30, text_style(22, font=FROZEN_DARK, bold=True, align="center"), parent=box)
+    for idx in range(3):
+        d.vertex(
+            "",
+            28,
+            52 + idx * 16,
+            w - 56,
+            8,
+            rect_style(WHITE, RULE, rounded=0, stroke_width=1),
+            parent=box,
+        )
+    return box
+
+
+def feature_stack(d: Diagram, x: int, y: int) -> str:
+    stack = d.vertex(
+        "",
+        x,
+        y,
+        190,
+        92,
+        rect_style(WHITE, RULE, extra="container=1;pointerEvents=0"),
+        cell_id="frozen_feature_stack",
+    )
+    d.vertex("Frozen features", 10, 8, 170, 28, text_style(20, font=FROZEN_DARK, bold=True, align="center"), parent=stack)
+    for idx, width in enumerate([150, 136, 122]):
+        d.vertex(
+            "",
+            20,
+            46 + idx * 13,
+            width,
+            7,
+            rect_style(FROZEN_FILL, FROZEN, rounded=0, stroke_width=1),
+            parent=stack,
+        )
+    return stack
+
+
+def draw_panel_a(d: Diagram) -> dict[str, str]:
+    panel_title(d, "A", "Construct training signals", 40, 28, 540)
+    d.vertex("TRAINING GROUPS", 40, 88, 400, 34, text_style(23, font=MUTED, bold=True))
+
+    cards = [
+        document_card(d, "Domain 1", 40, 130, "domain_1"),
+        document_card(d, "Domain 2", 200, 130, "domain_2"),
+        document_card(d, "⋯", 360, 130, "domain_mid"),
+        document_card(d, "Domain N", 520, 130, "domain_n"),
+    ]
+    pair = d.vertex(
+        "<b>Request–response</b><br>pair (x, y)",
+        90,
+        250,
+        280,
+        78,
+        rect_style(WHITE, RULE, size=22, bold=True),
+        cell_id="xy_pair",
+    )
+    assignment = d.vertex(
+        "<b>Prompt-only</b><br>training label d",
+        400,
+        250,
+        250,
+        78,
+        rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=21, bold=True, stroke_width=2.0),
+        cell_id="offline_assignment",
+    )
+    for idx, card in enumerate(cards):
+        d.edge(card, pair, width=1.5, exit_xy=(0.5, 1), entry_xy=((idx + 1) / 5, 0))
+    d.edge(pair, assignment, color=ROUTE, width=2.0, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+
+    d.vertex("Response y", 40, 360, 130, 32, text_style(22, font=MUTED, bold=True))
+    response = token_row(
+        d,
+        "response",
+        210,
+        350,
+        ["y<sub>1</sub>", "y<sub>2</sub>", "y<sub>a</sub>", "y<sub>4</sub>", "⋯"],
+        ["plain", "plain", "anchor", "plain", "plain"],
+        cell_w=60,
+        gap=10,
+        font_size=24,
+    )
+    d.edge(pair, response[0], width=1.8, exit_xy=(0.5, 1), entry_xy=(0.5, 0), points=[(230, 340), (240, 340)])
+    d.vertex("sampled anchor", 340, 405, 150, 28, text_style(20, font=AMBER, bold=True, align="center"))
+
+    masked_box = d.vertex(
+        "",
+        250,
+        440,
+        390,
+        112,
+        rect_style(WHITE, GOOD, stroke_width=1.8, extra="container=1;pointerEvents=0"),
+        cell_id="masked_candidate_block",
+    )
+    d.vertex("Masked candidate block", 15, 10, 360, 30, text_style(22, font=GOOD, bold=True, align="center"), parent=masked_box)
+    token_row(
+        d,
+        "masked",
+        28,
+        52,
+        ["y<sub>a</sub>", "[M]", "[M]", "[M]"],
+        ["anchor", "mask", "mask", "mask"],
+        cell_w=68,
+        gap=12,
+        parent=masked_box,
+        font_size=24,
+    )
+    d.edge(response[2], masked_box, color=AMBER, width=2.0, exit_xy=(0.5, 1), entry_xy=(0.33, 0))
+
+    target = frozen_target(d, 40, 560, 180, 110, "frozen_target_train")
+    features = feature_stack(d, 250, 570)
+    z_port = d.vertex(
+        "Target context<br><b>z<sub>t</sub></b>",
+        480,
+        555,
+        170,
+        54,
+        rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=21, bold=True),
+        cell_id="train_zt_port",
+    )
+    q_port = d.vertex(
+        "Prompt features<br><b>q(x)</b>",
+        480,
+        620,
+        170,
+        54,
+        rect_style(FROZEN_FILL, FROZEN, font=FROZEN_DARK, size=21, bold=True),
+        cell_id="train_q_port",
+    )
+    d.edge(pair, target, color=FROZEN_DARK, width=1.7, exit_xy=(0, 0.55), entry_xy=(0.25, 0), points=[(20, 295), (20, 540)])
+    d.edge(target, features, color=FROZEN_DARK, width=1.7, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(features, z_port, color=SHARED, width=1.8, exit_xy=(1, 0.35), entry_xy=(0, 0.5))
+    d.edge(features, q_port, color=FROZEN_DARK, width=1.6, dashed=True, exit_xy=(1, 0.75), entry_xy=(0, 0.5))
+    return {"assignment": assignment, "z": z_port, "q": q_port}
+
+
+def mlp_bank(d: Diagram, x: int, y: int) -> tuple[str, str]:
+    bank = d.vertex(
+        "",
+        x,
+        y,
+        300,
+        260,
+        rect_style(WHITE, RULE, extra="container=1;pointerEvents=0"),
+        cell_id="train_mlp_bank",
+    )
+    d.vertex("MLP bank", 15, 10, 270, 34, text_style(25, bold=True, align="center"), parent=bank)
+    labels = ["MLP<sub>ℓ,1</sub>", "MLP<sub>ℓ,2</sub>", "MLP<sub>ℓ,d</sub>", "⋯", "MLP<sub>ℓ,N</sub>"]
+    selected = ""
+    for idx, label in enumerate(labels):
+        active = idx == 2
+        row = d.vertex(
+            label,
+            18,
+            54 + idx * 38,
+            264,
+            30,
+            rect_style(
+                MLP_FILL if active else MLP_FAINT,
+                MLP if active else RULE,
+                font=MLP_DARK if active else MUTED,
+                size=22,
+                bold=active,
+                stroke_width=2.2 if active else 1.1,
+            ),
+            cell_id=f"train_mlp_{idx + 1}",
+            parent=bank,
+        )
+        if active:
+            selected = row
+    return bank, selected
+
+
+def draw_panel_b(d: Diagram) -> dict[str, str]:
+    panel_title(d, "B", "Jointly train MoS", 750, 28, 520)
+    d.vertex("ONE EXPANDED DRAFT LAYER", 750, 88, 500, 34, text_style(23, font=SHARED_DARK, bold=True))
+
+    h_in = d.vertex("h<sub>ℓ</sub>", 760, 250, 90, 66, rect_style(WHITE, RULE, size=30, bold=True), cell_id="train_h_in")
+    norm_1 = d.vertex("Norm", 890, 250, 110, 66, rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=25, bold=True), cell_id="train_norm_1")
+    attn = d.vertex(
+        "Shared<br>attention",
+        1040,
+        215,
+        210,
+        136,
+        rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=27, bold=True, stroke_width=2.2),
+        cell_id="train_attention",
+    )
+    plus_1 = d.vertex("+", 1280, 257, 52, 52, rect_style(WHITE, INK, size=32, bold=True, extra="ellipse"), cell_id="train_plus_1")
+    norm_2 = d.vertex("Norm", 1370, 250, 100, 66, rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=25, bold=True), cell_id="train_norm_2")
+    bank, selected = mlp_bank(d, 1500, 135)
+    plus_2 = d.vertex("+", 1840, 257, 52, 52, rect_style(WHITE, INK, size=32, bold=True, extra="ellipse"), cell_id="train_plus_2")
+    h_out = d.vertex("h<sub>ℓ+1</sub>", 1930, 250, 100, 66, rect_style(WHITE, RULE, size=28, bold=True), cell_id="train_h_out")
+
+    z_tokens = token_row(
+        d,
+        "zt_train",
+        790,
+        145,
+        ["z<sub>t</sub><sup>1</sup>", "z<sub>t</sub><sup>2</sup>", "⋯"],
+        ["shared", "shared", "shared"],
+        cell_w=56,
+        gap=8,
+        font_size=22,
+    )
+    projection = d.vertex(
+        "Target-context<br>projection",
+        1040,
+        130,
+        210,
+        66,
+        rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=23, bold=True),
+        cell_id="train_projection",
+    )
+    selector = d.vertex(
+        "Training label d",
+        1280,
+        120,
+        180,
+        44,
+        rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=21, bold=True, stroke_width=2.0),
+        cell_id="training_selector",
+    )
+
+    d.edge(h_in, norm_1, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(norm_1, attn, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(attn, plus_1, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(plus_1, norm_2, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(norm_2, selected, color=MLP, width=2.6, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(selected, plus_2, color=MLP, width=2.6, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(plus_2, h_out, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(z_tokens[-1], projection, color=SHARED, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(projection, attn, color=SHARED, width=1.9, exit_xy=(0.5, 1), entry_xy=(0.5, 0))
+    d.edge(
+        selector,
+        selected,
+        color=ROUTE,
+        width=1.9,
+        dashed=True,
+        exit_xy=(1, 0.5),
+        entry_xy=(0, 0.5),
+        points=[(1480, 142), (1480, 280)],
+    )
+
+    d.edge(h_in, plus_1, width=1.7, exit_xy=(0.5, 1), entry_xy=(0.5, 1), points=[(805, 380), (1305, 380)])
+    d.vertex("residual", 940, 390, 120, 28, text_style(20, font=MUTED, align="center"))
+    d.edge(plus_1, plus_2, width=1.7, exit_xy=(0.5, 1), entry_xy=(0.5, 1), points=[(1305, 420), (1865, 420)])
+    d.vertex("residual", 1540, 388, 120, 28, text_style(20, font=MUTED, align="center"))
+
+    d.vertex("Same d across layers", 760, 445, 240, 34, text_style(22, font=MLP_DARK, bold=True))
+    group_boxes = []
+    for idx, (x, label) in enumerate([(1060, "MLP<sub>1,d</sub>"), (1270, "MLP<sub>2,d</sub>"), (1550, "MLP<sub>L,d</sub>")]):
+        group_boxes.append(
+            d.vertex(
+                label,
+                x,
+                438,
+                160,
+                52,
+                rect_style(MLP_FILL, MLP, font=MLP_DARK, size=22, bold=True, stroke_width=2.0),
+                cell_id=f"same_group_{idx}",
+            )
+        )
+    dots = d.vertex("⋯", 1455, 445, 60, 34, text_style(28, font=MLP_DARK, bold=True, align="center"))
+    d.edge(group_boxes[0], group_boxes[1], color=MLP, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(group_boxes[1], dots, color=MLP, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5), end_arrow="none")
+    d.edge(dots, group_boxes[2], color=MLP, width=2.2, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+
+    init = d.vertex(
+        "",
+        2050,
+        80,
+        310,
+        160,
+        rect_style(SURFACE, RULE, extra="container=1;pointerEvents=0"),
+        cell_id="initialization_inset",
+    )
+    d.vertex("Initialization", 15, 10, 280, 30, text_style(24, bold=True, align="center"), parent=init)
+    d.vertex("Public DFlash or trained generalist", 15, 50, 280, 30, text_style(19, font=MUTED, align="center"), parent=init)
+    d.vertex("↓", 135, 80, 40, 20, text_style(22, font=MUTED, bold=True, align="center"), parent=init)
+    d.vertex("Shared modules +<br>copied MLP groups", 15, 108, 280, 42, text_style(20, font=MLP_DARK, bold=True, align="center"), parent=init)
+
+    logits = d.vertex("Parallel predictions", 2070, 280, 270, 52, rect_style(MLP_FAINT, MLP, font=MLP_DARK, size=21, bold=True), cell_id="block_logits")
+    block_loss = d.vertex("Block-prediction loss", 2070, 360, 270, 52, rect_style(WHITE, MLP, font=MLP_DARK, size=21, bold=True, stroke_width=2.0), cell_id="block_loss")
+    block_update = d.vertex("Updates shared modules<br>and selected MLP", 2070, 440, 270, 62, rect_style(MLP_FILL, MLP, font=MLP_DARK, size=19, bold=True, stroke_width=1.8), cell_id="block_update")
+    d.edge(h_out, logits, color=MLP, width=2.0, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(logits, block_loss, color=MLP, width=2.0, exit_xy=(0.5, 1), entry_xy=(0.5, 0))
+    d.edge(block_loss, block_update, color=MLP, width=2.0, exit_xy=(0.5, 1), entry_xy=(0.5, 0))
+
+    d.vertex("ROUTER SUPERVISION", 760, 535, 300, 30, text_style(22, font=ROUTE_DARK, bold=True))
+    detached = d.vertex("Prompt features<br>(no gradient)", 760, 575, 230, 64, rect_style(FROZEN_FILL, FROZEN, font=FROZEN_DARK, size=19, bold=True, dashed=True), cell_id="router_detached_features")
+    pool = d.vertex("Mean pool", 1020, 580, 130, 58, rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=21, bold=True), cell_id="router_pool")
