export m2/soft-pool-scout

exports/m2/soft-pool-scout/_assets/provenance.json

@@ -1,20 +1,703 @@
 {
  "refs": {
+  "reports/m2/edit-effects/arrays/falloff-s704": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:27.745107+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s705": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:29.289254+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s706": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:30.206788+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s707": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:30.913068+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s708": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:31.591037+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s709": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:32.357757+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s710": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:33.056452+00:00"
+  },
+  "reports/m2/edit-effects/arrays/falloff-s711": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:33.793483+00:00"
+  },
+  "reports/m2/edit-effects/fits": {
+   "experiment": "edit-effects",
+   "git_describe": "835649c-dirty",
+   "git_dirty": true,
+   "git_sha": "835649c7ff951548d0d098987bbc46e66125032c",
+   "run_at": "2026-10-21T03:23:55.151469+00:00",
+   "task": "publish-f6d20f2311fb",
+   "written_at": "2026-10-21T03:35:40.120498+00:00"
+  },
+  "reports/m2/embedding-lean/results": {
+   "experiment": "embedding-lean",
+   "git_describe": "fb9b762",
+   "git_sha": "fb9b762d4511cfc203c7c0fc900c4338ca326aab",
+   "run_at": "2026-10-09T05:07:46.906138+00:00",
+   "task": "publish-d18f59dd23c4",
+   "written_at": "2026-10-09T05:35:29.953556+00:00"
+  },
+  "reports/m2/loss-rate/trajectories": {
+   "experiment": "loss-rate",
+   "git_describe": "9a37845-dirty",
+   "git_dirty": true,
+   "git_sha": "9a37845875e95595470c87736f759ea0b8545b9f",
+   "run_at": "2026-10-21T03:31:15.451431+00:00",
+   "task": "publish_screen-c94dd1528e55",
+   "written_at": "2026-10-23T01:13:44.219373+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s704": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:40.040694+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s705": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:40.967295+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s706": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:41.942577+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s707": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:42.864585+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s708": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:43.767611+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s709": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:44.668364+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s710": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:45.583180+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/pool-123-soft-s711": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:46.571968+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s704": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:27:15.463808+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s705": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:33.163134+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s706": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:34.218755+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s707": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:35.152510+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s708": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:36.110116+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s709": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:37.084273+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s710": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:38.174476+00:00"
+  },
+  "reports/m2/recipe-compare/write-arrays/step-soft-s711": {
+   "experiment": "recipe-compare",
+   "git_describe": "60a5c40",
+   "git_dirty": true,
+   "git_sha": "60a5c40501379be75da8381a96547db153f56001",
+   "run_at": "2026-10-24T11:16:35.921193+00:00",
+   "task": "publish-2abb510145a2",
+   "written_at": "2026-10-24T11:28:39.044136+00:00"
+  },
+  "reports/m2/reference-baseline/axes": {
+   "experiment": "reference-baseline",
+   "git_describe": "dc95e0f-dirty",
+   "git_dirty": true,
+   "git_sha": "dc95e0fef6af42ebfafc518b0a8ca777687c9bf2",
+   "run_at": "2026-10-24T03:17:34.857273+00:00",
+   "task": "publish-98c00b47880b",
+   "written_at": "2026-10-24T07:37:19.307965+00:00"
+  },
+  "reports/m2/soft-pool-scout/along": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:55.321603+00:00"
+  },
+  "reports/m2/soft-pool-scout/axes": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:28.369911+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:25.298701+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s704": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:02.670644+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s705": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:03.978101+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s706": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:05.310398+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s707": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:06.692304+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s708": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:08.058055+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s709": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:09.435166+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s710": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:10.761020+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s711": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:12.073970+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s704": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:13.382430+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s705": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:14.700396+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s706": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:15.985363+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s707": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:17.301489+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s708": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:18.636995+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s709": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:19.947552+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s710": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:21.230777+00:00"
+  },
+  "reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s711": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:22.561240+00:00"
+  },
+  "reports/m2/soft-pool-scout/results": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:01.292321+00:00"
+  },
   "reports/m2/soft-pool-scout/screen": {
    "experiment": "soft-pool-scout",
-   "git_describe": "7696eea",
-   "git_sha": "7696eea9c4e6854754a5f72bd44565bd288c5701",
+   "git_describe": "aa45e7f",
+   "git_sha": "aa45e7f6a2083242d99a554c2b5f62d556d17880",
    "run_at": "2026-10-26T05:35:59.041082+00:00",
    "task": "publish_screen-1ecf85a8af1f",
-   "written_at": "2026-10-26T06:57:09.955235+00:00"
+   "written_at": "2026-10-26T08:42:35.364750+00:00"
   },
   "reports/m2/soft-pool-scout/screen-trajectories": {
    "experiment": "soft-pool-scout",
-   "git_describe": "7696eea",
-   "git_sha": "7696eea9c4e6854754a5f72bd44565bd288c5701",
+   "git_describe": "aa45e7f",
+   "git_sha": "aa45e7f6a2083242d99a554c2b5f62d556d17880",
    "run_at": "2026-10-26T05:35:59.041082+00:00",
    "task": "publish_screen-1ecf85a8af1f",
-   "written_at": "2026-10-26T06:57:03.780201+00:00"
+   "written_at": "2026-10-26T08:42:31.938615+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s704": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:29.701227+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s705": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:31.015484+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s706": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:32.317200+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s707": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:33.638379+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s708": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:34.959237+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s709": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:36.307946+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s710": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:37.664376+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s711": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:39.062690+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s704": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:40.378356+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s705": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:41.699796+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s706": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:43.049267+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s707": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:44.359198+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s708": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:45.771589+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s709": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:47.172153+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s710": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:48.478156+00:00"
+  },
+  "reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s711": {
+   "experiment": "soft-pool-scout",
+   "git_describe": "4017bcf",
+   "git_sha": "4017bcf54b32143093e391119741adcbd0b37a22",
+   "run_at": "2026-10-26T05:35:59.041082+00:00",
+   "task": "publish_finished-37e3addf23a6",
+   "written_at": "2026-10-26T10:17:49.800645+00:00"
+  },
+  "reports/m2/soft-pool/dose": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:23.488978+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s704": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:17.221510+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s705": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:17.745713+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s706": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:18.320940+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s707": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:18.905287+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s708": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:19.455498+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s709": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:20.002577+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s710": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:20.546708+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/pool-123-soft-s711": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:21.168612+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s704": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:03.464100+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s705": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:04.003767+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s706": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:04.586284+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s707": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:05.121296+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s708": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:05.699928+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s709": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:06.251953+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s710": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:06.771424+00:00"
+  },
+  "reports/m2/soft-pool/dose-arrays/step-soft-s711": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:07.288701+00:00"
+  },
+  "reports/m2/soft-pool/results": {
+   "experiment": "soft-pool",
+   "git_describe": "434a7a7",
+   "git_sha": "434a7a7a7dd70fb070d778bddefac5d9ac98ba2b",
+   "run_at": "2026-10-20T02:58:20.784880+00:00",
+   "task": "publish-7e29dbaee0fe",
+   "written_at": "2026-10-20T04:42:02.859483+00:00"
   },
   "reports/m2/spill-timing/stored-checkpoints": {
    "experiment": "spill-timing",
@@ -31,6 +714,150 @@
    "run_at": "2026-10-23T06:39:05.717702+00:00",
    "task": "publish-59e8f5540fe0",
    "written_at": "2026-10-23T07:29:21.916350+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:46.236064+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s704": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:37.531774+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s705": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:38.521720+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s706": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:39.448114+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s707": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:40.390841+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s708": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:41.360512+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s709": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:42.335240+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s710": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:43.280197+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/control-s711": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:44.215312+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s704": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:29.717595+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s705": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:30.685197+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s706": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:31.651638+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s707": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:32.614755+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s708": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:33.623860+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s709": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:34.595525+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s710": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:35.603892+00:00"
+  },
+  "reports/m2/split-anti-seeds/dose-arrays/falloff-s711": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:36.566340+00:00"
+  },
+  "reports/m2/split-anti-seeds/results": {
+   "experiment": "split-anti-seeds",
+   "git_describe": "87826ef",
+   "git_sha": "87826ef418b2824e9e6a3d02f6abc1db8063ff7f",
+   "run_at": "2026-10-13T02:15:33.586485+00:00",
+   "task": "publish-2a8f91e85a47",
+   "written_at": "2026-10-13T04:39:04.233299+00:00"
   }
  }
 }