+    router = d.vertex("Request router", 1190, 580, 170, 58, rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=20, bold=True), cell_id="request_router_train")
+    route_loss = d.vertex("Routing loss", 1390, 580, 150, 58, rect_style(WHITE, ROUTE, font=ROUTE_DARK, size=20, bold=True, stroke_width=2.0), cell_id="route_loss")
+    router_update = d.vertex("Router update", 1570, 580, 180, 58, rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=20, bold=True), cell_id="router_update")
+    d.edge(detached, pool, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(pool, router, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(router, route_loss, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(route_loss, router_update, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+
+    checkpoint = d.vertex(
+        "",
+        2070,
+        535,
+        270,
+        125,
+        rect_style(WHITE, INK, stroke_width=2.0, extra="container=1;pointerEvents=0"),
+        cell_id="mos_checkpoint",
+    )
+    d.vertex("MoS checkpoint", 15, 12, 240, 32, text_style(23, bold=True, align="center"), parent=checkpoint)
+    d.vertex("Shared modules · MLP bank<br>Request router", 15, 52, 240, 54, text_style(19, font=MUTED, align="center"), parent=checkpoint)
+    d.edge(block_update, checkpoint, color=MLP, width=1.8, exit_xy=(0.5, 1), entry_xy=(0.65, 0))
+    d.edge(router_update, checkpoint, color=ROUTE, width=1.8, exit_xy=(1, 0.5), entry_xy=(0, 0.75), points=[(1870, 609), (1870, 630)])
+    return {"checkpoint": checkpoint}
+
+
+def inference_layer(d: Diagram, prefix: str, x: int, label: str, selected_label: str, *, parent: str) -> tuple[str, str, str]:
+    layer = d.vertex(
+        "",
+        x,
+        62,
+        340,
+        170,
+        rect_style(WHITE, RULE, extra="container=1;pointerEvents=0"),
+        cell_id=f"{prefix}_layer",
+        parent=parent,
+    )
+    d.vertex(label, 15, 10, 310, 34, text_style(25, bold=True, align="center"), parent=layer)
+    shared = d.vertex(
+        "Shared<br>modules",
+        18,
+        60,
+        165,
+        88,
+        rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=22, bold=True, stroke_width=2.0),
+        cell_id=f"{prefix}_shared",
+        parent=layer,
+    )
+    selected = d.vertex(
+        selected_label,
+        205,
+        64,
+        115,
+        56,
+        rect_style(MLP_FILL, MLP, font=MLP_DARK, size=20, bold=True, stroke_width=2.2),
+        cell_id=f"{prefix}_selected",
+        parent=layer,
+    )
+    d.vertex("", 215, 132, 95, 7, rect_style(MLP_FAINT, RULE, rounded=0, stroke_width=1.0), cell_id=f"{prefix}_inactive_1", parent=layer)
+    d.vertex("", 215, 145, 95, 7, rect_style(MLP_FAINT, RULE, rounded=0, stroke_width=1.0), cell_id=f"{prefix}_inactive_2", parent=layer)
+    d.edge(shared, selected, color=MLP, width=2.3, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    return layer, shared, selected
+
+
+def candidate_block(d: Diagram, x: int, y: int) -> tuple[str, str]:
+    box = d.vertex(
+        "",
+        x,
+        y,
+        340,
+        210,
+        rect_style(WHITE, MLP, stroke_width=2.0, extra="container=1;pointerEvents=0"),
+        cell_id="candidate_block",
+    )
+    d.vertex("Candidate block (parallel)", 15, 12, 310, 34, text_style(21, font=MLP_DARK, bold=True, align="center"), parent=box)
+    tokens = token_row(
+        d,
+        "candidate_tokens",
+        18,
+        70,
+        ["t<sub>1</sub>", "t<sub>2</sub>", "t<sub>3</sub>", "⋯", "t<sub>16</sub>"],
+        ["plain"] * 5,
+        cell_w=48,
+        gap=10,
+        parent=box,
+        font_size=21,
+    )
+    bus = d.vertex("", 35, 166, 270, 3, rect_style(MLP, MLP, rounded=0, stroke_width=0), cell_id="candidate_parallel_bus", parent=box)
+    for idx in [0, 2, 4]:
+        d.edge(bus, tokens[idx], color=MLP, width=1.6, exit_xy=((idx + 1) / 6, 0), entry_xy=(0.5, 1))
+    return box, bus
+
+
+def verification_box(d: Diagram, x: int, y: int) -> str:
+    box = d.vertex(
+        "",
+        x,
+        y,
+        320,
+        230,
+        rect_style(FROZEN_FILL, FROZEN, stroke_width=2.0, extra="container=1;pointerEvents=0"),
+        cell_id="target_verification",
+    )
+    d.vertex("Target verification", 15, 12, 290, 36, text_style(24, font=FROZEN_DARK, bold=True, align="center"), parent=box)
+    token_row(
+        d,
+        "verify_tokens",
+        32,
+        66,
+        ["t<sub>1</sub>", "t<sub>2</sub>", "t<sub>3</sub>", "t<sub>4</sub>"],
+        ["good", "good", "good", "correct"],
+        cell_w=54,
+        gap=10,
+        parent=box,
+        font_size=21,
+    )
+    d.vertex("✓       ✓       ✓", 40, 122, 190, 30, text_style(24, font=GOOD, bold=True, align="center"), parent=box)
+    d.vertex("fix", 245, 122, 45, 30, text_style(19, font=AMBER, bold=True, align="center"), parent=box)
+    d.vertex("Accept matching prefix", 25, 164, 270, 28, text_style(20, font=GOOD, bold=True, align="center"), parent=box)
+    d.vertex("Correct first mismatch", 25, 198, 270, 28, text_style(20, font=AMBER, bold=True, align="center"), parent=box)
+    return box
+
+
+def draw_panel_c(d: Diagram, checkpoint: str) -> None:
+    panel_title(d, "C", "Request-routed speculative decoding", 40, 730, 760)
+    d.vertex("ROUTE ONCE", 40, 792, 220, 32, text_style(23, font=ROUTE_DARK, bold=True))
+
+    prompt = d.vertex("Prompt x", 40, 845, 160, 68, rect_style(WHITE, RULE, size=26, bold=True), cell_id="infer_prompt")
+    prefill = d.vertex("Frozen target<br>prompt pass", 240, 835, 230, 88, rect_style(FROZEN_FILL, FROZEN, font=FROZEN_DARK, size=22, bold=True), cell_id="target_prefill")
+    feature_box = d.vertex(
+        "Prompt features",
+        510,
+        835,
+        180,
+        88,
+        rect_style(FROZEN_FILL, FROZEN, font=FROZEN_DARK, size=20, bold=True, extra="verticalAlign=top;spacingTop=10"),
+        cell_id="prompt_features",
+    )
+    for idx, width in enumerate([130, 110, 90]):
+        d.vertex(
+            "",
+            25,
+            48 + idx * 10,
+            width,
+            5,
+            rect_style(WHITE, RULE, rounded=0, stroke_width=1),
+            cell_id=f"prompt_feature_strip_{idx}",
+            parent=feature_box,
+        )
+    pool = d.vertex("Mean pool", 730, 845, 150, 68, rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=22, bold=True), cell_id="infer_mean_pool")
+    router = d.vertex("Request router", 920, 845, 190, 68, rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=22, bold=True), cell_id="infer_request_router")
+    choice = d.vertex(
+        "Predicted<br>group d<sub>pred</sub>",
+        1150,
+        835,
+        130,
+        88,
+        rect_style(ROUTE_FILL, ROUTE, font=ROUTE_DARK, size=20, bold=True, stroke_width=2.2),
+        cell_id="route_choice",
+    )
+    lock = d.vertex(
+        "LOCK",
+        1320,
+        845,
+        90,
+        68,
+        rect_style(WHITE, ROUTE, font=ROUTE_DARK, size=20, bold=True, stroke_width=2.2),
+        cell_id="route_lock",
+    )
+    d.edge(prompt, prefill, width=2.1, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(prefill, feature_box, color=FROZEN_DARK, width=1.8, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(feature_box, pool, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(pool, router, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(router, choice, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(choice, lock, color=ROUTE, width=1.9, exit_xy=(1, 0.5), entry_xy=(0, 0.5), end_arrow="none")
+
+    d.vertex("One decision per request", 1460, 825, 360, 34, text_style(25, font=ROUTE_DARK, bold=True))
+    d.vertex("cycle 1   —   cycle 2   —   ⋯   —   cycle T", 1460, 866, 520, 32, text_style(23, font=ROUTE_DARK))
+    d.vertex("Fixed for all layers and cycles", 1460, 904, 420, 30, text_style(22, font=ROUTE_DARK, bold=True))
+
+    loaded = d.vertex("Trained MoS checkpoint", 2110, 835, 250, 72, rect_style(SURFACE, RULE, font=MUTED, size=21, bold=True), cell_id="loaded_checkpoint")
+    d.edge(checkpoint, loaded, color=MUTED, width=1.6, dashed=True, exit_xy=(0.75, 1), entry_xy=(0.75, 0), points=[(2300, 690), (2300, 810)])
+
+    d.vertex("DATA PATH", 40, 950, 180, 32, text_style(23, font=MUTED, bold=True))
+    context = d.vertex("Target context<br><b>z<sub>t</sub></b>", 40, 1000, 190, 70, rect_style(SHARED_FILL, SHARED, font=SHARED_DARK, size=22, bold=True), cell_id="infer_context")
+    masked = d.vertex("Masked block<br><b>[M] [M] [M]</b>", 40, 1110, 190, 70, rect_style(GOOD_FILL, GOOD, font=GOOD, size=21, bold=True), cell_id="infer_masked")
+    d.edge(
+        prefill,
+        context,
+        color=SHARED,
+        width=1.8,
+        exit_xy=(0.05, 1),
+        entry_xy=(1, 0.5),
+        points=[(250, 975), (250, 1035)],
+    )
+
+    path = d.vertex(
+        "",
+        290,
+        950,
+        1220,
+        280,
+        rect_style(SURFACE, INK, stroke_width=2.0, extra="container=1;pointerEvents=0"),
+        cell_id="selected_path",
+    )
+    d.vertex("Selected MoS path", 20, 12, 300, 34, text_style(27, bold=True), parent=path)
+    d.vertex("fixed group d<sub>pred</sub> for the whole request", 670, 14, 520, 30, text_style(21, font=MLP_DARK, bold=True, align="right"), parent=path)
+    layer_1, shared_1, selected_1 = inference_layer(d, "infer_1", 20, "Layer 1", "MLP<sub>1</sub>", parent=path)
+    layer_2, shared_2, selected_2 = inference_layer(d, "infer_2", 420, "Layer 2", "MLP<sub>2</sub>", parent=path)
+    layer_l, shared_l, selected_l = inference_layer(d, "infer_l", 820, "Layer L", "MLP<sub>L</sub>", parent=path)
+    d.edge(context, shared_1, color=SHARED, width=2.0, exit_xy=(1, 0.5), entry_xy=(0, 0.3), points=[(260, 1035), (290, 1035)])
+    d.edge(masked, shared_1, color=GOOD, width=2.0, exit_xy=(1, 0.5), entry_xy=(0, 0.75), points=[(260, 1145), (290, 1145)])
+    d.edge(selected_1, shared_2, color=MLP, width=2.6, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.edge(selected_2, shared_l, color=MLP, width=2.6, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    d.vertex("Other MLP groups are not evaluated or merged", 250, 242, 720, 26, text_style(20, font=MUTED, bold=True, align="center"), parent=path)
+
+    candidate, candidate_bus = candidate_block(d, 1550, 970)
+    d.edge(selected_l, candidate_bus, color=MLP, width=2.6, exit_xy=(1, 0.5), entry_xy=(0, 0.5), points=[(1530, 1110)])
+    verify = verification_box(d, 1910, 950)
+    d.edge(candidate, verify, width=2.1, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+    response = d.vertex("Response", 2250, 1025, 140, 72, rect_style(WHITE, RULE, size=21, bold=True), cell_id="response_output")
+    d.edge(verify, response, width=2.1, exit_xy=(1, 0.5), entry_xy=(0, 0.5))
+
+    d.vertex("Accepted or corrected tokens start the next verification cycle", 560, 1260, 900, 32, text_style(22, font=MUTED, bold=True, align="center"), cell_id="cycle_note")
+    d.edge(verify, context, color=FROZEN_DARK, width=1.6, exit_xy=(0.5, 1), entry_xy=(0, 0.5), points=[(2070, 1310), (20, 1310), (20, 1035)])
+
+    roles = [
+        (SHARED_FILL, SHARED, "shared"),
+        (MLP_FILL, MLP, "selected MLP"),
+        (ROUTE_FILL, ROUTE, "routing"),
+        (FROZEN_FILL, FROZEN, "frozen / inactive"),
+    ]
+    for idx, (fill, stroke, label) in enumerate(roles):
+        x = 40 + idx * 260
+        d.vertex("", x, 1360, 34, 24, rect_style(fill, stroke, rounded=0, stroke_width=1.6), cell_id=f"legend_swatch_{idx}")
+        d.vertex(label, x + 46, 1354, 200, 36, text_style(21, font=MUTED, bold=True), cell_id=f"legend_label_{idx}")
+    d.vertex("Exact verification preserves the target distribution", 1500, 1354, 860, 36, text_style(22, font=MUTED, bold=True, align="right"), cell_id="exactness_note")
+
+
+def build() -> Diagram:
+    diagram = Diagram()
+    add_separator(diagram, 700, 20, 2, 650)
+    add_separator(diagram, 30, 700, 2240, 2)
+    add_separator(diagram, 2330, 700, 40, 2)
+    draw_panel_a(diagram)
+    panel_b = draw_panel_b(diagram)
+    draw_panel_c(diagram, panel_b["checkpoint"])
+    return diagram
+
+
+def main() -> None:
+    args = parse_args()
+    build().save(args.output)
+
+
+if __name__ == "__main__":
+    main()