exports/m2/soft-pool-scout/index.html

@@ -1063,31 +1063,36 @@ main.lit details:not(:open) {
 @media print{[data-mini-banner],[data-mini-provenance]{display:none}}</style>
 </head>
 <body>
-    <details data-mini-provenance><summary>Data provenance</summary><div><strong>soft-pool-scout</strong> · <code>7696eea</code> · run 2026-10-26<div>via reports/m2/soft-pool-scout/screen, reports/m2/soft-pool-scout/screen-trajectories</div></div><div><strong>spill-timing</strong> · <code>bb3dc2d</code> · run 2026-10-23<div>via reports/m2/spill-timing/stored-checkpoints, reports/m2/spill-timing/trajectories</div></div></details>
+    <details data-mini-provenance><summary>Data provenance</summary><div><strong>edit-effects</strong> · <code>835649c-dirty</code> (dirty) · run 2026-10-21<div>via reports/m2/edit-effects/arrays/falloff-s704, reports/m2/edit-effects/arrays/falloff-s705, reports/m2/edit-effects/arrays/falloff-s706, reports/m2/edit-effects/arrays/falloff-s707, reports/m2/edit-effects/arrays/falloff-s708, reports/m2/edit-effects/arrays/falloff-s709, reports/m2/edit-effects/arrays/falloff-s710, reports/m2/edit-effects/arrays/falloff-s711, reports/m2/edit-effects/fits</div></div><div><strong>embedding-lean</strong> · <code>fb9b762</code> · run 2026-10-09<div>via reports/m2/embedding-lean/results</div></div><div><strong>loss-rate</strong> · <code>9a37845-dirty</code> (dirty) · run 2026-10-21<div>via reports/m2/loss-rate/trajectories</div></div><div><strong>recipe-compare</strong> · <code>60a5c40</code> (dirty) · run 2026-10-24<div>via reports/m2/recipe-compare/write-arrays/pool-123-soft-s704, reports/m2/recipe-compare/write-arrays/pool-123-soft-s705, reports/m2/recipe-compare/write-arrays/pool-123-soft-s706, reports/m2/recipe-compare/write-arrays/pool-123-soft-s707, reports/m2/recipe-compare/write-arrays/pool-123-soft-s708, reports/m2/recipe-compare/write-arrays/pool-123-soft-s709, reports/m2/recipe-compare/write-arrays/pool-123-soft-s710, reports/m2/recipe-compare/write-arrays/pool-123-soft-s711, reports/m2/recipe-compare/write-arrays/step-soft-s704, reports/m2/recipe-compare/write-arrays/step-soft-s705, reports/m2/recipe-compare/write-arrays/step-soft-s706, reports/m2/recipe-compare/write-arrays/step-soft-s707, reports/m2/recipe-compare/write-arrays/step-soft-s708, reports/m2/recipe-compare/write-arrays/step-soft-s709, reports/m2/recipe-compare/write-arrays/step-soft-s710, reports/m2/recipe-compare/write-arrays/step-soft-s711</div></div><div><strong>reference-baseline</strong> · <code>dc95e0f-dirty</code> (dirty) · run 2026-10-24<div>via reports/m2/reference-baseline/axes</div></div><div><strong>soft-pool</strong> · <code>434a7a7</code> · run 2026-10-20<div>via reports/m2/soft-pool/dose, reports/m2/soft-pool/dose-arrays/pool-123-soft-s704, reports/m2/soft-pool/dose-arrays/pool-123-soft-s705, reports/m2/soft-pool/dose-arrays/pool-123-soft-s706, reports/m2/soft-pool/dose-arrays/pool-123-soft-s707, reports/m2/soft-pool/dose-arrays/pool-123-soft-s708, reports/m2/soft-pool/dose-arrays/pool-123-soft-s709, reports/m2/soft-pool/dose-arrays/pool-123-soft-s710, reports/m2/soft-pool/dose-arrays/pool-123-soft-s711, reports/m2/soft-pool/dose-arrays/step-soft-s704, reports/m2/soft-pool/dose-arrays/step-soft-s705, reports/m2/soft-pool/dose-arrays/step-soft-s706, reports/m2/soft-pool/dose-arrays/step-soft-s707, reports/m2/soft-pool/dose-arrays/step-soft-s708, reports/m2/soft-pool/dose-arrays/step-soft-s709, reports/m2/soft-pool/dose-arrays/step-soft-s710, reports/m2/soft-pool/dose-arrays/step-soft-s711, reports/m2/soft-pool/results</div></div><div><strong>soft-pool-scout</strong> · <code>4017bcf</code> · run 2026-10-26<div>via reports/m2/soft-pool-scout/along, reports/m2/soft-pool-scout/axes, reports/m2/soft-pool-scout/dose, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s704, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s705, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s706, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s707, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s708, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s709, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s710, reports/m2/soft-pool-scout/dose-arrays/lt0.1-by40-s711, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s704, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s705, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s706, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s707, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s708, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s709, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s710, reports/m2/soft-pool-scout/dose-arrays/lt0.1-fixed-lt-s711, reports/m2/soft-pool-scout/results, reports/m2/soft-pool-scout/screen, reports/m2/soft-pool-scout/screen-trajectories, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s704, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s705, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s706, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s707, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s708, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s709, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s710, reports/m2/soft-pool-scout/write-arrays/lt0.1-by40-s711, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s704, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s705, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s706, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s707, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s708, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s709, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s710, reports/m2/soft-pool-scout/write-arrays/lt0.1-fixed-lt-s711</div></div><div><strong>spill-timing</strong> · <code>bb3dc2d</code> · run 2026-10-23<div>via reports/m2/spill-timing/stored-checkpoints, reports/m2/spill-timing/trajectories</div></div><div><strong>split-anti-seeds</strong> · <code>87826ef</code> · run 2026-10-13<div>via reports/m2/split-anti-seeds/dose, reports/m2/split-anti-seeds/dose-arrays/control-s704, reports/m2/split-anti-seeds/dose-arrays/control-s705, reports/m2/split-anti-seeds/dose-arrays/control-s706, reports/m2/split-anti-seeds/dose-arrays/control-s707, reports/m2/split-anti-seeds/dose-arrays/control-s708, reports/m2/split-anti-seeds/dose-arrays/control-s709, reports/m2/split-anti-seeds/dose-arrays/control-s710, reports/m2/split-anti-seeds/dose-arrays/control-s711, reports/m2/split-anti-seeds/dose-arrays/falloff-s704, reports/m2/split-anti-seeds/dose-arrays/falloff-s705, reports/m2/split-anti-seeds/dose-arrays/falloff-s706, reports/m2/split-anti-seeds/dose-arrays/falloff-s707, reports/m2/split-anti-seeds/dose-arrays/falloff-s708, reports/m2/split-anti-seeds/dose-arrays/falloff-s709, reports/m2/split-anti-seeds/dose-arrays/falloff-s710, reports/m2/split-anti-seeds/dose-arrays/falloff-s711, reports/m2/split-anti-seeds/results</div></div></details>
 <main class="lit">
 <h1 id="soft-pool-scout">Soft pool scout<a class="anchor-link" href="#soft-pool-scout" title="Link to this heading">&para;</a></h1>
 <div class="admonition tip">
 <!-- lede -->
-<p>We screened sixteen schedules of the pooled pull at epoch 64, looking for one that stops the early <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> of <code>pool 1–3, soft τ</code>. A sharper pool over slices did it: with the slice temperature ending at 0.1, every run spilled about as little as the quiet recipes, whenever the temperatures sharpened. Moving the anchor warm-up did not help, and fixing the temperatures from the start spilled more.</p>
+<p>A sharper pool over slices stopped the early <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> of <code>pool 1–3, soft τ</code>. Of sixteen schedules of the pooled pull screened at epoch 64, those with the slice temperature ending at 0.1 had no run spill early, however soon or late the temperatures sharpened.</p>
+<p>We carried two of these conditions on to epoch 400. Both stay within the spill criterion on every seed, and both remove a little more than the recipe of record at no cost to the task. Holding the slice temperature at 0.1 from the start is the cleaner of the two, and a candidate for the recipe of record.</p>
 </div>
 <p><a href="/docs/m2/recipe-compare/report.py">Recipe-compare</a> left one thing between <code>pool 1–3, soft τ</code> and the recipe of record: two of its eight runs go over the spill criterion, and the spill enters at block 1, the one slice the pool pulls and the step leaves alone. <a href="/docs/m2/spill-timing/report.py">Spill-timing</a> found that both of those runs spill early, while the pool sharpens, and that this early kind already shows at epoch 64. So a schedule that keeps block 1 closed should show it at epoch 64, at a sixth of the cost of a full run.</p>
-<p>This report trains sixteen conditions of the pool at its eight seeds (704 to 711), each changing when the two pool temperatures sharpen, how sharp the pool over slices ends, or when the anchor weight warms up. Every run stops at epoch 64, where we score whether it is keeping the anchor and how much the edit spills. This is stage 1; the leaders will be carried on to epoch 400 from where they stopped, in a second stage still to run.</p>
+<p>This report trains sixteen conditions of the pool at its eight seeds (704 to 711), each changing when the two pool temperatures sharpen, how sharp the pool over slices ends, or when the anchor weight warms up. Every run stops at epoch 64, where we score whether it is keeping the anchor and how much the edit spills. This is stage 1. In stage 2, the two conditions with <span class="arithmatex">\(\tau_\ell\)</span> ending at 0.1 that change least from the base (sharpening by epoch 40 as the base does, and held at 0.1 from the start) are carried on to epoch 400 from where they stopped, and measured as recipe-compare measured the three recipes, with the spill along training from checkpoints kept on the way.</p>
 <h2 id="observations">Observations<a class="anchor-link" href="#observations" title="Link to this heading">&para;</a></h2>
 <ul>
-<li><a href="#lost-anchors-e1">Lost anchors (E1)</a>: no schedule lost the anchor. All 136 runs, the base included, are called kept, the lowest well above the line, though the fixed temperatures and the stretched warm-up keep it a little less firmly.</li>
-<li><a href="#spill-at-epoch-64-e2">Spill at epoch 64 (E2)</a>: the slice temperature at the end decides the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span>, and when the temperatures sharpen hardly matters. At τ_ℓ 0.1 no run is over the criterion, and the four conditions sit at or below the quiet recipes. At 0.4 (the base) about a quarter of the runs are over. The delayed and stretched warm-ups spill as much as the base, and fixing the temperatures from the start spills the most.</li>
-<li><a href="#where-the-op-is-heading-e3">Where the op is heading (E3)</a>: every condition peaks at slice 3, and none leans on the query answer. What changes is slice 1: a sharper pool over slices puts less of the anchor there, and the runs that spill are the ones with more of it at slice 1.</li>
+<li><a href="#lost-anchors-e1">Lost anchors (E1)</a>: no schedule lost the anchor. All 136 runs, the base included, are called kept, the lowest well above the line. The fixed temperatures and the stretched warm-up keep it a little less firmly.</li>
+<li><a href="#spill-at-epoch-64-e2">Spill at epoch 64 (E2)</a>: the final slice temperature decides the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span>; when the temperatures sharpen hardly matters. At τ_ℓ 0.1 no run is over the criterion, and the four conditions sit at or below the quiet recipes. At 0.4 (the base) about a quarter of the runs are over. Moving the <span class="arithmatex">\(\lambda_a\)</span> warm-up shifts the spill by less than the seeds vary, and fixing the temperatures from the start spills the most.</li>
+<li><a href="#where-the-evidence-is-heading-e3">Where the evidence is heading (E3)</a>: every condition peaks at slice 3, and none leans on the query answer. What changes is slice 1: a sharper pool over slices puts less of the anchor there, and the runs that spill are the ones with more of it at slice 1.</li>
+<li><a href="#removal-spill-and-task-at-epoch-400-e4">Removal, spill, and task at epoch 400 (E4)</a>: both finished conditions keep the anchor, remove a little more than <code>falloff</code> and the base, and stay within the spill criterion on every seed, with one or two seeds at the line. Neither costs the task. Holding <span class="arithmatex">\(\tau_\ell\)</span> fixed is the tighter of the two on removal and spill.</li>
+<li><a href="#spill-along-training-e5">Spill along training (E5)</a>: no finished run spills early. The spill that remains comes late, after epoch 128, on one or two seeds, and reaches about the criterion by the end. So the screen at epoch 64 called the early kind correctly, and only the finished runs show the late kind.</li>
+<li><a href="#where-the-spill-enters-e6">Where the spill enters (E6)</a>: the evidence still peaks at slice 3 on every seed. The edit at block 1 alone costs the other ops about what it costs them under the step, and less than under the base. With <span class="arithmatex">\(\tau_\ell\)</span> fixed that holds seed by seed; when it sharpens by 40 it holds on the mean only. The other ops sit as far from e₁ at the first two slices as under the step, and the lean of the color embeddings along lightness falls back to the level of the step.</li>
 </ul>
 <h2 id="scope-and-method">Scope and method<a class="anchor-link" href="#scope-and-method" title="Link to this heading">&para;</a></h2>
-<p>This is an exploratory study, with no preregistration and no gate. Every condition starts from <code>pool 1–3, soft τ</code> (the base): the pull is pooled over the slice and position pairs of slices 1 to 3, the position pool temperature <span class="arithmatex">\(\tau_p\)</span> sharpens from 1 to 0.1 and the slice pool temperature <span class="arithmatex">\(\tau_\ell\)</span> from 4 to 0.4 over the first forty epochs, and the anchor weight <span class="arithmatex">\(\lambda_a\)</span> warms up over the same forty epochs to its peak of 0.1. The anti weight is that of <code>falloff</code>, the recipe of record, and the rest of training is as in <a href="/docs/m2/soft-pool/report.py">soft-pool</a>.</p>
+<p>This is an exploratory study, with no preregistration and no gate. Every condition starts from <code>pool 1–3, soft τ</code> (the base, which is not the recipe of record): the pull is pooled over the slice and position pairs of slices 1 to 3, the position pool temperature <span class="arithmatex">\(\tau_p\)</span> sharpens from 1 to 0.1 and the slice pool temperature <span class="arithmatex">\(\tau_\ell\)</span> from 4 to 0.4 over the first forty epochs, and the anchor weight <span class="arithmatex">\(\lambda_a\)</span> warms up over the same forty epochs to its peak of 0.1. The anti weight is that of <code>falloff</code>, the recipe of record, and the rest of training is as in <a href="/docs/m2/soft-pool/report.py">soft-pool</a>.</p>
 <p>Twelve conditions cross two factors. One is <span class="arithmatex">\(\tau_\ell\)</span> at the end, with levels 0.1, 0.2 and 0.4 (the base), each starting from 4. The other is when the temperatures are sharp: by epoch 10, 20 or 40 (the base), both on the same window, or <span class="arithmatex">\(\tau_\ell\)</span> at its end value from epoch 0 while <span class="arithmatex">\(\tau_p\)</span> still sharpens by 40, which tells the two temperatures apart. Five more conditions move the <span class="arithmatex">\(\lambda_a\)</span> warm-up at the base temperatures: a delay (zero until epoch 20, then up by 60) and a stretch (up by 80), each with the soft temperatures and with both temperatures fixed at their end values from the start, and the fixed temperatures with the base warm-up. The fixed-temperature conditions might have lost the anchor, which is what they were there to show: if a delayed pull keeps it at a fixed τ, the soft start was a delay by other means. The schedules are drawn below.</p>
 <figure class="mini-themed-figure">
-<img class="mini-themed-img-light" src="_assets/soft-pool-scout-schedules-light.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="452" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
-<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-schedules-dark.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="452" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-schedules-light.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="548" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-schedules-dark.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="548" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
 <figcaption><p><strong>The schedules the conditions change.</strong> The anchor weight <span class="arithmatex">\(\lambda_a\)</span>, the position pool temperature
 <span class="arithmatex">\(\tau_p\)</span> and the slice pool temperature <span class="arithmatex">\(\tau_\ell\)</span> (both on a log scale) over the first hundred
 epochs, one line per condition, in its ink and marker; where conditions share a schedule, the later line is
-drawn over the earlier. After epoch 100 every condition is the same.</p></figcaption></figure>
+drawn over the earlier, with the markers staggered. The dotted line is the stop at epoch 64. After
+epoch 100 every condition is the same.</p></figcaption></figure>
 