recipes/plotting/v1/fig_5arm_per_domain.py

@@ -0,0 +1,49 @@
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+domains = ["code", "math", "factual_qa", "creative_writing", "general"]
+nan = np.nan
+series = [
+    ("D0 (base)",          [2.44, 3.38, 2.23, 2.11, 2.43], "#cfd8dc"),
+    ("small-data spec.",   [3.16, 4.49, 2.68, 2.54, 2.91], "#90a4ae"),
+    ("warm spec.",         [3.20, 4.68, 2.86, 2.67, 3.06], "#607d8b"),
+    ("Generalist",         [3.205, 4.698, 2.814, 2.651, 3.086], "#1565c0"),
+    ("big-data spec. 250k",[3.340, 5.190, 2.932, 2.861, 3.117],  "#2e7d32"),
+]
+x = np.arange(len(domains)); n = len(series); w = 0.16
+fig, ax = plt.subplots(figsize=(10.5, 5.2))
+for i, (name, vals, c) in enumerate(series):
+    off = (i - (n-1)/2) * w
+    bars = ax.bar(x + off, vals, w, label=name, color=c)
+    for b, v in zip(bars, vals):
+        if not np.isnan(v):
+            ax.text(b.get_x()+b.get_width()/2, v+0.02, f"{v:.2f}",
+                    ha="center", va="bottom", fontsize=7, rotation=90)
+# mark big-data wins (code, math)
+ax.annotate("+0.135", xy=(x[0]+2*w, 3.34), xytext=(x[0]+2*w, 3.72),
+            ha="center", fontsize=10, fontweight="bold",
+            arrowprops=dict(arrowstyle="->", color="#2e7d32", lw=1.4))
+ax.annotate("+0.49", xy=(x[1]+2*w, 5.19), xytext=(x[1]+2*w, 5.60),
+            ha="center", fontsize=10, fontweight="bold",
+            arrowprops=dict(arrowstyle="->", color="#2e7d32", lw=1.4))
+ax.annotate("+0.118", xy=(x[2]+2*w, 2.93), xytext=(x[2]+2*w, 3.30),
+            ha="center", fontsize=10, fontweight="bold",
+            arrowprops=dict(arrowstyle="->", color="#2e7d32", lw=1.4))
+ax.annotate("+0.21", xy=(x[3]+2*w, 2.86), xytext=(x[3]+2*w, 3.25),
+            ha="center", fontsize=10, fontweight="bold",
+            arrowprops=dict(arrowstyle="->", color="#2e7d32", lw=1.4))
+ax.annotate("+0.03", xy=(x[4]+2*w, 3.12), xytext=(x[4]+2*w, 3.50),
+            ha="center", fontsize=9, color="#555",
+            arrowprops=dict(arrowstyle="->", color="#999", lw=1.0))
+ax.set_xticks(x); ax.set_xticklabels(domains)
+ax.set_ylabel("held-out accept length (AL)")
+ax.set_ylim(2.0, 5.85)
+ax.set_title("Per-domain AL: D0 / small-data / warm / Generalist / big-data(250k) specialist\n"
+             "(big-data 5/5: math +0.49, creative_writing +0.21, code +0.135, factual_qa +0.118, general +0.03)")
+ax.legend(ncol=5, loc="upper center", fontsize=8.5, framealpha=0.9)
+ax.grid(axis="y", ls=":", alpha=0.5)
+fig.tight_layout()
+fig.savefig("reasonmix_5arm_per_domain.png", dpi=130)
+print("OK")

recipes/plotting/v1/fig_code250k_vs_gen.py

@@ -0,0 +1,33 @@
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+epochs = ["ep0", "ep1", "ep2", "ep3"]
+gen   = [3.072, 3.150, 3.205, 3.204]   # generalist on code held-out
+spec  = [3.191, 3.268, 3.340, 3.339]   # code250k specialist on code held-out
+x = np.arange(len(epochs)); w = 0.38
+
+fig, ax = plt.subplots(figsize=(7.2, 4.6))
+b1 = ax.bar(x - w/2, gen,  w, label="generalist (mixed 250k, 91k code)", color="#9aa7b8")
+b2 = ax.bar(x + w/2, spec, w, label="code specialist (from D0, 250k code)", color="#2e7d32")
+
+ax.axhline(3.205, ls="--", lw=1.2, color="#9aa7b8")
+ax.axhline(3.340, ls="--", lw=1.2, color="#2e7d32")
+ax.annotate("", xy=(3.34, 3.340), xytext=(3.34, 3.205),
+            arrowprops=dict(arrowstyle="<->", color="black", lw=1.3))
+ax.text(3.0, (3.205+3.340)/2, "+0.135", fontsize=12, fontweight="bold", va="center")
+
+for b in list(b1)+list(b2):
+    ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.006,
+            f"{b.get_height():.3f}", ha="center", va="bottom", fontsize=8.5)
+
+ax.set_xticks(x); ax.set_xticklabels(epochs)
+ax.set_ylim(3.0, 3.42)
+ax.set_ylabel("code held-out accept length (AL)")
+ax.set_title("code: 250k single-domain specialist vs generalist (per-epoch, held-out=89)")
+ax.legend(loc="lower right", fontsize=9)
+ax.grid(axis="y", ls=":", alpha=0.5)
+fig.tight_layout()
+fig.savefig("code250k_vs_gen.png", dpi=130)
+print("OK")

recipes/plotting/v1/fig_exp1_forgetting.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+# Exp 1 forgetting trajectories: continue-train the GENERALIST on ONE domain (math / code),
+# full-param (B) vs MLP-only/frozen-attn (A), baseline (k5clean) data, eval every epoch on all 5 domains.
+# Shows: target domain rises; OTHER domains drop (forgetting) EVEN WITH MLP-only -> forgetting lives in the MLP.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+EP = list(range(6))
+DOMS = ["code", "math", "factual_qa", "creative_writing", "general"]
+GEN = {"code": 3.209, "math": 4.698, "factual_qa": 2.814, "creative_writing": 2.636, "general": 3.049}
+
+# [trained_domain][arm][eval_domain] = per-epoch AL (ep0..ep5)
+DATA = {
+  "math": {
+    "B": {"code":[3.183,3.163,3.143,3.166,3.161,3.157],"math":[4.777,4.758,4.843,4.915,4.907,4.912],
+          "factual_qa":[2.844,2.741,2.760,2.743,2.758,2.809],"creative_writing":[2.606,2.618,2.567,2.599,2.584,2.579],
+          "general":[3.040,3.029,3.031,3.033,3.064,3.048]},
+    "A": {"code":[3.177,3.183,3.153,3.178,3.169,3.188],"math":[4.748,4.781,4.798,4.800,4.842,4.847],
+          "factual_qa":[2.834,2.765,2.766,2.747,2.830,2.769],"creative_writing":[2.663,2.596,2.599,2.628,2.629,2.606],
+          "general":[3.067,3.046,3.039,3.075,3.090,3.056]},
+  },
+  "code": {
+    "B": {"code":[3.208,3.239,3.296,3.303,3.329,3.338],"math":[4.630,4.646,4.583,4.574,4.559,4.578],
+          "factual_qa":[2.781,2.739,2.820,2.747,2.753,2.744],"creative_writing":[2.602,2.569,2.572,2.590,2.633,2.598],
+          "general":[3.017,3.029,3.030,3.021,3.022,3.048]},
+    "A": {"code":[3.226,3.261,3.263,3.308,3.304,3.334],"math":[4.665,4.613,4.580,4.600,4.575,4.585],
+          "factual_qa":[2.815,2.839,2.766,2.767,2.746,2.812],"creative_writing":[2.630,2.631,2.622,2.617,2.613,2.609],
+          "general":[3.055,3.042,3.048,3.065,3.042,3.033]},
+  },
+}
+
+fig, axes = plt.subplots(2, 5, figsize=(18, 6.8), sharex=True)
+for r, trained in enumerate(["math", "code"]):
+    for c, ed in enumerate(DOMS):
+        ax = axes[r][c]
+        is_target = (ed == trained)
+        ax.plot(EP, DATA[trained]["B"][ed], "o-", color="#2e7d32", lw=1.8, ms=4, label="full-param (B)")
+        ax.plot(EP, DATA[trained]["A"][ed], "s-", color="#e67e22", lw=1.8, ms=4, label="MLP-only / frozen-attn (A)")
+        ax.axhline(GEN[ed], ls="--", color="#888", lw=1.2, label="Generalist (start)")
+        ax.set_title(f"eval={ed}" + ("  ◀ TARGET" if is_target else ""),
+                     fontsize=9.5, fontweight=("bold" if is_target else "normal"),
+                     color=("#1a1a1a" if is_target else "#555"))
+        ax.grid(alpha=0.3)
+        if is_target:
+            ax.set_facecolor("#eef7ee")
+        if c == 0:
+            ax.set_ylabel(f"train on {trained.upper()}\nheld-out AL", fontsize=10)
+        if r == 1:
+            ax.set_xlabel("epoch")
+        ax.tick_params(labelsize=8)
+axes[0][0].legend(fontsize=7.5, loc="best")
+fig.suptitle("Exp 1 — Forgetting trajectory: continue-train Generalist on ONE domain (baseline data), eval all 5 every epoch.\n"
+             "TARGET (green bg) rises; OTHER domains fall below Generalist (forgetting). Code→math forgetting is the same for full-param and MLP-only "
+             "→ forgetting lives in the MLP (router to per-domain MLP isolates it).", fontsize=11, y=1.02)
+fig.tight_layout()
+fig.savefig("fig_exp1_forgetting.png", dpi=135, bbox_inches="tight")
+print("OK wrote /tmp/fig_exp1_forgetting.png")

recipes/plotting/v1/fig_exp1_three_recipes.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python3
+# Exp1: three specialist recipes (big-data / small-data / warm-MLP) vs generalist, on each domain's own data.
+# One grouped bar chart. Only big-data specialist beats generalist.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+DOMS = ["math", "code", "fqa", "cw", "general"]
+GEN      = [4.698, 3.209, 2.814, 2.647, 3.086]   # generalist (best single ckpt)
+BIGDATA  = [5.190, 3.340, 2.932, 2.861, 3.117]   # big-data specialist (单域 250k), peak
+SMALLDATA= [4.493, 3.162, 2.683, 2.536, 2.908]   # small-data specialist (单域自然量), peak
+WARM     = [4.748, 3.215, 2.861, 2.671, 3.106]   # warm MLP-only specialist, peak
+
+x = np.arange(len(DOMS)); w = 0.2
+fig, ax = plt.subplots(figsize=(11, 5.4))
+bars = [
+    ax.bar(x - 1.5*w, GEN,       w, label="generalist (monolithic baseline)", color="#9aa7b8"),
+    ax.bar(x - 0.5*w, BIGDATA,   w, label="big-data specialist (250k/domain)", color="#1b5e20"),
+    ax.bar(x + 0.5*w, SMALLDATA, w, label="small-data specialist (natural share)", color="#c0392b"),
+    ax.bar(x + 1.5*w, WARM,      w, label="warm specialist (MLP-only)", color="#e0a030"),
+]
+for bs in bars:
+    for b in bs:
+        ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.02, f"{b.get_height():.2f}",
+                ha="center", va="bottom", fontsize=7)
+ax.set_xticks(x); ax.set_xticklabels(DOMS)
+ax.set_ylabel("held-out accept length (AL)")
+ax.set_ylim(2.0, 5.7)
+ax.set_title("Only the big-data specialist (250k/domain) beats the generalist on all 5 domains;\nsmall-data and warm specialists do not", fontsize=11)
+ax.legend(loc="upper right", fontsize=9, ncol=2); ax.grid(axis="y", ls=":", alpha=0.4)
+fig.tight_layout()
+fig.savefig("fig_exp1_three_recipes.png", dpi=140, bbox_inches="tight")
+print("OK wrote /tmp/fig_exp1_three_recipes.png")

recipes/plotting/v1/fig_exp4_inference.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Exp 4 inference cost: (A) end-to-end single-stream tok/s merged~specialist>gen; (B) serve-step component
+# breakdown (decode regime) — self_attn dominates, the lm_head "giant" is only ~6.5%.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+ARMS = ["generalist", "merged", "specialist"]
+TOKS = [505.1, 567.3, 559.8]      # single-stream tok/s (batch=1, fa3, math held-out)
+AL   = [4.69, 5.32, 5.24]
+COL  = {"generalist": "#9aa7b8", "merged": "#2e7d32", "specialist": "#b8860b"}
+
+fig, (axA, axB) = plt.subplots(1, 2, figsize=(12.5, 5.0))
+
+# --- Panel A: end-to-end tok/s ---
+x = np.arange(len(ARMS))
+bars = axA.bar(x, TOKS, width=0.6, color=[COL[a] for a in ARMS])
+for i, b in enumerate(bars):
+    axA.text(b.get_x()+b.get_width()/2, b.get_height()+3, f"{TOKS[i]:.0f}\nAL {AL[i]:.2f}",
+             ha="center", va="bottom", fontsize=9)
+axA.set_xticks(x); axA.set_xticklabels(ARMS)
+axA.set_ylabel("single-stream throughput (tok/s)")
+axA.set_ylim(0, 640)
+axA.set_title("A. End-to-end speed (math, batch=1)\nmerged 567 ≈ specialist 560,  +12% over generalist 505", fontsize=10.5)
+axA.text(0.5, 0.04, "merged delivers specialist-level speed at single-model per-token cost",
+         transform=axA.transAxes, ha="center", fontsize=8.5, style="italic", color="#444")
+axA.grid(axis="y", ls=":", alpha=0.4)
+
+# --- Panel B: serve-step component breakdown (ms), all arms ~identical -> show one ---
+comp_names = ["self_attn\n(5L)", "draft_other\n(embed/norm/\nresidual/cache)", "mlp\n(5L)", "lm_head\n(GIANT)"]
+comp_ms    = [2.806, 1.487, 0.641, 0.342]
+comp_col   = ["#c0392b", "#7f8c8d", "#2980b9", "#f1c40f"]
+full = sum(comp_ms)
+xb = np.arange(len(comp_names))
+bb = axB.bar(xb, comp_ms, width=0.6, color=comp_col)
+for i, b in enumerate(bb):
+    axB.text(b.get_x()+b.get_width()/2, b.get_height()+0.03, f"{comp_ms[i]:.2f}ms\n{100*comp_ms[i]/full:.0f}%",
+             ha="center", va="bottom", fontsize=8.5)
+axB.set_xticks(xb); axB.set_xticklabels(comp_names, fontsize=8)
+axB.set_ylabel("serve-step time (ms, decode block of 16)")
+axB.set_ylim(0, 3.4)
+axB.set_title("B. Where the drafter step spends time (decode regime)\nself_attn dominates (53%); the lm_head 'giant' is only 6.5%", fontsize=10.5)
+axB.text(0.5, 0.92, "per-forward step ≈ identical across all 3 arms (within 1.5%)",
+         transform=axB.transAxes, ha="center", fontsize=8.5, style="italic", color="#444")
+axB.grid(axis="y", ls=":", alpha=0.4)
+
+fig.suptitle("Exp 4 — Merged drafter: single-model per-token cost, specialist-level throughput, negligible router (0.0002 ms/tok)",
+             fontsize=11.5, y=1.02)
+fig.tight_layout()
+fig.savefig("fig_exp4_inference.png", dpi=140, bbox_inches="tight")
+print("OK wrote /tmp/fig_exp4_inference.png")