 <p>Each run trains on the full 400-epoch schedule and stops at epoch 64, keeping what it needs to carry on (a resumed run matches a straight run bit for bit). The base is not trained again: <a href="/docs/m2/spill-timing/report.py">spill-timing</a> trained it at these seeds with the same scoring at every record, so its epoch-64 record is the base here. The epoch-64 checkpoints of <code>falloff</code> and <code>step, soft τ</code> left by <a href="/docs/m2/loss-rate/report.py">the loss-rate screen</a> are reference points on the same scale.</p>
 <p>A screen at epoch 64 can rank the conditions by the early kind of <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> and by lost anchors. It cannot clear a condition: the late kind of spill, which grows on <code>value-hsv</code> after epoch 100 on seeds that learn it late, is invisible here. Spill-timing found the base about 0.02 above the quiet recipes at epoch 64, pooled over seeds, so the seed mean of eight runs can tell an early-spilling condition from a quiet one at about that resolution, and a single run says little. We compare many conditions, so a mean a little apart from its neighbors is a lead only. The pattern across a factor, which pools 32 runs at each level, is firmer.</p>
@@ -1095,10 +1100,21 @@ drawn over the earlier. After epoch 100 every condition is the same.</p></figcap
 <ul>
 <li>The call: <dfn class="term" tabindex="0">α</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Alignment</span> <span class="gloss">How far a state points along e₁<span class="lead-colon">:</span></span><span class="more"> the cosine between the state and e₁, which for a unit-length state is its component along e₁; 1 on the anchor direction and 0 off it.</span></span> (the cosine of the state with e₁) at the example answers at slice 3, averaged over epochs 44 to 64. As in the loss-rate screen, a run is called lost below 0.15, kept from 0.28, and uncertain between.</li>
 <li>Spill at epoch 64: the largest drop on another op when e₁ is projected out at every position and slice, less the median drop under eight reference edits (each projecting out one other basis direction), over the whole held-out set, as in spill-timing. A drop is in <dfn class="term" tabindex="0">expected exact match</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Expected exact match</span> <span class="gloss">The task score, as the chance that a drawn answer agrees with a true one.</span><span class="more"> With drawn answers, the correct answer to a line is spread over two or more colors, so we take the chance that an answer drawn from the model agrees with one drawn from the true distribution. Its ceiling is that same chance for the true distribution against itself (Σq²), below 1 on the ops that round.</span></span>. Lower is better. The criterion is 0.02, though at epoch 64 it is a guide only, since it is set for a finished run.</li>
-<li>Where the op sits: α at the example answers and at the query answer, by slice, at epoch 64.</li>
+<li>Where the evidence for the op sits: α at the example answers and at the query answer, by slice, at epoch 64.</li>
 </ul>
+<h3 id="stage-2">Stage 2<a class="anchor-link" href="#stage-2" title="Link to this heading">&para;</a></h3>
+<p>Each run of the two finished conditions (<span class="arithmatex">\(\tau_\ell\)</span> ending at 0.1 and sharpening by epoch 40, and <span class="arithmatex">\(\tau_\ell\)</span> held at 0.1 from the start while <span class="arithmatex">\(\tau_p\)</span> sharpens by 40) carries on from epoch 64 to epoch 400. The checkpoints kept at epochs 16, 32, 64, 128, and 256 along the way are scored for the spill.</p>
+<p>We measure the finished runs as <a href="/docs/m2/recipe-compare/report.py">recipe-compare</a> measured <code>falloff</code>, <code>step, soft τ</code>, and the base at the same seeds, and set them beside its numbers for those three. Each run at a seed shares its initialization, batches, and label draws with the base and with the control of ex-2.2.23 at that seed. So each comparison is paired by seed, and the paired difference from the base says whether a gap is larger than the variation between seeds.</p>
+<ul>
+<li>Removal: the drop in the score of <code>difference</code> under the full edit, as a share of the way from the clean score to the <dfn class="term" tabindex="0">target null</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Target null</span> <span class="gloss">Where a removal should land<span class="lead-colon">:</span></span><span class="more"> what an ideal predictor that had lost one op and nothing else would answer. It is the Bayes posterior over ops given the examples, with the removed op set to zero and the rest renormalized, so it still weighs the other ops by how well they fit.</span></span>. It is net of the median drop under the same edit along each of the other 63 coordinate axes of the run (the rule of the D2.2 design). A run that removes less than 0.5 has lost the anchor.</li>
+<li>Spill: the largest drop on another op at any dose, net of the same median. The criterion is 0.02.</li>
+<li>Task change: the clean score per op, less that of the control at the seed. The task gate fails below -0.02 on the mean over ops, or below -0.05 on any one op.</li>
+<li>Spill along training: the spill at each kept checkpoint and at the end, against eight reference edits, as at the stop.</li>
+<li>Restricted edits, the write, the grading, and the landing, as in recipe-compare: the full edit made at a set of slices only, net of the same edit on the control; how much e₁ the blocks after the edit put back on the other ops, against the bound the clean pass sets; the share removed on <code>difference</code> contexts; and how far the edited answers move toward the target null.</li>
+</ul>
+<p>Eight seeds are too few to add to a count of lost anchors. The loss-rate screen took that count at forty seeds for the step recipes, and nobody has taken it for the pool.</p>
 <h2 id="lost-anchors-e1">Lost anchors (E1)<a class="anchor-link" href="#lost-anchors-e1" title="Link to this heading">&para;</a></h2>
-<p>The fixed-temperature conditions were the ones most likely to lose the anchor, since a sharp pool from the first step is what let the pull settle on the query answer on one seed in slice-terms. Did any schedule lose it by epoch 64?</p>
+<p>The fixed-temperature conditions were the ones most likely to lose the anchor, since a sharp pool from the first step is what let the pull latch onto the query answer on one seed in slice-terms. Did any schedule lose it by epoch 64?</p>
 <p>None did. Every run is called kept, and the lowest call (0.36, on a stretched warm-up) is still above the line by a wide margin. The figure (top) shows the calls close together across all sixteen conditions. The fixed temperatures and the stretch sit a little lower than the rest, which fits a pull that has had less time at full strength by epoch 64, but none comes near the uncertain band.</p>
 <figure class="mini-themed-figure">
 <img class="mini-themed-img-light" src="_assets/soft-pool-scout-screen-light.png" alt="Two panels with one column per condition. Top: the call sits between about 0.5 and 0.6 for every condition, far above the kept line at 0.28; the lowest single run, on the stretched warm-up, is at 0.36. Bottom: the four τ_ℓ 0.1 conditions sit just below zero with every seed under 0.01, like the falloff and step, soft τ columns at the right. The τ_ℓ 0.2 conditions sit near zero, with one seed at 0.13 (τ_ℓ fixed) and two near 0.03. The τ_ℓ 0.4 conditions and the λ_a conditions have means between about 0 and 0.02 and seeds up to 0.05 to 0.18." width="644" height="452" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-screen" />
@@ -1106,15 +1122,17 @@ drawn over the earlier. After epoch 100 every condition is the same.</p></figcap
 <figcaption><p><strong>Each condition at epoch 64.</strong> One column per condition, one small dot per seed, the seed mean as
 the large mark and the seed range as a thin bar. Top: α at the example answers at slice
 3, averaged over epochs 44 to 64 (the loss-rate call);
-below the dotted line a run is called lost, above the dashed line kept. Bottom: the spill at epoch
+below the dotted line a run is called lost (cross-hatched), above the dashed line kept, and between
+the two uncertain (lightly hatched). Bottom: the spill at epoch
 64, the largest drop on another op under the full edit less the median drop under eight reference
-edits; the dashed line is the criterion. The gray columns at the right are the loss-rate screen's
+edits; the dashed line is the criterion, with the side over it hatched. The gray columns at the right are the loss-rate screen's
 epoch-64 checkpoints of the two quiet recipes, at 34 seeds each.</p></figcaption></figure>
 
-<p>So the soft start is not what keeps the anchor in this pool, at least to epoch 64 and at these seeds: the pull pooled over three slices holds the anchor at a fixed temperature too. Eight seeds would miss a loss rate of one in ten or so, and a run can still lose the anchor late, so the fixed temperatures are not cleared. But the soft start was added to the pool to keep the anchor, and at epoch 64 the pool keeps it without one.</p>
+<p>Two soft starts are in play: <span class="arithmatex">\(\tau_p\)</span> sharpening from 1, and <span class="arithmatex">\(\tau_\ell\)</span> sharpening from 4. The conditions with <span class="arithmatex">\(\tau_\ell\)</span> fixed drop the second, the fixed-τ conditions drop both, and all of them kept the anchor on every seed. So in this pool neither soft start is what keeps the anchor, at least to epoch 64 and at these seeds. The step may differ: in the loss-rate screen, the step at a fixed <span class="arithmatex">\(\tau_p\)</span> (<code>falloff</code>) lost the anchor on one seed of forty and the step with the soft start on none, which fits a soft <span class="arithmatex">\(\tau_p\)</span> that helps the step, though one loss is too few to show it. Eight seeds would miss a loss rate of one in ten or so, and a run can still lose the anchor late, so the fixed temperatures are not cleared for this pool either.</p>
 <h2 id="spill-at-epoch-64-e2">Spill at epoch 64 (E2)<a class="anchor-link" href="#spill-at-epoch-64-e2" title="Link to this heading">&para;</a></h2>
 <p>If <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> enters through block 1 while the pool is soft, then sharpening sooner, sharpening the slice pool further, or holding the pull back until e₁ is cleared might each close it. Which of them lowers the spill at epoch 64?</p>
-<p>The slice temperature at the end does, and the others barely move it. The figure (bottom) groups the crossing by <span class="arithmatex">\(\tau_\ell\)</span>: at 0.1 the four timing levels all sit at or below the quiet recipes, with no run of 32 over the criterion; at 0.2 the spill is a little higher, with 3 runs over; at 0.4, the base level, 9 of 32 are over. Within each level, sharpening by epoch 10, 20 or 40, or holding <span class="arithmatex">\(\tau_\ell\)</span> fixed from the start, makes no difference we can see at eight seeds. So for the early spill, what matters is how sharp the slice pool ends, and how soon it gets there barely does.</p>
+<p>The slice temperature at the end does, and the others barely move it. The figure (bottom) groups the crossing by <span class="arithmatex">\(\tau_\ell\)</span>: at 0.1 the four timing levels all sit at or below the quiet recipes, with no run of 32 over the criterion; at 0.2 the early spill is a little higher, with 3 runs over; at 0.4, the base level, 9 of 32 are over. Within each level, sharpening by epoch 10, 20 or 40, or holding <span class="arithmatex">\(\tau_\ell\)</span> fixed from the start, makes no difference we can see at eight seeds. So for the early spill, what matters is how sharp the slice pool ends, and how soon it gets there barely does. That also covers where the slice pool starts: at 0.1, sharpening from 4 and holding at 0.1 from epoch 0 spill about the same, so a start in between would be unlikely to differ.</p>
+<p>This is a stronger effect of <span class="arithmatex">\(\tau_\ell\)</span> than <a href="/docs/m2/slice-terms/report.py">slice-terms</a> saw. Its E4 trained the pool over slices with <span class="arithmatex">\(\tau_\ell\)</span> from 0.1 to 1.6 and found no trend in removal, with the spill a little higher at the two softest levels. But that was one seed, chosen because its pull had settled on block 4, at a fixed <span class="arithmatex">\(\tau_p\)</span> and with the pool over blocks 1 to 4. Here the pool leaves block 4 out, and the spill is a per-seed event: at <span class="arithmatex">\(\tau_\ell\)</span> 0.4 most seeds stay quiet, and one seed would most likely have been one of those.</p>
 <figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>kept ↑</th><th class=num>lowest call</th><th class=num>spill ↓</th><th class=num>seed range</th><th class=num>over</th><th class=num>peak slice</th><th class=num>α at peak</th><th class=num>α, query</th></tr></thead><tbody><tr><td><code>τ_ℓ 0.1, by 20</code></td><td class=num>8/8</td><td class=num>0.50</td><td class=num>−0.007</td><td class=num>−0.013 to 0.001</td><td class=num>0</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.89</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.006</td><td class=num>−0.016 to 0.003</td><td class=num>0</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.1, by 10</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.006</td><td class=num>−0.011 to 0.006</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.2, by 10</code></td><td class=num>8/8</td><td class=num>0.52</td><td class=num>−0.003</td><td class=num>−0.012 to 0.005</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.87</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.003</td><td class=num>−0.017 to 0.008</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.86</td></tr><tr><td><code>τ_ℓ 0.2, by 20</code></td><td class=num>8/8</td><td class=num>0.48</td><td class=num>−0.002</td><td class=num>−0.012 to 0.033</td><td class=num>1</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.2, by 40</code></td><td class=num>8/8</td><td class=num>0.51</td><td class=num>−0.002</td><td class=num>−0.013 to 0.004</td><td class=num>0</td><td class=num>3</td><td class=num>0.54</td><td class=num>0.86</td></tr><tr><td><code>fixed τ, λ_a delay</code></td><td class=num>8/8</td><td class=num>0.52</td><td class=num>0.004</td><td class=num>−0.014 to 0.035</td><td class=num>1</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.85</td></tr><tr><td><code>τ_ℓ 0.4, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.48</td><td class=num>0.006</td><td class=num>−0.010 to 0.033</td><td class=num>2</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.89</td></tr><tr><td><code>τ_ℓ 0.4, by 10</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.009</td><td class=num>−0.009 to 0.034</td><td class=num>2</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.85</td></tr><tr><td><code>λ_a delay</code></td><td class=num>8/8</td><td class=num>0.51</td><td class=num>0.009</td><td class=num>−0.010 to 0.052</td><td class=num>2</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.86</td></tr><tr><td><code>τ_ℓ 0.4, by 20</code></td><td class=num>8/8</td><td class=num>0.47</td><td class=num>0.010</td><td class=num>−0.010 to 0.041</td><td class=num>2</td><td class=num>3</td><td class=num>0.54</td><td class=num>0.85</td></tr><tr><td><code>τ_ℓ 0.2, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.53</td><td class=num>0.013</td><td class=num>−0.013 to 0.126</td><td class=num>2</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>8/8</td><td class=num>0.45</td><td class=num>0.014</td><td class=num>−0.013 to 0.054</td><td class=num>3</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.85</td></tr><tr><td><code>λ_a stretch</code></td><td class=num>8/8</td><td class=num>0.36</td><td class=num>0.019</td><td class=num>−0.007 to 0.082</td><td class=num>2</td><td class=num>3</td><td class=num>0.53</td><td class=num>0.83</td></tr><tr><td><code>fixed τ</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.020</td><td class=num>−0.014 to 0.123</td><td class=num>3</td><td class=num>3</td><td class=num>0.53</td><td class=num>0.82</td></tr><tr><td><code>fixed τ, λ_a stretch</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.021</td><td class=num>−0.011 to 0.177</td><td class=num>1</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.84</td></tr></tbody></table><figcaption><p><strong>The conditions at epoch 64, ranked.</strong> Ranked by whether any run was not called kept, then by the mean
 spill. Kept: the runs the loss-rate call keeps (α at the example answers at slice 3,
 averaged over epochs 44 to 64, at least 0.28). Lowest call: the
@@ -1124,11 +1142,11 @@ the criterion (0.02). Peak slice and α at peak: where the seed-mean α at the e
 highest at epoch 64, and its value. α, query: the seed-mean α at the query answer at slice
 3.</p></figcaption></figure>
 