recipes/plotting/v1/fig_exp5_method_vs_data.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+# Exp6 method>data: at the SAME data budget, three tiers per domain —
+#   naive (from D0, before our fix; LOSES gen) < generalist < CorDA-MoS warm-from-gen (OURS; beats gen).
+# Shows how much our method (warm-from-gen + CorDA fusion) advances over the naive same-data attempt.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+# ascending by our method -> tallest (math) at the right. OURS = saturated (8-epoch) per-domain peaks.
+DOMS  = ["cw", "fqa", "general", "code", "math"]
+NAIVE = [2.591, 2.775, 3.000, 3.211, 4.632]   # naive same-data CorDA-MoS from D0 (per-domain peak) — loses gen
+GEN   = [2.647, 2.814, 3.049, 3.204, 4.698]   # generalist, single best-avg checkpoint
+MOS   = [2.721, 2.907, 3.108, 3.338, 4.922]   # CorDA-MoS warm-from-gen, SATURATED per-domain peak — OURS
+
+x = np.arange(len(DOMS)); w = 0.27
+fig, ax = plt.subplots(figsize=(10.5, 5.4))
+b0 = ax.bar(x - w, NAIVE, w, label="naive same-data (from D0, before warm-start) — loses", color="#c0392b")
+b1 = ax.bar(x,     GEN,   w, label="generalist (same 250k data)", color="#9aa7b8")
+b2 = ax.bar(x + w, MOS,   w, label="CorDA-MoS warm-from-gen (OURS, same data) — wins", color="#1b5e20")
+for bars in (b0, b1, b2):
+    for b in bars:
+        ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.02, f"{b.get_height():.2f}", ha="center", va="bottom", fontsize=7.5)
+# show the advance our method makes over the naive version
+for j in range(len(DOMS)):
+    gain = MOS[j] - NAIVE[j]
+    ax.text(x[j]+w, MOS[j]+0.20, f"+{gain:.2f} vs naive", ha="center", fontsize=7, color="#1b5e20", fontweight="bold")
+ax.set_xticks(x); ax.set_xticklabels(DOMS)
+ax.set_ylabel("held-out accept length (AL)")
+ax.set_ylim(2.0, 5.4)
+ax.set_title("Method > Data (same 250k): naive same-data split (from D0) LOSES to gen;\n"
+             "our warm-from-gen CorDA-MoS BEATS gen on all 5 domains — the gap shows the method's contribution", fontsize=10.5)
+ax.legend(loc="upper left", fontsize=8.5); ax.grid(axis="y", ls=":", alpha=0.4)
+ax.text(0.58, 0.74, f"avg AL:  naive {np.mean(NAIVE):.3f}  <  gen {np.mean(GEN):.3f}  <  ours {np.mean(MOS):.3f}",
+        transform=ax.transAxes, fontsize=9.5, va="top", ha="center",
+        bbox=dict(boxstyle="round,pad=0.3", fc="#e8f3e8", ec="#1b5e20"))
+fig.tight_layout()
+fig.savefig("fig_exp5_method_vs_data.png", dpi=140, bbox_inches="tight")
+print("OK wrote /tmp/fig_exp5_method_vs_data.png")

recipes/plotting/v1/fig_fusion_combined.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env python3
+# Combined fusion figure (replaces separate Exp3 + Exp6 figs):
+# per domain, three bars — generalist (baseline) -> small-data fusion (CorDA-MoS warm, SAME data as gen)
+# -> big-data fusion (merged 5 big-data specialists). Shows the fusion ladder: method gain (same data),
+# then method+data gain.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+DOMS = ["cw", "fqa", "general", "code", "math"]            # ascending -> math tallest at right
+GEN      = [2.647, 2.814, 3.049, 3.204, 4.698]             # generalist (single ckpt)            avg 3.282
+SMALLFUS = [2.721, 2.907, 3.108, 3.338, 4.922]             # CorDA-MoS warm, SAME 250k data       avg 3.399
+BIGFUS   = [2.857, 3.061, 3.227, 3.392, 5.247]             # merged big-data drafter (250k/dom)   avg 3.557
+x = np.arange(len(DOMS)); w = 0.27
+fig, ax = plt.subplots(figsize=(11, 5.6))
+b0 = ax.bar(x - w, GEN,      w, label=f"generalist (baseline)            avg {np.mean(GEN):.3f}", color="#9aa7b8")
+b1 = ax.bar(x,     SMALLFUS, w, label=f"small-data fusion · CorDA-MoS (SAME data as gen)   avg {np.mean(SMALLFUS):.3f}", color="#2a7fb8")
+b2 = ax.bar(x + w, BIGFUS,   w, label=f"big-data fusion · merged drafter (250k/domain)     avg {np.mean(BIGFUS):.3f}", color="#1b5e20")
+for bars in (b0, b1, b2):
+    for b in bars:
+        ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.02, f"{b.get_height():.2f}", ha="center", va="bottom", fontsize=7.5)
+ax.set_xticks(x); ax.set_xticklabels(DOMS)
+ax.set_ylabel("held-out accept length (AL)"); ax.set_ylim(2.2, 5.6)
+ax.set_title("Fusion ladder: same-data fusion already beats the generalist on all 5 domains (method),\n"
+             "more data per domain lifts it further — both share ONE attention (inference cost = one drafter)", fontsize=10.5)
+ax.legend(loc="upper left", fontsize=8.6); ax.grid(axis="y", ls=":", alpha=0.4)
+ax.text(0.60, 0.70, f"avg AL:  gen {np.mean(GEN):.3f}  <  small-data fusion {np.mean(SMALLFUS):.3f}  <  big-data fusion {np.mean(BIGFUS):.3f}",
+        transform=ax.transAxes, fontsize=9.5, va="top", ha="center",
+        bbox=dict(boxstyle="round,pad=0.3", fc="#e8f3e8", ec="#1b5e20"))
+fig.tight_layout()
+fig.savefig("fig_fusion_combined.png", dpi=140, bbox_inches="tight")
+print("OK wrote fig_fusion_combined.png")

recipes/plotting/v1/fig_matrix_and_arms.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+# Forgetting matrix + per-arm epoch-matched comparisons (reasonmix DFlash).
+# All numbers baked from same-protocol bench (DFLASH 8,1,1,16, ROUTED=0, thinking-on, reasonmix held-out).
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from matplotlib.colors import LinearSegmentedColormap
+from matplotlib.patches import Rectangle
+import numpy as np
+import os
+
+OUT = os.environ.get("MOS_FIG_OUT", ".")
+DOM = ["code", "math", "factual_qa", "creative_writing", "general"]
+SHORT = ["code", "math", "fqa", "cw", "general"]
+
+# ---------- FIG 1: 5x5 forgetting matrix (same-protocol) ----------
+SPECS = ["code", "math", "factual_qa", "creative_writing", "general"]
+M = np.array([
+    [3.340, 4.094, 2.612, 2.419, 2.847],   # code-spec
+    [2.853, 5.190, 2.501, 2.296, 2.842],   # math-spec
+    [2.914, 4.278, 2.902, 2.491, 3.026],   # factual_qa-spec
+    [2.805, 3.892, 2.707, 2.861, 2.948],   # cw-spec
+    [2.992, 4.614, 2.773, 2.627, 3.117],   # general-spec
+])
+GEN = np.array([3.209, 4.698, 2.814, 2.636, 3.049])   # generalist (same protocol, ep3)
+D = M - GEN[None, :]                                    # delta vs generalist, per column
+
+matrix_short = ["Code", "Math", "Fact.", "Creat.", "Gen."]
+paper_fs = 9.6  # remains at least 9 pt after final single-column placement
+paper_diverging = LinearSegmentedColormap.from_list(
+    "paper_diverging", ["#c6dbef", "#ffffff", "#fdd0a2"]
+)
+fig, ax = plt.subplots(figsize=(3.3, 3.0))
+vmax = np.abs(D).max()
+im = ax.imshow(D, cmap=paper_diverging, vmin=-vmax, vmax=vmax, aspect="auto")
+for i in range(5):
+    for j in range(5):
+        delta = f"{D[i,j]:+.2f}".replace("+0.", "+.").replace("-0.", "-.")
+        ax.text(j, i, f"{M[i,j]:.2f}\n{delta}", ha="center", va="center",
+                fontsize=paper_fs, linespacing=0.95,
+                fontweight=("bold" if i == j else "normal"))
+        if i == j:
+            ax.add_patch(Rectangle((j - 0.46, i - 0.46), 0.92, 0.92,
+                                   fill=False, edgecolor="black", linewidth=1.0))
+ax.set_xticks(range(5)); ax.set_xticklabels(matrix_short, fontsize=paper_fs)
+ax.set_yticks(range(5)); ax.set_yticklabels(matrix_short, fontsize=paper_fs)
+ax.set_xlabel("Evaluation domain", fontsize=paper_fs, labelpad=3)
+ax.set_ylabel("Specialist", fontsize=paper_fs, labelpad=3)
+ax.tick_params(axis="both", labelsize=paper_fs, width=0.6, length=2.5)
+for spine in ax.spines.values():
+    spine.set_linewidth(0.6)
+cb = fig.colorbar(im, ax=ax, fraction=0.050, pad=0.025)
+cb.set_label(r"$\Delta$ AL vs. generalist", fontsize=paper_fs, labelpad=3)
+cb.ax.tick_params(labelsize=paper_fs, width=0.6, length=2.5)
+cb.outline.set_linewidth(0.6)
+fig.tight_layout(pad=0.25)
+fig.savefig(f"{OUT}/fig_forgetting_matrix.png", dpi=300, bbox_inches="tight")
+plt.close(fig)
+
+# ---------- generic per-arm grouped-bar (epoch-matched: arm@best vs gen@same epoch) ----------
+def arm_vs_gen(fname, title, spec, gen, ep_lbl, spec_name, spec_color):
+    x = np.arange(5); w = 0.38
+    fig, ax = plt.subplots(figsize=(8.4, 4.8))
+    b1 = ax.bar(x - w/2, gen,  w, label="Generalist (same epoch)", color="#9aa7b8")
+    b2 = ax.bar(x + w/2, spec, w, label=spec_name, color=spec_color)
+    for bars in (b1, b2):
+        for b in bars:
+            ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.02, f"{b.get_height():.2f}",
+                    ha="center", va="bottom", fontsize=8)
+    for j in range(5):
+        d = spec[j] - gen[j]
+        ax.text(x[j], max(spec[j], gen[j]) + 0.16, f"{d:+.3f}", ha="center", fontsize=9,
+                fontweight="bold", color=("#2e7d32" if d > 0 else "#c62828"))
+        ax.text(x[j], min(spec[j], gen[j]) - 0.001, ep_lbl[j], ha="center", va="top", fontsize=7, color="#555")
+    ax.set_xticks(x); ax.set_xticklabels(SHORT)
+    ax.set_ylabel("held-out accept length (AL)")
+    ax.set_ylim(2.0, max(spec.max(), gen.max()) + 0.5)
+    ax.set_title(title, fontsize=11)
+    ax.legend(loc="upper left", fontsize=9); ax.grid(axis="y", ls=":", alpha=0.4)
+    fig.tight_layout(); fig.savefig(f"{OUT}/{fname}", dpi=140, bbox_inches="tight"); plt.close(fig)
+
+# FIG 2: big-data specialist (250k, full params, from D0) — best epoch vs gen@same epoch
+bd_spec = np.array([3.340, 5.190, 2.932, 2.861, 3.117]);  bd_gen = np.array([3.205, 4.698, 2.767, 2.651, 3.075])
+bd_ep   = ["ep2", "ep3", "ep2", "ep3", "ep3"]
+arm_vs_gen("fig_bigdata_vs_gen.png",
+           "Big-data specialist (250k single-domain, full params, from D0)\nspecialist@best-epoch  vs  Generalist@same-epoch",
+           bd_spec, bd_gen, bd_ep, "Big-data specialist", "#2e7d32")
+
+# FIG 3: small-data specialist (oracle: baseline volume, full params, from D0) — SEPARATE
+sd_spec = np.array([3.162, 4.493, 2.683, 2.536, 2.908]);  sd_gen = np.array([3.205, 4.694, 2.814, 2.596, 3.075])
+sd_ep   = ["ep2", "ep2", "ep3", "ep1", "ep3"]
+arm_vs_gen("fig_smalldata_vs_gen.png",
+           "Small-data specialist (baseline volume, full params, from D0)\nspecialist@best-epoch  vs  Generalist@same-epoch  (loses on every domain)",
+           sd_spec, sd_gen, sd_ep, "Small-data specialist", "#607d8b")
+
+# ---------- FIG 4: warm specialist (cleanA: MLP-only, frozen backbone, from generalist) saturation ----------
+warm = {
+    "code":             ([700, 1400, 2100, 2800, 2840], [3.210, 3.214, 3.208, 3.215, 3.198]),
+    "math":             ([250, 500, 750, 984],          [4.748, 4.690, 4.681, 4.675]),
+    "factual_qa":       ([350, 700, 1050, 1390],        [2.789, 2.843, 2.790, 2.861]),
+    "creative_writing": ([230, 460, 690, 913],          [2.670, 2.654, 2.646, 2.671]),
+    "general":          ([420, 840, 1260, 1674],        [3.106, 3.093, 3.079, 3.058]),
+}
+warm_gen = {"code": 3.205, "math": 4.698, "factual_qa": 2.814, "creative_writing": 2.651, "general": 3.086}
+fig, axes = plt.subplots(1, 5, figsize=(15, 3.4))
+for ax, d, s in zip(axes, DOM, SHORT):
+    steps, al = warm[d]; frac = np.array(steps) / steps[-1]
+    ax.plot(frac, al, "o-", color="#b8860b", lw=1.8, label="warm spec (MLP-only)")
+    ax.axhline(warm_gen[d], ls="--", color="#1565c0", lw=1.3, label="Generalist (start)")
+    ax.set_title(s, fontsize=10); ax.set_xlabel("frac of 1 epoch"); ax.grid(alpha=0.3)
+    ax.set_xlim(0, 1.02)
+    lo = min(al + [warm_gen[d]]); hi = max(al + [warm_gen[d]])
+    ax.set_ylim(lo - 0.04, hi + 0.04)
+axes[0].set_ylabel("held-out AL"); axes[0].legend(fontsize=7, loc="lower right")
+fig.suptitle("Warm spec (continue Generalist's MLP only, backbone frozen): AL stays within ~±0.04 of Generalist "
+             "— flat / declining (math, general), peaks early. MLP already saturated → can't push AL up.", fontsize=10.5, y=1.04)
+fig.tight_layout(); fig.savefig(f"{OUT}/fig_warm_saturation.png", dpi=140, bbox_inches="tight"); plt.close(fig)
+
+print("OK wrote: fig_forgetting_matrix.png fig_bigdata_vs_gen.png fig_smalldata_vs_gen.png fig_warm_saturation.png")