-<p>The <span class="arithmatex">\(\lambda_a\)</span> conditions all stay well above the <span class="arithmatex">\(\tau_\ell\)</span> 0.1 level. A twenty-epoch delay, which gives e₁ time to clear under the full learning rate, lowered the mean spill somewhat, at the soft temperatures and at the fixed ones. The stretched warm-up raised it, and so did fixing both temperatures with the base warm-up. Each of these means rests on one or two seeds that spill far more than the rest, so at eight seeds the differences among them could be chance. None of them comes near what sharpening the slice pool does.</p>
+<p>The <span class="arithmatex">\(\lambda_a\)</span> conditions all stay well above the <span class="arithmatex">\(\tau_\ell\)</span> 0.1 level. One of them is not on an equal footing: under the stretched warm-up, <span class="arithmatex">\(\lambda_a\)</span> is still at 94% of its peak at the stop, so its spill at epoch 64 is of a pull not yet at full strength. A twenty-epoch delay, which gives e₁ time to clear under the full learning rate, lowered the mean spill somewhat, at the soft temperatures and at the fixed ones. The stretched warm-up raised it, and so did fixing both temperatures with the base warm-up. Each of these means rests on one or two seeds that spill far more than the rest, so at eight seeds the differences among them could be chance. None of them comes near what sharpening the slice pool does.</p>
 <p>Of the 21 runs over the criterion, 15 spill most on <code>darken</code>, the op on which the base spills early in spill-timing.</p>
 <p>These are early numbers. A sharper slice pool could still spill late, on <code>value-hsv</code> as the base does on its slow seeds, and only a finished run can show that. What we take from the screen is that <span class="arithmatex">\(\tau_\ell\)</span> is the lever on the early kind, and that the level at 0.1 brings the pool to the quiet recipes on it.</p>
-<h2 id="where-the-op-is-heading-e3">Where the op is heading (E3)<a class="anchor-link" href="#where-the-op-is-heading-e3" title="Link to this heading">&para;</a></h2>
-<p>A sharper slice pool could lower the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> in a way we would not want: by settling the pull on one slice and one role, as the joint pool did in <a href="/docs/m2/slice-terms/report.py">slice-terms</a>, where one seed settled on the query answer and the op could not be removed. Where does the evidence for the op sit at epoch 64?</p>
+<h2 id="where-the-evidence-is-heading-e3">Where the evidence is heading (E3)<a class="anchor-link" href="#where-the-evidence-is-heading-e3" title="Link to this heading">&para;</a></h2>
+<p>A sharper slice pool could lower the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> in a way we would not want: by settling the pull on one slice and one role, as the joint pool did in <a href="/docs/m2/slice-terms/report.py">slice-terms</a>, where one seed latched onto the query answer and the op could not be removed. Where does the evidence for the op sit at epoch 64?</p>
 <p>In the same place in every condition. The seed-mean <dfn class="term" tabindex="0">α</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Alignment</span> <span class="gloss">How far a state points along e₁<span class="lead-colon">:</span></span><span class="more"> the cosine between the state and e₁, which for a unit-length state is its component along e₁; 1 on the anchor direction and 0 off it.</span></span> at the example answers peaks at slice 3 in all seventeen conditions, and α at the query answer is about the same everywhere (the table above). The figure shows the three levels of <span class="arithmatex">\(\tau_\ell\)</span> at the base timing, beside the condition that spilled most: the shapes are nearly the same, with one difference at slice 1.</p>
 <figure class="mini-themed-figure">
 <img class="mini-themed-img-light" src="_assets/soft-pool-scout-depth-light.png" alt="Four panels of α against slice, from the embedding to slice 4. In every panel the example answers rise from about zero at the embedding to a peak of about 0.55 at slice 3 and fall to about 0.37 at slice 4, and the query answer peaks near 0.85 at slice 3. At slice 1 the example answers sit near 0.22 for τ_ℓ 0.1, against about 0.3 to 0.37 in the other three panels." width="644" height="242" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-depth" />
@@ -1137,6 +1155,15 @@ highest at epoch 64, and its value. α, query: the seed-mean α at the query ans
 at the query answer (dashed), at each slice, for the three slice temperatures at the base timing and for the condition with the highest mean spill.
 The hairlines behind every panel are the seed means of every condition.</p></figcaption></figure>
 
+<p>Beside the quiet recipes (below), the two conditions at <span class="arithmatex">\(\tau_\ell\)</span> 0.1 have about as much of the evidence at slice 1 as the step does, though the step pulls nothing there: some alignment reaches slice 1 from the pull on the slices after it. The base has half as much again. Past slice 1 the shapes differ as recipe-compare found at the end of training: the step holds the evidence level from slice 3 to slice 4, where the pool lets it fall back, at the example answers and at the query answer alike.</p>
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-quiet-light.png" alt="Two panels of seed-mean α against slice. Left, at the example answers: every line starts near zero at the embedding. At slice 1 the two τ_ℓ 0.1 conditions and the two quiet recipes are together near 0.23, and the base is at about 0.37. The pool conditions peak near 0.56 at slice 3 and fall to about 0.38 at slice 4; falloff and step, soft τ level off near 0.5 from slice 3 to slice 4. Right, at the query answer: the same ordering at slice 1 (about 0.27, 0.33 and 0.43), all near 0.85 at slice 3, then the pool falls to about 0.62 at slice 4 while the quiet recipes stay near 0.85." width="626" height="280" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-quiet" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-quiet-dark.png" alt="Two panels of seed-mean α against slice. Left, at the example answers: every line starts near zero at the embedding. At slice 1 the two τ_ℓ 0.1 conditions and the two quiet recipes are together near 0.23, and the base is at about 0.37. The pool conditions peak near 0.56 at slice 3 and fall to about 0.38 at slice 4; falloff and step, soft τ level off near 0.5 from slice 3 to slice 4. Right, at the query answer: the same ordering at slice 1 (about 0.27, 0.33 and 0.43), all near 0.85 at slice 3, then the pool falls to about 0.62 at slice 4 while the quiet recipes stay near 0.85." width="626" height="280" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-quiet" />
+<figcaption><p><strong>The pool beside the quiet recipes at epoch 64.</strong> Seed-mean α by slice at the example answers
+(left) and at the query answer (right): the two stage-2 conditions and the base in their ink and
+marker, and the loss-rate screen's <code>falloff</code> and <code>step, soft τ</code> at the same epoch
+(34 seeds each) in gray. The hairlines are the other conditions.</p></figcaption></figure>
+
 <p>That difference is the one that tracks the spill. At <span class="arithmatex">\(\tau_\ell\)</span> 0.1, α at the example answers at slice 1 averages 0.22, against 0.37 at 0.4. Across all 136 runs, the runs over the criterion are the ones with more of the anchor at slice 1, though at the same α some spill and some do not.</p>
 <figure class="mini-themed-figure">
 <img class="mini-themed-img-light" src="_assets/soft-pool-scout-slice1-light.png" alt="A scatter of 136 runs: α at the example answers at slice 1 on the horizontal axis, from about 0.15 to 0.45, and spill at epoch 64 on the vertical, from about −0.02 to 0.18. The τ_ℓ 0.1 runs (dark blue) sit at the left, between about 0.16 and 0.26, and all below the criterion. The runs over the criterion all sit right of about 0.26, most from the τ_ℓ 0.4 and λ_a conditions." width="491" height="299" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-slice1" />
@@ -1146,12 +1173,91 @@ slice 1 (horizontal), and the spill (vertical), in the ink and marker of its con
 is the criterion.</p></figcaption></figure>
 