recipes/plotting/v1/fig_merged_vs_specialist.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+# Headline: the merged drafter (1 shared CorDA-fused attention + per-domain MLP) vs 5 separate specialists vs generalist.
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+# ordered by merged AL ascending -> tallest (math) at the right end
+DOMS = ["creative_writing", "factual_qa", "general", "code", "math"]
+SHORT = ["cw", "fqa", "general", "code", "math"]
+GEN    = [2.636, 2.814, 3.049, 3.209, 4.698]
+MERGED = [2.857, 3.061, 3.227, 3.392, 5.247]   # all 5 domains: MLP retrained on the fused attention
+SPEC   = [2.861, 2.902, 3.117, 3.340, 5.190]
+RETRAINED = [True, True, True, True, True]      # all retrained -> all beat their specialist
+
+x = np.arange(len(DOMS)); w = 0.26
+fig, ax = plt.subplots(figsize=(9.5, 5.0))
+b1 = ax.bar(x - w, GEN,    w, label="Generalist (monolithic)", color="#9aa7b8")
+b2 = ax.bar(x,     MERGED, w, label="Merged drafter (1 shared attn + per-domain MLP)", color="#2e7d32")
+b3 = ax.bar(x + w, SPEC,   w, label="5 separate specialists", color="#b8860b")
+for bars in (b1, b2, b3):
+    for b in bars:
+        ax.text(b.get_x()+b.get_width()/2, b.get_height()+0.02, f"{b.get_height():.2f}",
+                ha="center", va="bottom", fontsize=7.5)
+for j in range(len(DOMS)):
+    if RETRAINED[j]:
+        ax.text(x[j], MERGED[j]+0.18, "retrained", ha="center", fontsize=6.5, color="#2e7d32")
+ax.set_xticks(x); ax.set_xticklabels(SHORT)
+ax.set_ylabel("held-out accept length (AL)")
+ax.set_ylim(2.0, 5.6)
+ax.set_title("Merged drafter > 5 separate specialists (avg 3.557 vs 3.482) with ONE shared attention, >> generalist (3.281)\n"
+             "merged = CorDA-fused shared attention + per-domain MLP (all 5 retrained on the fused attn; every domain beats its specialist)", fontsize=10.5)
+ax.legend(loc="upper left", fontsize=9); ax.grid(axis="y", ls=":", alpha=0.4)
+# avg annotation
+ax.text(0.46, 0.82, f"avg AL:  merged {np.mean(MERGED):.3f}  |  specialists {np.mean(SPEC):.3f}  |  gen {np.mean(GEN):.3f}",
+        transform=ax.transAxes, fontsize=9, va="top",
+        bbox=dict(boxstyle="round,pad=0.3", fc="#eef7ee", ec="#2e7d32"))
+fig.tight_layout()
+fig.savefig("fig_merged_vs_specialist.png", dpi=140, bbox_inches="tight")
+print("OK wrote /tmp/fig_merged_vs_specialist.png")

recipes/plotting/v1/fig_serving_specialist.py

@@ -0,0 +1,58 @@
+import matplotlib; matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+# ---- fig_serving: (a) single-stream tokens/s, (b) garbage tau tradeoff ----
+fig, (a, b) = plt.subplots(1, 2, figsize=(12.6, 4.4), gridspec_kw={"width_ratios": [1, 1.15]})
+for ax in (a, b):
+    for s in ["top", "right"]: ax.spines[s].set_visible(False)
+    for s in ["left", "bottom"]: ax.spines[s].set_color("#c9ced6")
+    ax.tick_params(colors="#444", labelsize=10.5)
+
+names = ["target only", "target +\ngeneralist drafter", "MoS, real router\n(5 experts + router)"]
+tps = [184, 391, 415]
+cols = ["#8a8f98", "#4DABF7", "#9C36B5"]
+bars = a.barh(names, tps, color=cols, height=0.62)
+a.invert_yaxis()
+for r, v, al in zip(bars, tps, ["", "AL 3.40", "AL 3.65 · acc 0.878"]):
+    a.text(v + 6, r.get_y() + r.get_height() / 2, f"{v}", va="center", fontsize=12, fontweight="bold", color=r.get_facecolor())
+    if al:
+        a.text(8, r.get_y() + r.get_height() / 2, al, va="center", fontsize=9.5, color="white", fontweight="bold")
+a.set_xlim(0, 500)
+a.set_xlabel("tokens/s  (single stream, H200, thinking on)", fontsize=11)
+a.set_title("Serving speed with the real router", fontsize=12, fontweight="bold", loc="left")
+
+taus = [0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+junk = [60.0, 66.7, 68.9, 73.3, 73.3, 75.6]
+real = [3.5, 5.5, 6.0, 6.0, 6.5, 9.5]
+b.axvspan(0.4, 0.5, color="#E9FAC8", alpha=0.7, zorder=0)
+b.plot(taus, junk, "o-", color="#2F9E44", lw=2.2, ms=5, label="junk requests sent to garbage (%)")
+b.plot(taus, real, "s-", color="#E8590C", lw=2.2, ms=5, label="real-domain requests mis-sent (%)")
+for t, j in zip(taus, junk): b.text(t, j + 2.5, f"{j:.0f}", ha="center", fontsize=9, color="#2F9E44")
+for t, r_ in zip(taus, real): b.text(t, r_ + 2.5, f"{r_:.0f}", ha="center", fontsize=9, color="#E8590C")
+b.text(0.45, 96, "τ = 0.4–0.5", ha="center", fontsize=10, color="#5c940d", fontweight="bold")
+b.set_ylim(0, 105); b.set_xlim(0.17, 0.73)
+b.set_xlabel("garbage threshold τ  (raw 4-domain max prob)", fontsize=11)
+b.set_ylabel("% of requests", fontsize=11)
+b.set_title("Garbage fallback v1 (no training): expected AL unchanged", fontsize=12, fontweight="bold", loc="left")
+b.legend(frameon=False, fontsize=9.5, loc="center right")
+plt.tight_layout()
+plt.savefig("fig_serving.png", dpi=150, bbox_inches="tight")
+
+# ---- fig_specialist_overall: weighted overall bars ----
+fig2, c = plt.subplots(figsize=(8.6, 4.2))
+for s in ["top", "right"]: c.spines[s].set_visible(False)
+for s in ["left", "bottom"]: c.spines[s].set_color("#c9ced6")
+c.tick_params(colors="#444", labelsize=10.5)
+labels = ["generalist", "code specialist\n(from generalist)", "code specialist\n(from scratch)", "MoS\n(from D0)", "MoS\n(from generalist)"]
+vals = [3.39, 3.40, 3.18, 3.59, 3.65]
+ccols = ["#8a8f98", "#C92A2A", "#5F3DC4", "#2F9E44", "#9C36B5"]
+bars = c.bar(labels, vals, color=ccols, width=0.62)
+for r, v in zip(bars, vals):
+    c.text(r.get_x() + r.get_width() / 2, v + 0.012, f"{v:.2f}", ha="center", fontsize=12, fontweight="bold", color=r.get_facecolor())
+c.axhline(3.39, ls=(0, (3, 3)), lw=1.1, color="#8a8f98", alpha=0.8)
+c.set_ylim(3.0, 3.78)
+c.set_ylabel("AL, weighted by traffic share", fontsize=11)
+c.set_title("One specialist ≈ generalist at best; MoS beats both (same active params)", fontsize=12, fontweight="bold", loc="left")
+plt.tight_layout()
+plt.savefig("fig_specialist_overall.png", dpi=150, bbox_inches="tight")
+print("saved both")

recipes/plotting/v1/plot_main_results.py

@@ -0,0 +1,380 @@
+#!/usr/bin/env python3
+"""Render the three-panel MoS main-results figure from frozen evidence.
+
+All three panels use the fixed-seed, five-domain R1 evaluation. Panel (a)
+compares the matched-domain profiles of the Generalist and the two MoS
+initializations; panels (b)--(c) show the corresponding MoS routing matrices.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import matplotlib as mpl
+import matplotlib.pyplot as plt
+import numpy as np
+from matplotlib.colors import LinearSegmentedColormap, Normalize
+from matplotlib.patches import Rectangle
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+DEFAULT_EVIDENCE = (
+    REPO_ROOT
+    / "paper"
+    / "submission"
+    / "evidence"
+    / "r1_mainfig_seed20260719_20260721T0602Z_cells_summary.json"
+)
+DEFAULT_OUTPUT = (
+    REPO_ROOT / "paper" / "submission" / "figures" / "fig_main_results.png"
+)
+
+DOMAINS = ["code", "math", "factual_qa", "creative_writing", "general"]
+DOMAIN_LABELS = ["Code", "Math", "Factual QA", "Creative", "General"]
+
+# Restrained, color-blind-safe palette. Shape and line style also distinguish
+# methods, so the figure remains legible in grayscale.
+INK = "#25313B"
+MUTED = "#68747E"
+GRID = "#E2E7EA"
+GENERALIST = "#7F8790"
+D0 = "#2F9E44"
+WARM = "#9C36B5"
+LIGHT_RULE = "#C8D0D5"
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--evidence", type=Path, default=DEFAULT_EVIDENCE)
+    parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
+    return parser.parse_args()
+
+
+def configure_style() -> None:
+    mpl.rcParams.update(
+        {
+            "font.family": "sans-serif",
+            "font.sans-serif": [
+                "Arial",
+                "Helvetica",
+                "Liberation Sans",
+                "DejaVu Sans",
+            ],
+            "font.size": 8.0,
+            "axes.titlesize": 9.3,
+            "axes.labelsize": 8.3,
+            "xtick.labelsize": 7.4,
+            "ytick.labelsize": 7.4,
+            "legend.fontsize": 7.3,
+            "pdf.fonttype": 42,
+            "ps.fonttype": 42,
+            "axes.linewidth": 0.65,
+            "savefig.bbox": "tight",
+            "savefig.pad_inches": 0.035,
+        }
+    )
+
+
+def matrix_from_evidence(evidence: dict, key: str) -> np.ndarray:
+    mapping = evidence[key]
+    matrix = np.asarray(
+        [[float(mapping[row][column]) for column in DOMAINS] for row in DOMAINS],
+        dtype=float,
+    )
+    for column in range(len(DOMAINS)):
+        if int(np.argmax(matrix[:, column])) != column:
+            raise ValueError(f"{key}: matched MLP is not best in column {DOMAINS[column]}")
+    return matrix
+
+
+def generalist_from_evidence(evidence: dict) -> np.ndarray:
+    mapping = evidence["panel_d_generalist"]
+    return np.asarray([float(mapping[domain]) for domain in DOMAINS], dtype=float)
+
+
+def panel_title(ax: plt.Axes, letter: str, title: str) -> None:
+    # Use a point-based offset so the letter-to-title gap is physically
+    # identical in the full-width trajectory and the half-width matrices.
+    origin = (-0.055, 1.075)
+    ax.text(
+        *origin,
+        letter,
+        transform=ax.transAxes,
+        ha="left",
+        va="bottom",
+        fontsize=9.8,
+        fontweight="bold",
+        color=INK,
+    )
+    ax.annotate(
+        title,
+        xy=origin,
+        xycoords="axes fraction",
+        xytext=(18, 0),
+        textcoords="offset points",
+        ha="left",
+        va="bottom",
+        fontsize=8.6,
+        fontweight="bold",
+        color=INK,
+    )
+
+
+def quiet_axes(ax: plt.Axes) -> None:
+    ax.spines["top"].set_visible(False)
+    ax.spines["right"].set_visible(False)
+    ax.spines["left"].set_color(LIGHT_RULE)
+    ax.spines["bottom"].set_color(LIGHT_RULE)
+    ax.tick_params(color=LIGHT_RULE, labelcolor=INK, width=0.65, length=2.8)
+
+
+def draw_profile_panel(
+    ax: plt.Axes,
+    generalist: np.ndarray,
+    d0_diag: np.ndarray,
+    warm_diag: np.ndarray,
+) -> None:
+    panel_title(ax, "A", "Matched-domain acceptance across five domains")
+    quiet_axes(ax)
+    ax.grid(axis="y", color=GRID, lw=0.6, alpha=0.9, zorder=0)
+    x = np.arange(len(DOMAINS))
+    ax.plot(
+        x,
+        generalist,
+        color=GENERALIST,
+        lw=1.65,
+        ls="-",
+        marker="o",
+        ms=4.2,
+        mfc="white",
+        mec=GENERALIST,
+        mew=1.0,
+        alpha=0.90,
+        label="Generalist",
+        zorder=3,
+    )
+    ax.plot(
+        x,
+        d0_diag,
+        color=D0,
+        lw=1.75,
+        marker="o",
+        ms=4.3,
+        mfc="white",
+        mec=D0,
+        mew=1.0,
+        label="D0-init MoS",
+        zorder=5,
+    )
+    ax.plot(
+        x,
+        warm_diag,
+        color=WARM,
+        lw=1.75,
+        marker="s",
+        ms=4.2,
+        mfc="white",
+        mec=WARM,
+        mew=1.0,
+        label="G-init MoS",
+        zorder=6,
+    )
+
+    for index, value in enumerate(generalist):
+        ax.annotate(
+            f"{value:.3f}",
+            xy=(x[index], value),
+            xytext=(0, -7),
+            textcoords="offset points",
+            ha="center",
+            va="top",
+            fontsize=5.7,
+            color=MUTED,
+        )
+    for index, value in enumerate(warm_diag):
+        ax.annotate(
+            f"{value:.3f}",
+            xy=(x[index], value),
+            xytext=(0, 6),
+            textcoords="offset points",
+            ha="center",
+            va="bottom",
+            fontsize=5.8,
+            color=INK,
+            fontweight="bold",
+        )
+
+    ax.set_xlim(-0.35, len(DOMAINS) - 0.65)
+    ax.set_ylim(2.55, 5.78)
+    ax.set_xticks(x, labels=DOMAIN_LABELS)
+    ax.set_yticks([2.8, 3.4, 4.0, 4.6, 5.2, 5.8])
+    ax.set_ylabel("Acceptance length")
+    ax.legend(
+        loc="upper right",
+        bbox_to_anchor=(1.0, 1.02),
+        frameon=False,
+        ncol=3,
+        handlelength=2.4,
+        borderaxespad=0.1,
+        columnspacing=1.15,
+        handletextpad=0.4,
+        labelspacing=0.25,
+        fontsize=5.8,
+    )
+
+
+HEATMAP_D0_CMAP = LinearSegmentedColormap.from_list(
+    "d0_matched_regret",
+    ["#F7FAF7", "#DDEFE1", "#A7D7B1", "#68B97A", D0],
+)
+HEATMAP_WARM_CMAP = LinearSegmentedColormap.from_list(
+    "warm_matched_regret",
+    ["#FBF8FC", "#F0E0F4", "#D9B7E2", "#BC79CB", WARM],
+)
+HEATMAP_NORM = Normalize(vmin=-1.30, vmax=0.0)
+
+
+def draw_matrix_panel(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    letter: str,
+    title: str,
+    cmap: LinearSegmentedColormap,
+    diagonal_edge: str,
+) -> mpl.image.AxesImage:
+    regret = matrix - np.diag(matrix)[None, :]
+    image = ax.imshow(regret, cmap=cmap, norm=HEATMAP_NORM, aspect="equal")
+    panel_title(ax, letter, title)
+    ax.set_xticks(range(len(DOMAINS)), labels=DOMAIN_LABELS)
+    ax.set_yticks(range(len(DOMAINS)), labels=DOMAIN_LABELS)
+    ax.tick_params(axis="x", rotation=29, length=0, pad=2.2, labelsize=6.2)
+    ax.tick_params(axis="y", length=0, pad=2.6, labelsize=6.5)
+    ax.set_ylabel("Selected MLP", labelpad=2.5, fontsize=7.0)
+
+    for row in range(len(DOMAINS)):
+        for column in range(len(DOMAINS)):
+            value = matrix[row, column]
+            normalized = HEATMAP_NORM(regret[row, column])
+            text_color = "white" if normalized > 0.68 else INK
+            ax.text(
+                column,
+                row,
+                f"{value:.3f}",
+                ha="center",
+                va="center",
+                fontsize=6.0,
+                color=text_color,
+                fontweight="bold" if row == column else "normal",
+            )
+            if row == column:
+                ax.add_patch(
+                    Rectangle(
+                        (column - 0.48, row - 0.48),
+                        0.96,
+                        0.96,
+                        facecolor="none",
+                        edgecolor=diagonal_edge,
+                        linewidth=1.45,
+                    )
+                )
+
+    ax.set_xticks(np.arange(-0.5, len(DOMAINS), 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, len(DOMAINS), 1), minor=True)
+    ax.grid(which="minor", color="white", linestyle="-", linewidth=1.15)
+    ax.tick_params(which="minor", bottom=False, left=False)
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+    return image
+
+
+def main() -> None:
+    args = parse_args()
+    configure_style()
+    evidence = json.loads(args.evidence.read_text())
+    if not evidence.get("passed"):
+        raise ValueError("R1 evidence is not marked passed")
+    if evidence.get("cells_total") != 52 or evidence.get("cells_passed") != 52:
+        raise ValueError("R1 evidence is not complete (expected 52/52 cells)")
+
+    d0_matrix = matrix_from_evidence(evidence, "panel_b_matrix_dflash_init")
+    warm_matrix = matrix_from_evidence(evidence, "panel_c_matrix_warm_start")
+    generalist = generalist_from_evidence(evidence)
+
+    if not np.all(np.diag(d0_matrix) > generalist):
+        raise ValueError("DFlash-initialized MoS is not above Generalist in every domain")
+    if not np.all(np.diag(warm_matrix) > generalist):
+        raise ValueError("warm-started MoS is not above Generalist in every domain")
+
+    # Match the intended AAAI double-column physical width. Raising DPI, rather
+    # than drawing an oversized canvas and shrinking it in LaTeX, preserves the
+    # configured 7--10 pt typography at publication size.
+    fig = plt.figure(figsize=(7.15, 5.00), facecolor="white")
+    outer = fig.add_gridspec(
+        2,
+        1,
+        height_ratios=[0.82, 1.18],
+        hspace=0.46,
+        left=0.075,
+        right=0.985,
+        top=0.945,
+        bottom=0.180,
+    )
+    ax_a = fig.add_subplot(outer[0, 0])
+    matrices = outer[1, 0].subgridspec(1, 2, wspace=0.28)
+    ax_b = fig.add_subplot(matrices[0, 0])
+    ax_c = fig.add_subplot(matrices[0, 1])
+
+    draw_profile_panel(
+        ax_a,
+        generalist,
+        np.diag(d0_matrix),
+        np.diag(warm_matrix),
+    )
+    d0_image = draw_matrix_panel(
+        ax_b,
+        d0_matrix,
+        "B",
+        "DFlash-initialized MoS",
+        HEATMAP_D0_CMAP,
+        "#226F32",
+    )
+    warm_image = draw_matrix_panel(
+        ax_c,
+        warm_matrix,
+        "C",
+        "Generalist-warm-started MoS",
+        HEATMAP_WARM_CMAP,
+        "#6F277D",
+    )
+
+    # Separate color strips preserve the original green/purple recipe identity
+    # while keeping an identical quantitative scale in both matrices.
+    for image, position in (
+        (d0_image, [0.145, 0.065, 0.29, 0.010]),
+        (warm_image, [0.575, 0.065, 0.29, 0.010]),
+    ):
+        cbar_ax = fig.add_axes(position)
+        cbar = fig.colorbar(image, cax=cbar_ax, orientation="horizontal")
+        cbar.set_ticks([-1.2, -0.6, 0.0])
+        cbar.ax.tick_params(labelsize=5.7, length=1.8, color=LIGHT_RULE)
+        cbar.outline.set_visible(False)
+    fig.text(
+        0.505,
+        0.014,
+        r"Shade: column-wise $\Delta$AL from the matched MLP",
+        ha="center",
+        va="bottom",
+        fontsize=6.0,
+        color=INK,
+    )
+
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(args.output, dpi=420, facecolor="white")
+    plt.close(fig)
+    print(f"saved {args.output}")
+
+
+if __name__ == "__main__":
+    main()