 <p>This fits the account recipe-compare gave: the pool spills when it puts part of the pull on block 1, and the edit then takes something from the other ops there. A sharper slice pool takes more of the pull off slice 1 and puts it on slice 3, so block 1 holds less of the anchor and the edit has less to take. It fits without proving it; the screen shows the two moving together, and a restricted edit on the finished runs (block 1 alone) would test it.</p>
+<h2 id="removal-spill-and-task-at-epoch-400-e4">Removal, spill, and task at epoch 400 (E4)<a class="anchor-link" href="#removal-spill-and-task-at-epoch-400-e4" title="Link to this heading">&para;</a></h2>
+<p>Do the two finished conditions keep the anchor, stay within the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> criterion, and leave the task alone at the end of training, seed by seed, against the base and the two quiet recipes?</p>
+<p>They do. No finished run lost the anchor: the lowest removal is 0.83. Both conditions remove a little more than the base on most seeds (the paired column of the table), and a little more than <code>falloff</code> on the mean. Holding <span class="arithmatex">\(\tau_\ell\)</span> fixed gives the tightest removal of the five (the figure, left), with all eight seeds between 0.88 and 0.91.</p>
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-finished-light.png" alt="Three panels with one column per condition. Left: removal sits between about 0.85 and 0.9 for every condition, with the two τ_ℓ 0.1 conditions highest and τ_ℓ 0.1, τ_ℓ fixed the tightest. Middle: spill sits near zero for every condition but the base, whose seeds reach 0.055; the two τ_ℓ 0.1 conditions have two seeds and one seed at the criterion line, none over it. Right: the task change sits near zero, the pool conditions a little higher than falloff and step, soft τ, all above the gate." width="644" height="290" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-finished" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-finished-dark.png" alt="Three panels with one column per condition. Left: removal sits between about 0.85 and 0.9 for every condition, with the two τ_ℓ 0.1 conditions highest and τ_ℓ 0.1, τ_ℓ fixed the tightest. Middle: spill sits near zero for every condition but the base, whose seeds reach 0.055; the two τ_ℓ 0.1 conditions have two seeds and one seed at the criterion line, none over it. Right: the task change sits near zero, the pool conditions a little higher than falloff and step, soft τ, all above the gate." width="644" height="290" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-finished" />
+<figcaption><p><strong>Removal, spill, and task at epoch 400, by condition.</strong> One column per condition, the
+8 seeds as small dots over a thin bar for their range, and the seed mean as the large mark.
+<strong>Left:</strong> removal at the full dose, net of the median over the other axes of the run. <strong>Middle:</strong> spill,
+the largest net drop on another op at any dose; the dashed line is the criterion, with the side over it
+hatched. <strong>Right:</strong> the task change, the mean clean score over ops less that of the control at the seed;
+the dashed line is the task gate on that mean, with the failing side hatched.</p></figcaption></figure>
+
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>removal ↑</th><th class=num>removal less base, paired</th><th class=num>spill ↓</th><th class=num>spill less base, paired</th><th class=num>over criterion ↓</th><th class=num>over, control median ↓</th><th class=num>task change ↑</th><th class=num>worst op change ↑</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>0.87 (0.77 to 0.91)</td><td class=num>—</td><td class=num>−0.002 (−0.009 to 0.014)</td><td class=num>—</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.007 (−0.032 to 0.005)</td><td class=num>−0.066 (<code>hue-hsv</code>, 708)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>0.86 (0.76 to 0.91)</td><td class=num>—</td><td class=num>−0.001 (−0.011 to 0.012)</td><td class=num>—</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.005 (−0.014 to 0.007)</td><td class=num>−0.036 (<code>hue-hsv</code>, 705)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>0.85 (0.81 to 0.91)</td><td class=num>—</td><td class=num>0.015 (−0.005 to 0.055)</td><td class=num>—</td><td class=num>2 of 8</td><td class=num>3 of 8</td><td class=num>0.001 (−0.003 to 0.008)</td><td class=num>−0.015 (<code>hue-hsv</code>, 708)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>0.89 (0.83 to 0.92)</td><td class=num>0.04 (−0.01 to 0.09); 6 up, 2 down</td><td class=num>0.003 (−0.011 to 0.020)</td><td class=num>−0.012 (−0.059 to 0.014); 3 up, 5 down</td><td class=num>0 of 8</td><td class=num>1 of 8</td><td class=num>−0.001 (−0.023 to 0.009)</td><td class=num>−0.037 (<code>darken</code>, 708)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>0.90 (0.88 to 0.91)</td><td class=num>0.05 (−0.01 to 0.08); 7 up, 1 down</td><td class=num>−0.001 (−0.011 to 0.020)</td><td class=num>−0.015 (−0.061 to 0.006); 3 up, 5 down</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.000 (−0.008 to 0.010)</td><td class=num>−0.023 (<code>hue-hsv</code>, 709)</td></tr></tbody></table><figcaption><p><strong>Removal, spill, and task at epoch 400</strong>, seed mean and range over 8 runs. Removal and
+spill are net of the median drop over the other 63 coordinate axes of the run, and "over, control median"
+counts the runs over the criterion with the median taken in the control at the seed instead. The paired
+columns are the condition less the base at the same seed, with the count of seeds on each side of zero. The
+task change is the mean over ops of the clean score less that of the control at the seed; the worst op change
+is the largest loss on one op at any seed. The rows for <code>falloff</code>, <code>step, soft τ</code> and the base are those of
+recipe-compare.</p></figcaption></figure>
+
+<p>On spill (middle), neither condition has a run over the criterion, where the base has 2. The change shows best on the two seeds the base spills on most (the table by seed, below): at 711 both conditions sit below zero, and at 707 both come down to the line.</p>
+<p>That leaves one or two seeds per condition at the criterion, each within a thousandth of it: 705 and 707 under <code>τ_ℓ 0.1, by 40</code>, and 707 under <code>τ_ℓ 0.1, τ_ℓ fixed</code>. Taking the median over the other axes of the control at the seed, in place of those of the run, puts one run of <code>τ_ℓ 0.1, by 40</code> over the line (705), and none of <code>τ_ℓ 0.1, τ_ℓ fixed</code>.</p>
+<p>On the task (right), both pass the gate on the seed mean, and no op at any seed goes past the per-op line, so the task change stays near zero, as recipe-compare found for the pool. One run of <code>τ_ℓ 0.1, by 40</code> (708) is a little under the line for the mean over ops, at the seed where <code>falloff</code> is lowest too.</p>
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>measure</th><th class=num>704</th><th class=num>705</th><th class=num>706</th><th class=num>707</th><th class=num>708</th><th class=num>709</th><th class=num>710</th><th class=num>711</th></tr></thead><tbody><tr><td><code>τ_ℓ 0.4, by 40</code></td><td class=num>removal</td><td class=num>0.83</td><td class=num>0.91</td><td class=num>0.82</td><td class=num>0.87</td><td class=num>0.83</td><td class=num>0.87</td><td class=num>0.84</td><td class=num>0.81</td></tr><tr><td></td><td class=num>spill</td><td class=num>−0.003</td><td class=num>0.020</td><td class=num>−0.005</td><td class=num><strong>0.027</strong></td><td class=num>0.012</td><td class=num>−0.003</td><td class=num>0.016</td><td class=num><strong>0.055</strong></td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>darken</code></td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>removal</td><td class=num>0.92</td><td class=num>0.90</td><td class=num>0.87</td><td class=num>0.89</td><td class=num>0.83</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.90</td></tr><tr><td></td><td class=num>spill</td><td class=num>0.010</td><td class=num>0.020</td><td class=num>−0.004</td><td class=num>0.019</td><td class=num>−0.002</td><td class=num>−0.003</td><td class=num>−0.011</td><td class=num>−0.004</td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>hue-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>darken</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>removal</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.89</td><td class=num>0.90</td><td class=num>0.91</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.88</td></tr><tr><td></td><td class=num>spill</td><td class=num>−0.002</td><td class=num>−0.005</td><td class=num>0.001</td><td class=num>0.020</td><td class=num>−0.006</td><td class=num>0.003</td><td class=num>−0.011</td><td class=num>−0.006</td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>lighten</code></td><td class=num><code>mix</code></td></tr></tbody></table><figcaption><p><strong>Removal and spill by seed</strong>, as in the table above, with the op the spill is largest on. Bold marks a run
+over the criterion.</p></figcaption></figure>
+
+<h2 id="spill-along-training-e5">Spill along training (E5)<a class="anchor-link" href="#spill-along-training-e5" title="Link to this heading">&para;</a></h2>
+<p><a href="/docs/m2/spill-timing/report.py">Spill-timing</a> found two kinds of <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> under the base: an early kind, which rises while the pool sharpens and already shows at epoch 64, and a late kind, which grows after epoch 100 on seeds that learn <code>value-hsv</code> late. Which kind is left on the seeds that come near the criterion at epoch 400?</p>
+<p>The late kind. Under the base, three seeds go over the criterion by epoch 64, and 711 does so already at epoch 32. Under the two finished conditions no seed goes over it by then, and up to epoch 128 the largest spill is 0.015 under <code>τ_ℓ 0.1, by 40</code> and 0.003 under <code>τ_ℓ 0.1, τ_ℓ fixed</code>. After that, two seeds under <code>τ_ℓ 0.1, by 40</code> (705 and 707) and one under <code>τ_ℓ 0.1, τ_ℓ fixed</code> (707) climb to about the criterion by the end.</p>
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-along-light.png" alt="Three panels of spill against epoch, from 16 to 400 on a log scale, one line per seed. Base: three seeds rise over the criterion by epoch 64 (711 already at 32), and 707 and 711 stay over it to the end; two more seeds climb to near the criterion after epoch 128. τ_ℓ 0.1, by 40, and τ_ℓ 0.1, τ_ℓ fixed: every seed stays under 0.016 up to epoch 128; after that two seeds (705 and 707) and one seed (707) rise to about the criterion by epoch 400." width="644" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-along" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-along-dark.png" alt="Three panels of spill against epoch, from 16 to 400 on a log scale, one line per seed. Base: three seeds rise over the criterion by epoch 64 (711 already at 32), and 707 and 711 stay over it to the end; two more seeds climb to near the criterion after epoch 128. τ_ℓ 0.1, by 40, and τ_ℓ 0.1, τ_ℓ fixed: every seed stays under 0.016 up to epoch 128; after that two seeds (705 and 707) and one seed (707) rise to about the criterion by epoch 400." width="644" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-along" />
+<figcaption><p><strong>Spill along training.</strong> The spill of each run (one line per seed) at the checkpoints kept at epochs
+16, 32, 64, 128, 256 and at the end, scored as at the stop, against the
+median of eight reference edits rather than the 63 axes of E4, so a run at the line there can fall on
+either side of it here. The base is spill-timing's rerun of soft-pool, scored at every record.
+The dashed line is the criterion and the dotted line the stop of stage 1. The seeds named are the ones
+that go over the criterion or come within a quarter of it.</p></figcaption></figure>
+
+<p>So the screen at epoch 64 called the early kind correctly: the conditions it found quiet stayed quiet through the part of training it covers.</p>
+<p>The late kind here is about as large as on the late seeds of the base (705 and 710 come to about 0.02 there too). It lands on different ops from those spill-timing saw, though: on <code>hue-hsv</code> and <code>darken</code> at 705 and 707 under <code>τ_ℓ 0.1, by 40</code>, and on <code>mix</code> at 707 under <code>τ_ℓ 0.1, τ_ℓ fixed</code>, where the late kind in the base was on <code>value-hsv</code>. Two or three runs are too few to say whether that is a different kind of late spill or the same one landing elsewhere.</p>
+<h2 id="where-the-spill-enters-e6">Where the spill enters (E6)<a class="anchor-link" href="#where-the-spill-enters-e6" title="Link to this heading">&para;</a></h2>
+<p>Recipe-compare traced the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> of the base to the edit at block 1, and E3 found that a sharper slice pool puts less of the anchor at slice 1. Does the edit at block 1 alone now cost the other ops less, and does the rest of what the edit does stay as it was?</p>
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>peak α, seeds</th><th class=num>removal, block 1 alone ↑</th><th class=num>removal, blocks 2 to 4 ↑</th><th class=num>spill, embedding alone ↓</th><th class=num>spill, block 1 alone ↓</th><th class=num>spill, blocks 1 to 4 ↓</th><th class=num>spill, every slice ↓</th><th class=num>lightness on e₁, embedding table</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>slice 4: 8</td><td class=num>0.71 (0.38 to 0.95)</td><td class=num>0.81 (0.11 to 0.94)</td><td class=num>−0.000 (−0.003 to 0.002); 0 over</td><td class=num>0.005 (−0.004 to 0.022); 1 over</td><td class=num>0.004 (−0.004 to 0.018); 0 over</td><td class=num>0.001 (−0.013 to 0.018); 0 over</td><td class=num>−0.41 (−0.64 to −0.20)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>slice 4: 8</td><td class=num>0.69 (0.22 to 0.95)</td><td class=num>0.78 (0.06 to 0.91)</td><td class=num>0.000 (−0.003 to 0.004); 0 over</td><td class=num>0.016 (−0.003 to 0.031); 3 over</td><td class=num>0.011 (−0.017 to 0.038); 2 over</td><td class=num>0.006 (−0.020 to 0.029); 2 over</td><td class=num>−0.38 (−0.63 to −0.04)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>slice 3: 8</td><td class=num>0.77 (0.71 to 0.93)</td><td class=num>0.80 (0.04 to 0.93)</td><td class=num>0.003 (−0.002 to 0.010); 0 over</td><td class=num>0.027 (−0.008 to 0.090); 4 over</td><td class=num>0.021 (−0.007 to 0.079); 4 over</td><td class=num>0.021 (−0.006 to 0.071); 4 over</td><td class=num>−0.61 (−0.65 to −0.57)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>slice 3: 8</td><td class=num>0.68 (0.25 to 0.96)</td><td class=num>0.82 (0.09 to 0.94)</td><td class=num>0.001 (−0.002 to 0.004); 0 over</td><td class=num>0.019 (−0.003 to 0.043); 3 over</td><td class=num>0.014 (−0.007 to 0.038); 2 over</td><td class=num>0.010 (−0.005 to 0.030); 3 over</td><td class=num>−0.36 (−0.53 to −0.05)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>slice 3: 8</td><td class=num>0.75 (0.52 to 0.95)</td><td class=num>0.82 (0.17 to 0.96)</td><td class=num>−0.001 (−0.003 to 0.002); 0 over</td><td class=num>0.011 (−0.020 to 0.044); 2 over</td><td class=num>0.005 (−0.017 to 0.033); 1 over</td><td class=num>0.002 (−0.027 to 0.014); 0 over</td><td class=num>−0.41 (−0.59 to −0.15)</td></tr><tr><td>control</td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num>−0.02 (−0.33 to 0.62)</td></tr></tbody></table><figcaption><p><strong>Where the evidence sits, and where the edit finds it.</strong> The slice at which α at the example answers of
+<code>difference</code> contexts peaks, with the count of seeds. Removal and spill under the full edit restricted to a
+set of slices, at every position, net of the same edit on the control at the seed (the median baseline exists
+only for the edit at every slice), with the count of runs over the criterion; seed mean and range. The last
+column is the correlation, over the color embeddings, of the e₁ component with lightness (the lean of
+embedding-lean). The first three rows are those of recipe-compare.</p></figcaption></figure>
+
+<p>The evidence settles where it did. Under both conditions <dfn class="term" tabindex="0">α</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Alignment</span> <span class="gloss">How far a state points along e₁<span class="lead-colon">:</span></span><span class="more"> the cosine between the state and e₁, which for a unit-length state is its component along e₁; 1 on the anchor direction and 0 off it.</span></span> at the example answers peaks at slice 3 on every seed, as under the base. The edit at blocks 2 to 4 removes most of the op on every seed but 710, which falls short under every recipe. The edit at block 4 alone removes nothing. The edit at the embedding alone spills nothing, and neither does the edit at blocks 2 to 4 (this one and the edit at block 4 are not in the table).</p>
+<p>The edit at block 1 alone costs the other ops less than under the base on the mean, and about what it costs them under <code>step, soft τ</code>: the mean is between the step and the base under <code>τ_ℓ 0.1, by 40</code> and below the step under <code>τ_ℓ 0.1, τ_ℓ fixed</code>.</p>
+<p>Seed by seed the two conditions differ. <code>τ_ℓ 0.1, τ_ℓ fixed</code> is below the base on six seeds of eight, where <code>τ_ℓ 0.1, by 40</code> is below it on three, and the lower mean under <code>τ_ℓ 0.1, by 40</code> comes from 711, the seed the base spills on most at block 1. So holding <span class="arithmatex">\(\tau_\ell\)</span> fixed closes block 1 about as far as leaving it out of the pull does, and sharpening to 0.1 by epoch 40 does so on the mean, carried by one seed.</p>
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-block1-light.png" alt="Two panels, one point per run at each seed from 704 to 711, for the base and the two finished conditions. Left, the edit at block 1 alone: the base is highest at 707, 708 and 711 (0.04 to 0.09); τ_ℓ 0.1, by 40 sits near the base or above it on most seeds, between about 0.01 and 0.04, and well below it at 711; τ_ℓ 0.1, τ_ℓ fixed sits below the base on most seeds, with one high point at 709. Right, the edit at every slice: the base reaches 0.07 at 711; by 40 is over the criterion at 704, 705 and 707; τ_ℓ fixed stays under it everywhere." width="626" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-block1" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-block1-dark.png" alt="Two panels, one point per run at each seed from 704 to 711, for the base and the two finished conditions. Left, the edit at block 1 alone: the base is highest at 707, 708 and 711 (0.04 to 0.09); τ_ℓ 0.1, by 40 sits near the base or above it on most seeds, between about 0.01 and 0.04, and well below it at 711; τ_ℓ 0.1, τ_ℓ fixed sits below the base on most seeds, with one high point at 709. Right, the edit at every slice: the base reaches 0.07 at 711; by 40 is over the criterion at 704, 705 and 707; τ_ℓ fixed stays under it everywhere." width="626" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-block1" />
+<figcaption><p><strong>Spill under the edit at block 1 alone, and at every slice, seed by seed.</strong> The largest drop on another
+op under the full edit made at those slices only, net of the same edit on the control at the seed, for
+the base (hollow) and the two finished conditions, offset a little at each seed. The dashed line is the
+criterion, with the side over it hatched.</p></figcaption></figure>
+
+<p>As recipe-compare noted, the block-1 edit on its own overstates what the full edit costs: under <code>τ_ℓ 0.1, τ_ℓ fixed</code> it goes over the criterion on two seeds whose full edit stays within it.</p>
+<!-- REVIEW: added the paired-by-seed reading of the block-1 edit. On the mean both conditions sit below the base, but `τ_ℓ 0.1, by 40` is above the base on five seeds of eight (SLICE_SPILL[n]["block 1"] against the base's, seed by seed), and without 711 its mean is above the base's, so the mean alone overstated it. The same holds for the control-netted edit at every slice, below (above the base on five seeds). Verify: the per-seed block-1 and every-slice spill of the base and the two leaders. -->
+
+<p>With the edit at every slice, net of the control rather than of the median, <code>τ_ℓ 0.1, by 40</code> has three runs over the criterion and <code>τ_ℓ 0.1, τ_ℓ fixed</code> none. Under this baseline <code>step, soft τ</code> has two over, so these counts are best compared with one another, and on them holding <span class="arithmatex">\(\tau_\ell\)</span> fixed is the quieter of the two.</p>
+<p>The embedding table moves with the slice pool as well. Under the base the e₁ components of the color embeddings follow lightness closely on every seed (the lean of <a href="/docs/m2/embedding-lean/report.py">embedding-lean</a>). Under both finished conditions the lean is back at about the level of the step, and varies from seed to seed as it does there, which fits a pull that reaches slice 1 less.</p>
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>largest write over bound ↓</th><th class=num>sites over 1.25×, all runs ↓</th><th class=num>bound at example answers, emb and slice 1 ↓</th><th class=num>share removed ↑</th><th class=num>landing ↑</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>1.08 (slice 1, 0.3°)</td><td class=num>28 (≤ 4°)</td><td class=num>5°, 32°</td><td class=num>0.85 (0.79 to 0.90)</td><td class=num>0.52 (0.49 to 0.56)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>1.15 (slice 1, 3.6°)</td><td class=num>23 (≤ 6°)</td><td class=num>6°, 32°</td><td class=num>0.83 (0.77 to 0.90)</td><td class=num>0.49 (0.47 to 0.53)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>1.33 (slice 1, 4.9°)</td><td class=num>48 (≤ 5°)</td><td class=num>14°, 43°</td><td class=num>0.83 (0.79 to 0.88)</td><td class=num>0.51 (0.47 to 0.55)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>1.03 (slice 1, 4.9°)</td><td class=num>13 (≤ 5°)</td><td class=num>4°, 27°</td><td class=num>0.86 (0.79 to 0.91)</td><td class=num>0.50 (0.45 to 0.55)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>1.06 (slice 1, 4.5°)</td><td class=num>22 (≤ 4°)</td><td class=num>5°, 29°</td><td class=num>0.87 (0.78 to 0.91)</td><td class=num>0.52 (0.47 to 0.54)</td></tr></tbody></table><figcaption><p><strong>The write, the grading and the landing.</strong> The largest ratio of write to bound in angle over the slices past
+the embedding and every position, from the seed-mean quantiles as in edit-effects, with where it falls and
+the bound there. The sites over 1.25× count slice and position pairs over all 8 runs, each
+run taken on its own, with the largest bound among them. The bound at the example answers is over those that
+do not fit <code>difference</code>. The share removed is over every <code>difference</code> context, net of the control at the
+seed, and the landing is the share of the way to the target null in total variation; seed mean and range.
+The first three rows are those of recipe-compare.</p></figcaption></figure>
+
+<p>What the edit does downstream barely changes, and where it changes it moves toward the step. The write stays close to the bound past the embedding, with fewer sites over the tolerance than under the base, and the share removed and the landing are about the same as under every recipe.</p>
+<p>Recipe-compare found the bound at the example answers of the other ops about 3× as wide at the embedding under the base as under <code>falloff</code>. Under both finished conditions it is back at the width of <code>falloff</code> at the embedding, and a little narrower at slice 1. So the other ops sit as far from e₁ at the first two slices as under the step.</p>
 <h2 id="discussion">Discussion<a class="anchor-link" href="#discussion" title="Link to this heading">&para;</a></h2>
-<p>TODO after stage 2.</p>
-<div class="admonition">
-<p class="admonition-title">Open decision</p>
-<p>Which conditions go to stage 2. The plan finishes two or three leaders, ranked by no lost anchors and the lowest mean <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span>, preferring the fewest changes from the base when two are close. The four conditions at <span class="arithmatex">\(\tau_\ell\)</span> 0.1 are close together. A proposal: <code>τ_ℓ 0.1, by 40</code> (one change from the base) and <code>τ_ℓ 0.1, τ_ℓ fixed</code> (the simplest schedule, with only <span class="arithmatex">\(\tau_p\)</span> moving), and <code>τ_ℓ 0.2, by 40</code> as a middle step that would show whether 0.1 is needed or 0.2 is enough once the late kind is counted. To check: whether the finished runs keep their removal, since a sharper slice pool concentrates the pull.</p>
-</div>
+<p>With <span class="arithmatex">\(\tau_\ell\)</span> at 0.1, the pool over slices 1 to 3 keeps the anchor, removes a little more than the recipe of record, costs the task nothing, and stays within the <dfn class="term" tabindex="0">spill</dfn><span class="sidenote glossnote" role="note" tabindex="-1"><span class="glossnote-term">Spill</span> <span class="gloss">The side effect of an edit on the other ops<span class="lead-colon">:</span></span><span class="more"> the largest drop in the score of any op other than the one removed, net of what the same edit does to a control trained without the anchor. The selectivity criterion of D2.2 allows at most 0.02 at any dose.</span></span> criterion on every seed. The early spill that put two runs of the base over the criterion is gone.</p>
+<p>So are most of the signs that block 1 is open to the edit. The wider bound at the first two slices and the lean in the embedding table are gone under both conditions, and the larger cost of the block-1 edit is gone under the condition with <span class="arithmatex">\(\tau_\ell\)</span> fixed.</p>
+<p>Of the two, holding <span class="arithmatex">\(\tau_\ell\)</span> at 0.1 from the start is the cleaner. It removes the most and the most evenly, has no run over the criterion under either median, and none under the control-netted edit at every slice. It also has one schedule fewer than the base, since only <span class="arithmatex">\(\tau_p\)</span> sharpens.</p>
+<p>The other condition is close behind on the median baseline of the D2.2 design, where the gap rests on two or three runs near the line. On the edits netted against the control the gap is larger: there the other condition sits at about the level of the base on most seeds, and only its mean is lower.</p>
+<p>What is left is the late kind of spill, which brings one or two seeds per condition to about the criterion by the end. It is about as large as under the base on its late seeds, so it does not seem to come from the sharper slice pool, and nothing in this study acts on it. When it appears, the term still changing is the anti weight, partway through its anneal, so if anything in the recipe acts on the late kind, that anneal is the likelier place.</p>
+<p>Before the pool with <span class="arithmatex">\(\tau_\ell\)</span> fixed at 0.1 could replace <code>falloff</code> as the recipe of record, it would need two things the step has: a count of lost anchors at forty seeds, which a loss-rate screen would give, and a set of fresh seeds, which a confirmation at seeds 712 to 719 with their controls would give. Since the pool chooses the slice within a range set by the architecture, it would also be the easier recipe to carry to another model, as recipe-compare argued.</p>
+<!-- REVIEW: was "replace `step, soft τ`"; the report's own Scope and method and the D2.2 design doc name `falloff` as the recipe of record, and the two things named (forty seeds, fresh seeds) are what `falloff` has. Verify: docs/m2/d2.2/design.md, the recipe block. -->
 </main>
 </body>
 </html>

exports/m2/soft-pool-scout/index.md

@@ -2,32 +2,38 @@
 
 /// tip |
 <!-- lede -->
-We screened sixteen schedules of the pooled pull at epoch 64, looking for one that stops the early spill of `pool 1–3, soft τ`. A sharper pool over slices did it: with the slice temperature ending at 0.1, every run spilled about as little as the quiet recipes, whenever the temperatures sharpened. Moving the anchor warm-up did not help, and fixing the temperatures from the start spilled more.
+A sharper pool over slices stopped the early spill of `pool 1–3, soft τ`. Of sixteen schedules of the pooled pull screened at epoch 64, those with the slice temperature ending at 0.1 had no run spill early, however soon or late the temperatures sharpened.
+
+We carried two of these conditions on to epoch 400. Both stay within the spill criterion on every seed, and both remove a little more than the recipe of record at no cost to the task. Holding the slice temperature at 0.1 from the start is the cleaner of the two, and a candidate for the recipe of record.
 ///
 
 [Recipe-compare](/docs/m2/recipe-compare/report.py) left one thing between `pool 1–3, soft τ` and the recipe of record: two of its eight runs go over the spill criterion, and the spill enters at block 1, the one slice the pool pulls and the step leaves alone. [Spill-timing](/docs/m2/spill-timing/report.py) found that both of those runs spill early, while the pool sharpens, and that this early kind already shows at epoch 64. So a schedule that keeps block 1 closed should show it at epoch 64, at a sixth of the cost of a full run.
 