recipes/plotting/v1/plot_mos_5x5_gains.py

@@ -0,0 +1,929 @@
+#!/usr/bin/env python3
+"""Render separate 5x5 MoS routing matrices with Generalist gains.
+
+The frozen R1 evidence contains two selected-checkpoint matrices:
+
+1. MoS initialized from the public DFlash drafter (D0-init).
+2. MoS warm-started from the trained Generalist (G-init).
+
+The first two figures show all selected-MLP x evaluation-domain AL cells.  The
+right panel reports the matched-domain diagonal's absolute and relative
+improvement over one fixed-seed evaluation of the selected Generalist
+checkpoint.  A third figure shows that Generalist baseline across the five
+evaluation domains.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import matplotlib as mpl
+import matplotlib.pyplot as plt
+import numpy as np
+from matplotlib.colors import LinearSegmentedColormap, Normalize
+from matplotlib.patches import Rectangle
+
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+DEFAULT_EVIDENCE = (
+    REPO_ROOT
+    / "paper"
+    / "submission"
+    / "evidence"
+    / "r1_mainfig_seed20260719_20260721T0602Z_cells_summary.json"
+)
+DEFAULT_OUTPUT_DIR = REPO_ROOT / "paper" / "submission" / "figures"
+
+DOMAINS = ["code", "math", "factual_qa", "creative_writing", "general"]
+DOMAIN_LABELS = ["Code", "Math", "Factual QA", "Creative", "General"]
+
+INK = "#25313B"
+MUTED = "#68747E"
+RULE = "#D9DFE3"
+ROW_FILL = "#F3F5F6"
+D0 = "#2F9E44"
+WARM = "#9C36B5"
+GENERALIST = "#7F8790"
+
+HEATMAP_NORM = Normalize(vmin=-1.30, vmax=0.0)
+D0_CMAP = LinearSegmentedColormap.from_list(
+    "d0_regret", ["#F7FAF7", "#DDEFE1", "#A7D7B1", "#68B97A", D0]
+)
+WARM_CMAP = LinearSegmentedColormap.from_list(
+    "ginit_regret", ["#FBF8FC", "#F0E0F4", "#D9B7E2", "#BC79CB", WARM]
+)
+ABSOLUTE_CMAP = LinearSegmentedColormap.from_list(
+    "absolute_al", ["#F2F7FB", "#C9DEEE", "#80B7D5", "#3182BD", "#12538A"]
+)
+ABSOLUTE_NORM = Normalize(vmin=2.25, vmax=5.60)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--evidence", type=Path, default=DEFAULT_EVIDENCE)
+    parser.add_argument("--output-dir", type=Path, default=DEFAULT_OUTPUT_DIR)
+    return parser.parse_args()
+
+
+def configure_style() -> None:
+    mpl.rcParams.update(
+        {
+            "font.family": "sans-serif",
+            "font.sans-serif": [
+                "Arial",
+                "Helvetica",
+                "Liberation Sans",
+                "DejaVu Sans",
+            ],
+            "font.size": 8.0,
+            "axes.titlesize": 10.0,
+            "axes.labelsize": 8.2,
+            "xtick.labelsize": 7.4,
+            "ytick.labelsize": 7.4,
+            "pdf.fonttype": 42,
+            "ps.fonttype": 42,
+            "savefig.bbox": "tight",
+            "savefig.pad_inches": 0.035,
+        }
+    )
+
+
+def load_evidence(path: Path) -> tuple[dict, np.ndarray]:
+    evidence = json.loads(path.read_text())
+    if not evidence.get("passed"):
+        raise ValueError("R1 evidence is not marked passed")
+    if evidence.get("cells_total") != 52 or evidence.get("cells_passed") != 52:
+        raise ValueError("R1 evidence is incomplete; expected 52/52 passed cells")
+    generalist = np.asarray(
+        [float(evidence["panel_d_generalist"][domain]) for domain in DOMAINS],
+        dtype=float,
+    )
+    return evidence, generalist
+
+
+def matrix_from_evidence(evidence: dict, key: str) -> np.ndarray:
+    mapping = evidence[key]
+    matrix = np.asarray(
+        [[float(mapping[row][column]) for column in DOMAINS] for row in DOMAINS],
+        dtype=float,
+    )
+    for column, domain in enumerate(DOMAINS):
+        if int(np.argmax(matrix[:, column])) != column:
+            raise ValueError(f"{key}: matched MLP is not best for {domain}")
+    return matrix
+
+
+def draw_matrix(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    cmap: LinearSegmentedColormap,
+    accent: str,
+) -> mpl.image.AxesImage:
+    regret = matrix - np.diag(matrix)[None, :]
+    image = ax.imshow(regret, cmap=cmap, norm=HEATMAP_NORM, aspect="equal")
+    ax.set_xticks(range(5), labels=DOMAIN_LABELS)
+    ax.set_yticks(range(5), labels=DOMAIN_LABELS)
+    ax.tick_params(axis="x", rotation=28, length=0, pad=3.0)
+    ax.tick_params(axis="y", length=0, pad=3.0)
+    ax.xaxis.set_label_position("top")
+    ax.set_xlabel("Evaluation domain", labelpad=8.5, fontweight="bold")
+    ax.set_ylabel("Selected MLP", labelpad=6.0, fontweight="bold")
+
+    for row in range(5):
+        for column in range(5):
+            value = matrix[row, column]
+            normalized = HEATMAP_NORM(regret[row, column])
+            text_color = "white" if normalized > 0.66 else INK
+            ax.text(
+                column,
+                row,
+                f"{value:.3f}",
+                ha="center",
+                va="center",
+                fontsize=7.7,
+                color=text_color,
+                fontweight="bold" if row == column else "normal",
+            )
+            if row == column:
+                ax.add_patch(
+                    Rectangle(
+                        (column - 0.48, row - 0.48),
+                        0.96,
+                        0.96,
+                        facecolor="none",
+                        edgecolor=accent,
+                        linewidth=1.7,
+                    )
+                )
+
+    ax.set_xticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.grid(which="minor", color="white", linewidth=1.25)
+    ax.tick_params(which="minor", bottom=False, left=False)
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+    return image
+
+
+def draw_gain_table(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    generalist: np.ndarray,
+    accent: str,
+) -> None:
+    diagonal = np.diag(matrix)
+    delta = diagonal - generalist
+    percent = 100.0 * delta / generalist
+    if not np.all(delta > 0):
+        raise ValueError("matched-domain MoS does not improve every domain")
+
+    mean_generalist = float(np.mean(generalist))
+    mean_diagonal = float(np.mean(diagonal))
+    mean_delta = mean_diagonal - mean_generalist
+    mean_percent = 100.0 * mean_delta / mean_generalist
+
+    labels = DOMAIN_LABELS + ["Mean"]
+    deltas = np.concatenate([delta, [mean_delta]])
+    percents = np.concatenate([percent, [mean_percent]])
+
+    ax.set_xlim(0.0, 1.0)
+    ax.set_ylim(0.0, 1.0)
+    ax.axis("off")
+    ax.text(
+        0.02,
+        0.965,
+        "Matched MLP gain vs Generalist",
+        ha="left",
+        va="top",
+        fontsize=9.1,
+        fontweight="bold",
+        color=INK,
+    )
+    ax.text(0.02, 0.855, "Domain", ha="left", va="center", color=MUTED, fontweight="bold")
+    ax.text(0.68, 0.855, "Δ AL", ha="right", va="center", color=MUTED, fontweight="bold")
+    ax.text(0.98, 0.855, "Δ %", ha="right", va="center", color=MUTED, fontweight="bold")
+    ax.plot([0.02, 0.98], [0.815, 0.815], color=RULE, lw=0.9)
+
+    ys = np.linspace(0.735, 0.175, len(labels))
+    for index, (label, value, pct, y) in enumerate(zip(labels, deltas, percents, ys)):
+        if index == len(labels) - 1:
+            ax.add_patch(
+                Rectangle(
+                    (0.01, y - 0.050),
+                    0.98,
+                    0.100,
+                    facecolor=ROW_FILL,
+                    edgecolor="none",
+                    zorder=0,
+                )
+            )
+        weight = "bold" if index == len(labels) - 1 else "normal"
+        ax.text(0.02, y, label, ha="left", va="center", color=INK, fontweight=weight)
+        ax.text(
+            0.68,
+            y,
+            f"+{value:.3f}",
+            ha="right",
+            va="center",
+            color=accent,
+            fontweight="bold",
+        )
+        ax.text(
+            0.98,
+            y,
+            f"+{pct:.1f}%",
+            ha="right",
+            va="center",
+            color=accent,
+            fontweight="bold",
+        )
+
+    ax.text(
+        0.02,
+        0.045,
+        "Mean is unweighted across the five domains.",
+        ha="left",
+        va="bottom",
+        fontsize=6.7,
+        color=MUTED,
+    )
+
+
+def render_one(
+    matrix: np.ndarray,
+    generalist: np.ndarray,
+    title: str,
+    subtitle: str,
+    cmap: LinearSegmentedColormap,
+    accent: str,
+    output: Path,
+) -> None:
+    fig = plt.figure(figsize=(7.15, 3.55), facecolor="white")
+    grid = fig.add_gridspec(
+        1,
+        2,
+        width_ratios=[1.20, 0.92],
+        wspace=0.22,
+        left=0.085,
+        right=0.985,
+        top=0.755,
+        bottom=0.21,
+    )
+    ax_matrix = fig.add_subplot(grid[0, 0])
+    ax_gain = fig.add_subplot(grid[0, 1])
+
+    image = draw_matrix(ax_matrix, matrix, cmap, accent)
+    draw_gain_table(ax_gain, matrix, generalist, accent)
+
+    fig.text(0.03, 0.970, title, ha="left", va="top", fontsize=11.2, fontweight="bold", color=INK)
+    fig.text(0.03, 0.862, subtitle, ha="left", va="top", fontsize=7.2, color=MUTED)
+
+    cbar_ax = fig.add_axes([0.137, 0.095, 0.355, 0.018])
+    cbar = fig.colorbar(image, cax=cbar_ax, orientation="horizontal")
+    cbar.set_ticks([-1.2, -0.6, 0.0], labels=["−1.2", "−0.6", "0"])
+    cbar.ax.tick_params(labelsize=6.5, length=2.0, color=RULE, pad=1.5)
+    cbar.outline.set_visible(False)
+    fig.text(
+        0.314,
+        0.040,
+        "Cell shade: AL difference from the matched MLP in each column",
+        ha="center",
+        va="bottom",
+        fontsize=6.5,
+        color=MUTED,
+    )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=420, facecolor="white")
+    plt.close(fig)
+    print(f"saved {output}")
+
+
+def render_generalist(
+    generalist: np.ndarray,
+    subtitle: str,
+    output: Path,
+) -> None:
+    mean_al = float(np.mean(generalist))
+    x = np.arange(len(DOMAINS))
+
+    fig, ax = plt.subplots(figsize=(7.15, 3.35), facecolor="white")
+    fig.subplots_adjust(left=0.095, right=0.975, top=0.755, bottom=0.205)
+    bars = ax.bar(
+        x,
+        generalist,
+        width=0.58,
+        color=GENERALIST,
+        edgecolor=INK,
+        linewidth=0.55,
+        zorder=3,
+    )
+    ax.bar_label(
+        bars,
+        labels=[f"{value:.3f}" for value in generalist],
+        padding=-16,
+        fontsize=8.1,
+        fontweight="bold",
+        color="white",
+    )
+    ax.axhline(
+        mean_al,
+        color=INK,
+        lw=1.15,
+        ls=(0, (4, 2)),
+        label=f"Five-domain mean = {mean_al:.3f}",
+        zorder=2,
+    )
+
+    ax.set_xlim(-0.55, len(DOMAINS) - 0.45)
+    ax.set_ylim(0.0, 5.55)
+    ax.set_xticks(x, labels=DOMAIN_LABELS)
+    ax.set_yticks(np.arange(0.0, 5.6, 1.0))
+    ax.set_ylabel("Acceptance length (AL)", fontweight="bold")
+    ax.grid(axis="y", color=RULE, linewidth=0.65, zorder=0)
+    ax.legend(loc="upper right", frameon=False, fontsize=7.4, handlelength=2.8)
+    ax.spines["top"].set_visible(False)
+    ax.spines["right"].set_visible(False)
+    ax.spines["left"].set_color(RULE)
+    ax.spines["bottom"].set_color(RULE)
+    ax.tick_params(color=RULE, labelcolor=INK, width=0.65, length=2.8)
+
+    fig.text(
+        0.03,
+        0.970,
+        "Generalist (DFlash baseline): AL across five domains",
+        ha="left",
+        va="top",
+        fontsize=11.2,
+        fontweight="bold",
+        color=INK,
+    )
+    fig.text(0.03, 0.862, subtitle, ha="left", va="top", fontsize=7.2, color=MUTED)
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=420, facecolor="white")
+    plt.close(fig)
+    print(f"saved {output}")
+
+
+def draw_compact_generalist(ax: plt.Axes, generalist: np.ndarray) -> None:
+    x = np.arange(len(DOMAINS))
+    bars = ax.bar(
+        x,
+        generalist,
+        width=0.66,
+        color=GENERALIST,
+        edgecolor=INK,
+        linewidth=0.45,
+        zorder=3,
+    )
+    ax.bar_label(
+        bars,
+        labels=[f"{value:.3f}" for value in generalist],
+        padding=-10,
+        fontsize=5.7,
+        fontweight="bold",
+        color="white",
+    )
+    ax.axhline(float(np.mean(generalist)), color=INK, lw=0.85, ls=(0, (3, 2)), zorder=2)
+    ax.set_xlim(-0.55, len(DOMAINS) - 0.45)
+    ax.set_ylim(0.0, 5.55)
+    ax.set_xticks(x, labels=["Code", "Math", "FQA", "Creat.", "Gen."])
+    ax.tick_params(axis="x", rotation=40, labelsize=5.5, pad=1.8)
+    ax.set_yticks([0, 2, 4], labels=["0", "2", "4"])
+    ax.tick_params(axis="y", labelsize=5.5)
+    ax.set_ylabel("AL", fontsize=6.5, fontweight="bold", labelpad=2.0)
+    ax.grid(axis="y", color=RULE, linewidth=0.5, zorder=0)
+    ax.spines["top"].set_visible(False)
+    ax.spines["right"].set_visible(False)
+    ax.spines["left"].set_color(RULE)
+    ax.spines["bottom"].set_color(RULE)
+    ax.tick_params(color=RULE, labelcolor=INK, width=0.5, length=2.0)
+    ax.text(
+        0.98,
+        0.96,
+        f"mean {np.mean(generalist):.3f}",
+        transform=ax.transAxes,
+        ha="right",
+        va="top",
+        fontsize=5.8,
+        color=INK,
+        fontweight="bold",
+    )
+
+
+def draw_compact_matrix(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    cmap: LinearSegmentedColormap,
+    accent: str,
+) -> None:
+    regret = matrix - np.diag(matrix)[None, :]
+    ax.imshow(regret, cmap=cmap, norm=HEATMAP_NORM, aspect="equal")
+    short_labels = ["Code", "Math", "FQA", "Creat.", "Gen."]
+    ax.set_xticks(range(5), labels=short_labels)
+    ax.set_yticks(range(5), labels=short_labels)
+    ax.tick_params(axis="x", rotation=40, length=0, pad=1.8, labelsize=5.3)
+    ax.tick_params(axis="y", length=0, pad=2.0, labelsize=5.3)
+    ax.set_ylabel("Selected MLP", fontsize=6.1, fontweight="bold", labelpad=2.0)
+
+    for row in range(5):
+        for column in range(5):
+            normalized = HEATMAP_NORM(regret[row, column])
+            ax.text(
+                column,
+                row,
+                f"{matrix[row, column]:.2f}",
+                ha="center",
+                va="center",
+                fontsize=5.4,
+                color="white" if normalized > 0.66 else INK,
+                fontweight="bold" if row == column else "normal",
+            )
+            if row == column:
+                ax.add_patch(
+                    Rectangle(
+                        (column - 0.47, row - 0.47),
+                        0.94,
+                        0.94,
+                        facecolor="none",
+                        edgecolor=accent,
+                        linewidth=1.15,
+                    )
+                )
+
+    ax.set_xticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.grid(which="minor", color="white", linewidth=0.9)
+    ax.tick_params(which="minor", bottom=False, left=False)
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+
+
+def draw_compact_gains(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    generalist: np.ndarray,
+    accent: str,
+) -> None:
+    delta = np.diag(matrix) - generalist
+    percent = 100.0 * delta / generalist
+    mean_delta = float(np.mean(np.diag(matrix)) - np.mean(generalist))
+    mean_percent = 100.0 * mean_delta / float(np.mean(generalist))
+
+    ax.set_xlim(0.0, 1.0)
+    ax.set_ylim(4.5, -0.5)
+    ax.axis("off")
+    ax.text(0.43, 1.045, "ΔAL", transform=ax.transAxes, ha="right", va="bottom", fontsize=5.5, color=MUTED, fontweight="bold")
+    ax.text(0.98, 1.045, "Δ%", transform=ax.transAxes, ha="right", va="bottom", fontsize=5.5, color=MUTED, fontweight="bold")
+    for row, (value, pct) in enumerate(zip(delta, percent)):
+        ax.text(0.43, row, f"+{value:.2f}", ha="right", va="center", fontsize=5.4, color=accent, fontweight="bold")