-This report trains sixteen conditions of the pool at its eight seeds (704 to 711), each changing when the two pool temperatures sharpen, how sharp the pool over slices ends, or when the anchor weight warms up. Every run stops at epoch 64, where we score whether it is keeping the anchor and how much the edit spills. This is stage 1; the leaders will be carried on to epoch 400 from where they stopped, in a second stage still to run.
+This report trains sixteen conditions of the pool at its eight seeds (704 to 711), each changing when the two pool temperatures sharpen, how sharp the pool over slices ends, or when the anchor weight warms up. Every run stops at epoch 64, where we score whether it is keeping the anchor and how much the edit spills. This is stage 1. In stage 2, the two conditions with $\tau_\ell$ ending at 0.1 that change least from the base (sharpening by epoch 40 as the base does, and held at 0.1 from the start) are carried on to epoch 400 from where they stopped, and measured as recipe-compare measured the three recipes, with the spill along training from checkpoints kept on the way.
 
 ## Observations
 
-- [Lost anchors (E1)](#lost-anchors-e1): no schedule lost the anchor. All 136 runs, the base included, are called kept, the lowest well above the line, though the fixed temperatures and the stretched warm-up keep it a little less firmly.
-- [Spill at epoch 64 (E2)](#spill-at-epoch-64-e2): the slice temperature at the end decides the spill, and when the temperatures sharpen hardly matters. At τ_ℓ 0.1 no run is over the criterion, and the four conditions sit at or below the quiet recipes. At 0.4 (the base) about a quarter of the runs are over. The delayed and stretched warm-ups spill as much as the base, and fixing the temperatures from the start spills the most.
-- [Where the op is heading (E3)](#where-the-op-is-heading-e3): every condition peaks at slice 3, and none leans on the query answer. What changes is slice 1: a sharper pool over slices puts less of the anchor there, and the runs that spill are the ones with more of it at slice 1.
+- [Lost anchors (E1)](#lost-anchors-e1): no schedule lost the anchor. All 136 runs, the base included, are called kept, the lowest well above the line. The fixed temperatures and the stretched warm-up keep it a little less firmly.
+- [Spill at epoch 64 (E2)](#spill-at-epoch-64-e2): the final slice temperature decides the spill; when the temperatures sharpen hardly matters. At τ_ℓ 0.1 no run is over the criterion, and the four conditions sit at or below the quiet recipes. At 0.4 (the base) about a quarter of the runs are over. Moving the $\lambda_a$ warm-up shifts the spill by less than the seeds vary, and fixing the temperatures from the start spills the most.
+- [Where the evidence is heading (E3)](#where-the-evidence-is-heading-e3): every condition peaks at slice 3, and none leans on the query answer. What changes is slice 1: a sharper pool over slices puts less of the anchor there, and the runs that spill are the ones with more of it at slice 1.
+- [Removal, spill, and task at epoch 400 (E4)](#removal-spill-and-task-at-epoch-400-e4): both finished conditions keep the anchor, remove a little more than `falloff` and the base, and stay within the spill criterion on every seed, with one or two seeds at the line. Neither costs the task. Holding $\tau_\ell$ fixed is the tighter of the two on removal and spill.
+- [Spill along training (E5)](#spill-along-training-e5): no finished run spills early. The spill that remains comes late, after epoch 128, on one or two seeds, and reaches about the criterion by the end. So the screen at epoch 64 called the early kind correctly, and only the finished runs show the late kind.
+- [Where the spill enters (E6)](#where-the-spill-enters-e6): the evidence still peaks at slice 3 on every seed. The edit at block 1 alone costs the other ops about what it costs them under the step, and less than under the base. With $\tau_\ell$ fixed that holds seed by seed; when it sharpens by 40 it holds on the mean only. The other ops sit as far from e₁ at the first two slices as under the step, and the lean of the color embeddings along lightness falls back to the level of the step.
 
 ## Scope and method
 
-This is an exploratory study, with no preregistration and no gate. Every condition starts from `pool 1–3, soft τ` (the base): the pull is pooled over the slice and position pairs of slices 1 to 3, the position pool temperature $\tau_p$ sharpens from 1 to 0.1 and the slice pool temperature $\tau_\ell$ from 4 to 0.4 over the first forty epochs, and the anchor weight $\lambda_a$ warms up over the same forty epochs to its peak of 0.1. The anti weight is that of `falloff`, the recipe of record, and the rest of training is as in [soft-pool](/docs/m2/soft-pool/report.py).
+This is an exploratory study, with no preregistration and no gate. Every condition starts from `pool 1–3, soft τ` (the base, which is not the recipe of record): the pull is pooled over the slice and position pairs of slices 1 to 3, the position pool temperature $\tau_p$ sharpens from 1 to 0.1 and the slice pool temperature $\tau_\ell$ from 4 to 0.4 over the first forty epochs, and the anchor weight $\lambda_a$ warms up over the same forty epochs to its peak of 0.1. The anti weight is that of `falloff`, the recipe of record, and the rest of training is as in [soft-pool](/docs/m2/soft-pool/report.py).
 
 Twelve conditions cross two factors. One is $\tau_\ell$ at the end, with levels 0.1, 0.2 and 0.4 (the base), each starting from 4. The other is when the temperatures are sharp: by epoch 10, 20 or 40 (the base), both on the same window, or $\tau_\ell$ at its end value from epoch 0 while $\tau_p$ still sharpens by 40, which tells the two temperatures apart. Five more conditions move the $\lambda_a$ warm-up at the base temperatures: a delay (zero until epoch 20, then up by 60) and a stretch (up by 80), each with the soft temperatures and with both temperatures fixed at their end values from the start, and the fixed temperatures with the base warm-up. The fixed-temperature conditions might have lost the anchor, which is what they were there to show: if a delayed pull keeps it at a fixed τ, the soft start was a delay by other means. The schedules are drawn below.
 
 <figure class="mini-themed-figure">
-<img class="mini-themed-img-light" src="_assets/soft-pool-scout-schedules-light.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="452" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
-<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-schedules-dark.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="452" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-schedules-light.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="548" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-schedules-dark.png" alt="Three panels over epochs 0 to 100. Top: λ_a rises from zero to 0.1 by epoch 40 for most conditions, by 60 from a start at 20 for the delay, and by 80 for the stretch. Middle: τ_p falls from 1 to 0.1 by epoch 10, 20 or 40, or sits at 0.1 throughout for fixed τ. Bottom: τ_ℓ falls from 4 to 0.1, 0.2 or 0.4 on the same windows, or sits at its end value throughout." width="626" height="548" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-schedules" />
 <figcaption><p><strong>The schedules the conditions change.</strong> The anchor weight <span class="arithmatex">\(\lambda_a\)</span>, the position pool temperature
 <span class="arithmatex">\(\tau_p\)</span> and the slice pool temperature <span class="arithmatex">\(\tau_\ell\)</span> (both on a log scale) over the first hundred
 epochs, one line per condition, in its ink and marker; where conditions share a schedule, the later line is
-drawn over the earlier. After epoch 100 every condition is the same.</p></figcaption></figure>
+drawn over the earlier, with the markers staggered. The dotted line is the stop at epoch 64. After
+epoch 100 every condition is the same.</p></figcaption></figure>
 
 Each run trains on the full 400-epoch schedule and stops at epoch 64, keeping what it needs to carry on (a resumed run matches a straight run bit for bit). The base is not trained again: [spill-timing](/docs/m2/spill-timing/report.py) trained it at these seeds with the same scoring at every record, so its epoch-64 record is the base here. The epoch-64 checkpoints of `falloff` and `step, soft τ` left by [the loss-rate screen](/docs/m2/loss-rate/report.py) are reference points on the same scale.
 
@@ -37,11 +43,25 @@ A screen at epoch 64 can rank the conditions by the early kind of spill and by l
 
 - The call: α (the cosine of the state with e₁) at the example answers at slice 3, averaged over epochs 44 to 64. As in the loss-rate screen, a run is called lost below 0.15, kept from 0.28, and uncertain between.
 - Spill at epoch 64: the largest drop on another op when e₁ is projected out at every position and slice, less the median drop under eight reference edits (each projecting out one other basis direction), over the whole held-out set, as in spill-timing. A drop is in expected exact match. Lower is better. The criterion is 0.02, though at epoch 64 it is a guide only, since it is set for a finished run.
-- Where the op sits: α at the example answers and at the query answer, by slice, at epoch 64.
+- Where the evidence for the op sits: α at the example answers and at the query answer, by slice, at epoch 64.
+
+### Stage 2
+
+Each run of the two finished conditions ($\tau_\ell$ ending at 0.1 and sharpening by epoch 40, and $\tau_\ell$ held at 0.1 from the start while $\tau_p$ sharpens by 40) carries on from epoch 64 to epoch 400. The checkpoints kept at epochs 16, 32, 64, 128, and 256 along the way are scored for the spill.
+
+We measure the finished runs as [recipe-compare](/docs/m2/recipe-compare/report.py) measured `falloff`, `step, soft τ`, and the base at the same seeds, and set them beside its numbers for those three. Each run at a seed shares its initialization, batches, and label draws with the base and with the control of ex-2.2.23 at that seed. So each comparison is paired by seed, and the paired difference from the base says whether a gap is larger than the variation between seeds.
+
+- Removal: the drop in the score of `difference` under the full edit, as a share of the way from the clean score to the target null. It is net of the median drop under the same edit along each of the other 63 coordinate axes of the run (the rule of the D2.2 design). A run that removes less than 0.5 has lost the anchor.
+- Spill: the largest drop on another op at any dose, net of the same median. The criterion is 0.02.
+- Task change: the clean score per op, less that of the control at the seed. The task gate fails below -0.02 on the mean over ops, or below -0.05 on any one op.
+- Spill along training: the spill at each kept checkpoint and at the end, against eight reference edits, as at the stop.
+- Restricted edits, the write, the grading, and the landing, as in recipe-compare: the full edit made at a set of slices only, net of the same edit on the control; how much e₁ the blocks after the edit put back on the other ops, against the bound the clean pass sets; the share removed on `difference` contexts; and how far the edited answers move toward the target null.
+
+Eight seeds are too few to add to a count of lost anchors. The loss-rate screen took that count at forty seeds for the step recipes, and nobody has taken it for the pool.
 
 ## Lost anchors (E1)
 
-The fixed-temperature conditions were the ones most likely to lose the anchor, since a sharp pool from the first step is what let the pull settle on the query answer on one seed in slice-terms. Did any schedule lose it by epoch 64?
+The fixed-temperature conditions were the ones most likely to lose the anchor, since a sharp pool from the first step is what let the pull latch onto the query answer on one seed in slice-terms. Did any schedule lose it by epoch 64?
 
 None did. Every run is called kept, and the lowest call (0.36, on a stretched warm-up) is still above the line by a wide margin. The figure (top) shows the calls close together across all sixteen conditions. The fixed temperatures and the stretch sit a little lower than the rest, which fits a pull that has had less time at full strength by epoch 64, but none comes near the uncertain band.
 