+        ax.text(0.98, row, f"+{pct:.1f}", ha="right", va="center", fontsize=5.4, color=accent, fontweight="bold")
+    ax.text(
+        0.98,
+        -0.16,
+        f"mean +{mean_delta:.2f} / +{mean_percent:.1f}%",
+        transform=ax.transAxes,
+        ha="right",
+        va="top",
+        fontsize=5.1,
+        color=accent,
+        fontweight="bold",
+    )
+
+
+def render_three_panel(
+    generalist: np.ndarray,
+    d0_matrix: np.ndarray,
+    warm_matrix: np.ndarray,
+    output: Path,
+) -> None:
+    fig = plt.figure(figsize=(7.15, 2.48), facecolor="white")
+    outer = fig.add_gridspec(
+        1,
+        3,
+        width_ratios=[0.78, 1.36, 1.36],
+        wspace=0.30,
+        left=0.055,
+        right=0.992,
+        top=0.78,
+        bottom=0.23,
+    )
+    ax_a = fig.add_subplot(outer[0, 0])
+    grid_b = outer[0, 1].subgridspec(1, 2, width_ratios=[1.0, 0.42], wspace=0.04)
+    ax_b = fig.add_subplot(grid_b[0, 0])
+    ax_b_gain = fig.add_subplot(grid_b[0, 1])
+    grid_c = outer[0, 2].subgridspec(1, 2, width_ratios=[1.0, 0.42], wspace=0.04)
+    ax_c = fig.add_subplot(grid_c[0, 0])
+    ax_c_gain = fig.add_subplot(grid_c[0, 1])
+
+    draw_compact_generalist(ax_a, generalist)
+    draw_compact_matrix(ax_b, d0_matrix, D0_CMAP, D0)
+    draw_compact_gains(ax_b_gain, d0_matrix, generalist, D0)
+    draw_compact_matrix(ax_c, warm_matrix, WARM_CMAP, WARM)
+    draw_compact_gains(ax_c_gain, warm_matrix, generalist, WARM)
+
+    panel_titles = (
+        (0.055, "A", "Generalist (DFlash)"),
+        (0.305, "B", "DFlash-init MoS"),
+        (0.661, "C", "Generalist-warm-start MoS"),
+    )
+    for x, letter, title in panel_titles:
+        fig.text(x, 0.935, letter, ha="left", va="top", fontsize=8.8, fontweight="bold", color=INK)
+        fig.text(x + 0.025, 0.935, title, ha="left", va="top", fontsize=8.0, fontweight="bold", color=INK)
+
+    fig.text(
+        0.63,
+        0.055,
+        "Rows select MLPs; columns are evaluation domains. Bold diagonal = matched MLP; gains are vs Generalist.",
+        ha="center",
+        va="bottom",
+        fontsize=5.3,
+        color=MUTED,
+    )
+    fig.text(
+        0.055,
+        0.055,
+        "Qwen3-8B target · fixed seed",
+        ha="left",
+        va="bottom",
+        fontsize=5.3,
+        color=MUTED,
+    )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=480, facecolor="white")
+    plt.close(fig)
+    print(f"saved {output}")
+
+
+def draw_baseline_aligned_panel(
+    ax_matrix: plt.Axes,
+    ax_gain: plt.Axes,
+    matrix: np.ndarray,
+    generalist: np.ndarray,
+) -> None:
+    aligned = np.vstack([generalist, matrix])
+    row_labels = ["Generalist", "Code MLP", "Math MLP", "FQA MLP", "Creat. MLP", "Gen. MLP"]
+    column_labels = ["Code", "Math", "FQA", "Creat.", "Gen."]
+    ax_matrix.imshow(aligned, cmap=ABSOLUTE_CMAP, norm=ABSOLUTE_NORM, aspect="equal")
+    ax_matrix.set_xticks(range(5), labels=column_labels)
+    ax_matrix.set_yticks(range(6), labels=row_labels)
+    ax_matrix.tick_params(axis="x", rotation=37, length=0, pad=2.0, labelsize=5.5)
+    ax_matrix.tick_params(axis="y", length=0, pad=2.4, labelsize=5.4)
+
+    for row in range(6):
+        for column in range(5):
+            value = aligned[row, column]
+            normalized = ABSOLUTE_NORM(value)
+            is_matched = row > 0 and row - 1 == column
+            ax_matrix.text(
+                column,
+                row,
+                f"{value:.2f}",
+                ha="center",
+                va="center",
+                fontsize=5.7,
+                color="white" if normalized > 0.58 else INK,
+                fontweight="bold" if is_matched else "normal",
+            )
+            if is_matched:
+                ax_matrix.add_patch(
+                    Rectangle(
+                        (column - 0.47, row - 0.47),
+                        0.94,
+                        0.94,
+                        facecolor="none",
+                        edgecolor=INK,
+                        linewidth=1.0,
+                    )
+                )
+
+    ax_matrix.axhline(0.5, color=INK, lw=1.15)
+    ax_matrix.set_xticks(np.arange(-0.5, 5, 1), minor=True)
+    ax_matrix.set_yticks(np.arange(-0.5, 6, 1), minor=True)
+    ax_matrix.grid(which="minor", color="white", linewidth=0.9)
+    ax_matrix.tick_params(which="minor", bottom=False, left=False)
+    for spine in ax_matrix.spines.values():
+        spine.set_visible(False)
+
+    delta = np.diag(matrix) - generalist
+    percent = 100.0 * delta / generalist
+    mean_delta = float(np.mean(np.diag(matrix)) - np.mean(generalist))
+    mean_percent = 100.0 * mean_delta / float(np.mean(generalist))
+    ax_gain.set_xlim(0.0, 1.0)
+    ax_gain.set_ylim(5.5, -0.5)
+    ax_gain.axis("off")
+    ax_gain.text(0.43, 1.04, "ΔAL", transform=ax_gain.transAxes, ha="right", va="bottom", fontsize=5.7, color=MUTED, fontweight="bold")
+    ax_gain.text(0.98, 1.04, "Δ%", transform=ax_gain.transAxes, ha="right", va="bottom", fontsize=5.7, color=MUTED, fontweight="bold")
+    ax_gain.text(0.43, 0, "—", ha="right", va="center", fontsize=5.4, color=MUTED)
+    ax_gain.text(0.98, 0, "—", ha="right", va="center", fontsize=5.4, color=MUTED)
+    for row, (value, pct) in enumerate(zip(delta, percent), start=1):
+        ax_gain.text(0.43, row, f"+{value:.2f}", ha="right", va="center", fontsize=5.5, color=INK, fontweight="bold")
+        ax_gain.text(0.98, row, f"+{pct:.1f}", ha="right", va="center", fontsize=5.5, color=INK, fontweight="bold")
+    ax_gain.axhline(0.5, color=INK, lw=1.15)
+    ax_gain.text(
+        0.98,
+        -0.14,
+        f"mean +{mean_delta:.2f} / +{mean_percent:.1f}%",
+        transform=ax_gain.transAxes,
+        ha="right",
+        va="top",
+        fontsize=5.2,
+        color=INK,
+        fontweight="bold",
+    )
+
+
+def render_baseline_aligned(
+    generalist: np.ndarray,
+    d0_matrix: np.ndarray,
+    warm_matrix: np.ndarray,
+    output: Path,
+) -> None:
+    fig = plt.figure(figsize=(7.15, 2.85), facecolor="white")
+    outer = fig.add_gridspec(
+        1,
+        2,
+        wspace=0.28,
+        left=0.105,
+        right=0.992,
+        top=0.72,
+        bottom=0.23,
+    )
+    grid_a = outer[0, 0].subgridspec(1, 2, width_ratios=[1.0, 0.35], wspace=0.04)
+    ax_a = fig.add_subplot(grid_a[0, 0])
+    ax_a_gain = fig.add_subplot(grid_a[0, 1])
+    grid_b = outer[0, 1].subgridspec(1, 2, width_ratios=[1.0, 0.35], wspace=0.04)
+    ax_b = fig.add_subplot(grid_b[0, 0])
+    ax_b_gain = fig.add_subplot(grid_b[0, 1])
+
+    draw_baseline_aligned_panel(ax_a, ax_a_gain, d0_matrix, generalist)
+    draw_baseline_aligned_panel(ax_b, ax_b_gain, warm_matrix, generalist)
+
+    fig.text(
+        0.055,
+        0.970,
+        "Generalist-aligned acceptance-length matrices · Qwen3-8B target",
+        ha="left",
+        va="top",
+        fontsize=9.2,
+        fontweight="bold",
+        color=INK,
+    )
+    fig.text(
+        0.055,
+        0.895,
+        f"Shared DFlash Generalist baseline mean = {np.mean(generalist):.3f}; fixed-seed selected-checkpoint evaluation",
+        ha="left",
+        va="top",
+        fontsize=6.0,
+        color=MUTED,
+    )
+    fig.text(0.105, 0.805, "A   DFlash-init MoS", ha="left", va="top", fontsize=7.5, fontweight="bold", color=INK)
+    fig.text(0.563, 0.805, "B   Generalist-warm-start MoS", ha="left", va="top", fontsize=7.5, fontweight="bold", color=INK)
+    fig.text(
+        0.50,
+        0.045,
+        "The shared Generalist row is repeated for direct comparison; it is one baseline, not five specialists. Bold boxes mark matched MLPs.",
+        ha="center",
+        va="bottom",
+        fontsize=5.2,
+        color=MUTED,
+    )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=480, facecolor="white")
+    plt.close(fig)
+    print(f"saved {output}")
+
+
+def draw_summary_matrix(
+    ax: plt.Axes,
+    matrix: np.ndarray,
+    cmap: LinearSegmentedColormap,
+    accent: str,
+) -> None:
+    regret = matrix - np.diag(matrix)[None, :]
+    ax.imshow(regret, cmap=cmap, norm=HEATMAP_NORM, aspect="equal")
+    labels = ["Code", "Math", "FQA", "Creat.", "Gen."]
+    ax.set_xticks(range(5), labels=labels)
+    ax.set_yticks(range(5), labels=labels)
+    ax.tick_params(axis="x", rotation=38, length=0, pad=2.0, labelsize=5.4)
+    ax.tick_params(axis="y", length=0, pad=2.2, labelsize=5.4)
+    ax.set_ylabel("Selected MLP", fontsize=6.1, fontweight="bold", labelpad=2.2)
+
+    for row in range(5):
+        for column in range(5):
+            normalized = HEATMAP_NORM(regret[row, column])
+            is_matched = row == column
+            ax.text(
+                column,
+                row,
+                f"{matrix[row, column]:.3f}",
+                ha="center",
+                va="center",
+                fontsize=5.2,
+                color="white" if normalized > 0.66 else INK,
+                fontweight="bold" if is_matched else "normal",
+            )
+            if is_matched:
+                ax.add_patch(
+                    Rectangle(
+                        (column - 0.47, row - 0.47),
+                        0.94,
+                        0.94,
+                        facecolor="none",
+                        edgecolor=accent,
+                        linewidth=1.15,
+                    )
+                )
+
+    ax.set_xticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, 5, 1), minor=True)
+    ax.grid(which="minor", color="white", linewidth=0.95)
+    ax.tick_params(which="minor", bottom=False, left=False)
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+
+
+def draw_three_method_table(
+    ax: plt.Axes,
+    generalist: np.ndarray,
+    d0_matrix: np.ndarray,
+    warm_matrix: np.ndarray,
+) -> None:
+    d0 = np.diag(d0_matrix)
+    warm = np.diag(warm_matrix)
+    labels = DOMAIN_LABELS + ["Mean"]
+    generalist_values = np.concatenate([generalist, [np.mean(generalist)]])
+    d0_values = np.concatenate([d0, [np.mean(d0)]])
+    warm_values = np.concatenate([warm, [np.mean(warm)]])
+
+    ax.set_xlim(0.0, 1.0)
+    ax.set_ylim(0.0, 1.0)
+    ax.axis("off")
+    header_y = 0.875
+    ax.text(0.01, header_y, "Domain", ha="left", va="center", fontsize=5.9, color=MUTED, fontweight="bold")
+    ax.text(0.48, header_y, "Generalist", ha="right", va="center", fontsize=5.7, color=MUTED, fontweight="bold")
+    ax.text(0.75, header_y, "D0 MoS", ha="right", va="center", fontsize=5.7, color=D0, fontweight="bold")
+    ax.text(0.99, header_y, "G-init", ha="right", va="center", fontsize=5.7, color=WARM, fontweight="bold")
+    ax.plot([0.01, 0.99], [0.825, 0.825], color=RULE, lw=0.8)
+
+    ys = np.linspace(0.745, 0.245, len(labels))
+    for index, (label, gen_value, d0_value, warm_value, y) in enumerate(
+        zip(labels, generalist_values, d0_values, warm_values, ys)
+    ):
+        is_mean = index == len(labels) - 1
+        if is_mean:
+            ax.add_patch(
+                Rectangle(
+                    (0.0, y - 0.045),
+                    1.0,
+                    0.090,
+                    facecolor=ROW_FILL,
+                    edgecolor="none",
+                    zorder=0,
+                )
+            )
+        weight = "bold" if is_mean else "normal"
+        ax.text(0.01, y, label, ha="left", va="center", fontsize=5.8, color=INK, fontweight=weight)
+        ax.text(0.48, y, f"{gen_value:.3f}", ha="right", va="center", fontsize=5.8, color=MUTED, fontweight=weight)
+        ax.text(0.75, y, f"{d0_value:.3f}", ha="right", va="center", fontsize=5.8, color=D0, fontweight="bold")
+        ax.text(0.99, y, f"{warm_value:.3f}", ha="right", va="center", fontsize=5.8, color=WARM, fontweight="bold")
+
+    d0_delta = float(np.mean(d0) - np.mean(generalist))
+    warm_delta = float(np.mean(warm) - np.mean(generalist))
+    d0_percent = 100.0 * d0_delta / float(np.mean(generalist))
+    warm_percent = 100.0 * warm_delta / float(np.mean(generalist))
+    ax.text(
+        0.99,
+        0.105,
+        f"D0 mean gain  +{d0_delta:.3f} / +{d0_percent:.1f}%",
+        ha="right",
+        va="center",
+        fontsize=5.4,
+        color=D0,
+        fontweight="bold",
+    )
+    ax.text(
+        0.99,
+        0.035,
+        f"G-init mean gain  +{warm_delta:.3f} / +{warm_percent:.1f}%",
+        ha="right",
+        va="center",
+        fontsize=5.4,
+        color=WARM,
+        fontweight="bold",
+    )
+
+
+def render_matrices_summary(
+    generalist: np.ndarray,
+    d0_matrix: np.ndarray,
+    warm_matrix: np.ndarray,
+    output: Path,
+) -> None:
+    fig = plt.figure(figsize=(7.15, 2.52), facecolor="white")
+    grid = fig.add_gridspec(
+        1,
+        3,
+        width_ratios=[1.0, 1.0, 1.18],
+        wspace=0.28,
+        left=0.065,
+        right=0.992,
+        top=0.77,
+        bottom=0.22,
+    )
+    ax_d0 = fig.add_subplot(grid[0, 0])
+    ax_warm = fig.add_subplot(grid[0, 1])
+    ax_table = fig.add_subplot(grid[0, 2])
+
+    draw_summary_matrix(ax_d0, d0_matrix, D0_CMAP, D0)
+    draw_summary_matrix(ax_warm, warm_matrix, WARM_CMAP, WARM)
+    draw_three_method_table(ax_table, generalist, d0_matrix, warm_matrix)
+
+    titles = (
+        (0.065, "A", "DFlash-init MoS"),
+        (0.360, "B", "Generalist-warm-start MoS"),
+        (0.670, "C", "Matched-domain AL"),
+    )
+    for x, letter, title in titles:
+        fig.text(x, 0.940, letter, ha="left", va="top", fontsize=8.7, fontweight="bold", color=INK)
+        fig.text(x + 0.025, 0.940, title, ha="left", va="top", fontsize=7.5, fontweight="bold", color=INK)
+
+    fig.text(
+        0.50,
+        0.045,
+        "Qwen3-8B target · fixed-seed selected checkpoints · matrix columns are evaluation domains; bold diagonal cells select the matched MLP.",
+        ha="center",
+        va="bottom",
+        fontsize=5.2,
+        color=MUTED,
+    )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=480, facecolor="white")
+    plt.close(fig)
+    print(f"saved {output}")
+
+
+def main() -> None:
+    args = parse_args()
+    configure_style()
+    evidence, generalist = load_evidence(args.evidence)
+
+    d0_matrix = matrix_from_evidence(evidence, "panel_b_matrix_dflash_init")
+    warm_matrix = matrix_from_evidence(evidence, "panel_c_matrix_warm_start")
+
+    subtitle = (
+        "Qwen3-8B target · selected-checkpoint, fixed-seed evaluation · "
+        "rows: selected MLP; columns: evaluation domain"
+    )
+    render_one(
+        d0_matrix,
+        generalist,
+        "DFlash-initialized MoS: 5×5 routing matrix",
+        subtitle,
+        D0_CMAP,
+        D0,
+        args.output_dir / "fig_mos_d0_matrix_gains.png",
+    )
+    render_one(
+        warm_matrix,
+        generalist,
+        "Generalist-warm-started MoS: 5×5 routing matrix",
+        subtitle,
+        WARM_CMAP,
+        WARM,
+        args.output_dir / "fig_mos_ginit_matrix_gains.png",
+    )
+    render_generalist(
+        generalist,
+        "Qwen3-8B target · selected-checkpoint, fixed-seed evaluation · standard single-model DFlash",
+        args.output_dir / "fig_generalist_domain_al.png",
+    )
+    render_three_panel(
+        generalist,
+        d0_matrix,
+        warm_matrix,
+        args.output_dir / "fig_mos_three_panel.png",
+    )
+    render_baseline_aligned(
+        generalist,
+        d0_matrix,
+        warm_matrix,
+        args.output_dir / "fig_mos_baseline_aligned.png",
+    )
+    render_matrices_summary(
+        generalist,
+        d0_matrix,
+        warm_matrix,
+        args.output_dir / "fig_mos_matrices_summary.png",
+    )
+
+
+if __name__ == "__main__":
+    main()