@@ -51,18 +71,21 @@ None did. Every run is called kept, and the lowest call (0.36, on a stretched wa
 <figcaption><p><strong>Each condition at epoch 64.</strong> One column per condition, one small dot per seed, the seed mean as
 the large mark and the seed range as a thin bar. Top: α at the example answers at slice
 3, averaged over epochs 44 to 64 (the loss-rate call);
-below the dotted line a run is called lost, above the dashed line kept. Bottom: the spill at epoch
+below the dotted line a run is called lost (cross-hatched), above the dashed line kept, and between
+the two uncertain (lightly hatched). Bottom: the spill at epoch
 64, the largest drop on another op under the full edit less the median drop under eight reference
-edits; the dashed line is the criterion. The gray columns at the right are the loss-rate screen's
+edits; the dashed line is the criterion, with the side over it hatched. The gray columns at the right are the loss-rate screen's
 epoch-64 checkpoints of the two quiet recipes, at 34 seeds each.</p></figcaption></figure>
 
-So the soft start is not what keeps the anchor in this pool, at least to epoch 64 and at these seeds: the pull pooled over three slices holds the anchor at a fixed temperature too. Eight seeds would miss a loss rate of one in ten or so, and a run can still lose the anchor late, so the fixed temperatures are not cleared. But the soft start was added to the pool to keep the anchor, and at epoch 64 the pool keeps it without one.
+Two soft starts are in play: $\tau_p$ sharpening from 1, and $\tau_\ell$ sharpening from 4. The conditions with $\tau_\ell$ fixed drop the second, the fixed-τ conditions drop both, and all of them kept the anchor on every seed. So in this pool neither soft start is what keeps the anchor, at least to epoch 64 and at these seeds. The step may differ: in the loss-rate screen, the step at a fixed $\tau_p$ (`falloff`) lost the anchor on one seed of forty and the step with the soft start on none, which fits a soft $\tau_p$ that helps the step, though one loss is too few to show it. Eight seeds would miss a loss rate of one in ten or so, and a run can still lose the anchor late, so the fixed temperatures are not cleared for this pool either.
 
 ## Spill at epoch 64 (E2)
 
 If spill enters through block 1 while the pool is soft, then sharpening sooner, sharpening the slice pool further, or holding the pull back until e₁ is cleared might each close it. Which of them lowers the spill at epoch 64?
 
-The slice temperature at the end does, and the others barely move it. The figure (bottom) groups the crossing by $\tau_\ell$: at 0.1 the four timing levels all sit at or below the quiet recipes, with no run of 32 over the criterion; at 0.2 the spill is a little higher, with 3 runs over; at 0.4, the base level, 9 of 32 are over. Within each level, sharpening by epoch 10, 20 or 40, or holding $\tau_\ell$ fixed from the start, makes no difference we can see at eight seeds. So for the early spill, what matters is how sharp the slice pool ends, and how soon it gets there barely does.
+The slice temperature at the end does, and the others barely move it. The figure (bottom) groups the crossing by $\tau_\ell$: at 0.1 the four timing levels all sit at or below the quiet recipes, with no run of 32 over the criterion; at 0.2 the early spill is a little higher, with 3 runs over; at 0.4, the base level, 9 of 32 are over. Within each level, sharpening by epoch 10, 20 or 40, or holding $\tau_\ell$ fixed from the start, makes no difference we can see at eight seeds. So for the early spill, what matters is how sharp the slice pool ends, and how soon it gets there barely does. That also covers where the slice pool starts: at 0.1, sharpening from 4 and holding at 0.1 from epoch 0 spill about the same, so a start in between would be unlikely to differ.
+
+This is a stronger effect of $\tau_\ell$ than [slice-terms](/docs/m2/slice-terms/report.py) saw. Its E4 trained the pool over slices with $\tau_\ell$ from 0.1 to 1.6 and found no trend in removal, with the spill a little higher at the two softest levels. But that was one seed, chosen because its pull had settled on block 4, at a fixed $\tau_p$ and with the pool over blocks 1 to 4. Here the pool leaves block 4 out, and the spill is a per-seed event: at $\tau_\ell$ 0.4 most seeds stay quiet, and one seed would most likely have been one of those.
 
 <figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>kept ↑</th><th class=num>lowest call</th><th class=num>spill ↓</th><th class=num>seed range</th><th class=num>over</th><th class=num>peak slice</th><th class=num>α at peak</th><th class=num>α, query</th></tr></thead><tbody><tr><td><code>τ_ℓ 0.1, by 20</code></td><td class=num>8/8</td><td class=num>0.50</td><td class=num>−0.007</td><td class=num>−0.013 to 0.001</td><td class=num>0</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.89</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.006</td><td class=num>−0.016 to 0.003</td><td class=num>0</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.1, by 10</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.006</td><td class=num>−0.011 to 0.006</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.2, by 10</code></td><td class=num>8/8</td><td class=num>0.52</td><td class=num>−0.003</td><td class=num>−0.012 to 0.005</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.87</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>8/8</td><td class=num>0.54</td><td class=num>−0.003</td><td class=num>−0.017 to 0.008</td><td class=num>0</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.86</td></tr><tr><td><code>τ_ℓ 0.2, by 20</code></td><td class=num>8/8</td><td class=num>0.48</td><td class=num>−0.002</td><td class=num>−0.012 to 0.033</td><td class=num>1</td><td class=num>3</td><td class=num>0.57</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.2, by 40</code></td><td class=num>8/8</td><td class=num>0.51</td><td class=num>−0.002</td><td class=num>−0.013 to 0.004</td><td class=num>0</td><td class=num>3</td><td class=num>0.54</td><td class=num>0.86</td></tr><tr><td><code>fixed τ, λ_a delay</code></td><td class=num>8/8</td><td class=num>0.52</td><td class=num>0.004</td><td class=num>−0.014 to 0.035</td><td class=num>1</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.85</td></tr><tr><td><code>τ_ℓ 0.4, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.48</td><td class=num>0.006</td><td class=num>−0.010 to 0.033</td><td class=num>2</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.89</td></tr><tr><td><code>τ_ℓ 0.4, by 10</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.009</td><td class=num>−0.009 to 0.034</td><td class=num>2</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.85</td></tr><tr><td><code>λ_a delay</code></td><td class=num>8/8</td><td class=num>0.51</td><td class=num>0.009</td><td class=num>−0.010 to 0.052</td><td class=num>2</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.86</td></tr><tr><td><code>τ_ℓ 0.4, by 20</code></td><td class=num>8/8</td><td class=num>0.47</td><td class=num>0.010</td><td class=num>−0.010 to 0.041</td><td class=num>2</td><td class=num>3</td><td class=num>0.54</td><td class=num>0.85</td></tr><tr><td><code>τ_ℓ 0.2, τ_ℓ fixed</code></td><td class=num>8/8</td><td class=num>0.53</td><td class=num>0.013</td><td class=num>−0.013 to 0.126</td><td class=num>2</td><td class=num>3</td><td class=num>0.56</td><td class=num>0.88</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>8/8</td><td class=num>0.45</td><td class=num>0.014</td><td class=num>−0.013 to 0.054</td><td class=num>3</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.85</td></tr><tr><td><code>λ_a stretch</code></td><td class=num>8/8</td><td class=num>0.36</td><td class=num>0.019</td><td class=num>−0.007 to 0.082</td><td class=num>2</td><td class=num>3</td><td class=num>0.53</td><td class=num>0.83</td></tr><tr><td><code>fixed τ</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.020</td><td class=num>−0.014 to 0.123</td><td class=num>3</td><td class=num>3</td><td class=num>0.53</td><td class=num>0.82</td></tr><tr><td><code>fixed τ, λ_a stretch</code></td><td class=num>8/8</td><td class=num>0.49</td><td class=num>0.021</td><td class=num>−0.011 to 0.177</td><td class=num>1</td><td class=num>3</td><td class=num>0.55</td><td class=num>0.84</td></tr></tbody></table><figcaption><p><strong>The conditions at epoch 64, ranked.</strong> Ranked by whether any run was not called kept, then by the mean
 spill. Kept: the runs the loss-rate call keeps (α at the example answers at slice 3,
@@ -73,15 +96,15 @@ the criterion (0.02). Peak slice and α at peak: where the seed-mean α at the e
 highest at epoch 64, and its value. α, query: the seed-mean α at the query answer at slice
 3.</p></figcaption></figure>
 
-The $\lambda_a$ conditions all stay well above the $\tau_\ell$ 0.1 level. A twenty-epoch delay, which gives e₁ time to clear under the full learning rate, lowered the mean spill somewhat, at the soft temperatures and at the fixed ones. The stretched warm-up raised it, and so did fixing both temperatures with the base warm-up. Each of these means rests on one or two seeds that spill far more than the rest, so at eight seeds the differences among them could be chance. None of them comes near what sharpening the slice pool does.
+The $\lambda_a$ conditions all stay well above the $\tau_\ell$ 0.1 level. One of them is not on an equal footing: under the stretched warm-up, $\lambda_a$ is still at 94% of its peak at the stop, so its spill at epoch 64 is of a pull not yet at full strength. A twenty-epoch delay, which gives e₁ time to clear under the full learning rate, lowered the mean spill somewhat, at the soft temperatures and at the fixed ones. The stretched warm-up raised it, and so did fixing both temperatures with the base warm-up. Each of these means rests on one or two seeds that spill far more than the rest, so at eight seeds the differences among them could be chance. None of them comes near what sharpening the slice pool does.
 
 Of the 21 runs over the criterion, 15 spill most on `darken`, the op on which the base spills early in spill-timing.
 
 These are early numbers. A sharper slice pool could still spill late, on `value-hsv` as the base does on its slow seeds, and only a finished run can show that. What we take from the screen is that $\tau_\ell$ is the lever on the early kind, and that the level at 0.1 brings the pool to the quiet recipes on it.
 
-## Where the op is heading (E3)
+## Where the evidence is heading (E3)
 
-A sharper slice pool could lower the spill in a way we would not want: by settling the pull on one slice and one role, as the joint pool did in [slice-terms](/docs/m2/slice-terms/report.py), where one seed settled on the query answer and the op could not be removed. Where does the evidence for the op sit at epoch 64?
+A sharper slice pool could lower the spill in a way we would not want: by settling the pull on one slice and one role, as the joint pool did in [slice-terms](/docs/m2/slice-terms/report.py), where one seed latched onto the query answer and the op could not be removed. Where does the evidence for the op sit at epoch 64?
 
 In the same place in every condition. The seed-mean α at the example answers peaks at slice 3 in all seventeen conditions, and α at the query answer is about the same everywhere (the table above). The figure shows the three levels of $\tau_\ell$ at the base timing, beside the condition that spilled most: the shapes are nearly the same, with one difference at slice 1.
 
@@ -92,6 +115,16 @@ In the same place in every condition. The seed-mean α at the example answers pe
 at the query answer (dashed), at each slice, for the three slice temperatures at the base timing and for the condition with the highest mean spill.
 The hairlines behind every panel are the seed means of every condition.</p></figcaption></figure>
 
+Beside the quiet recipes (below), the two conditions at $\tau_\ell$ 0.1 have about as much of the evidence at slice 1 as the step does, though the step pulls nothing there: some alignment reaches slice 1 from the pull on the slices after it. The base has half as much again. Past slice 1 the shapes differ as recipe-compare found at the end of training: the step holds the evidence level from slice 3 to slice 4, where the pool lets it fall back, at the example answers and at the query answer alike.
+
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-quiet-light.png" alt="Two panels of seed-mean α against slice. Left, at the example answers: every line starts near zero at the embedding. At slice 1 the two τ_ℓ 0.1 conditions and the two quiet recipes are together near 0.23, and the base is at about 0.37. The pool conditions peak near 0.56 at slice 3 and fall to about 0.38 at slice 4; falloff and step, soft τ level off near 0.5 from slice 3 to slice 4. Right, at the query answer: the same ordering at slice 1 (about 0.27, 0.33 and 0.43), all near 0.85 at slice 3, then the pool falls to about 0.62 at slice 4 while the quiet recipes stay near 0.85." width="626" height="280" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-quiet" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-quiet-dark.png" alt="Two panels of seed-mean α against slice. Left, at the example answers: every line starts near zero at the embedding. At slice 1 the two τ_ℓ 0.1 conditions and the two quiet recipes are together near 0.23, and the base is at about 0.37. The pool conditions peak near 0.56 at slice 3 and fall to about 0.38 at slice 4; falloff and step, soft τ level off near 0.5 from slice 3 to slice 4. Right, at the query answer: the same ordering at slice 1 (about 0.27, 0.33 and 0.43), all near 0.85 at slice 3, then the pool falls to about 0.62 at slice 4 while the quiet recipes stay near 0.85." width="626" height="280" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-quiet" />
+<figcaption><p><strong>The pool beside the quiet recipes at epoch 64.</strong> Seed-mean α by slice at the example answers
+(left) and at the query answer (right): the two stage-2 conditions and the base in their ink and
+marker, and the loss-rate screen's <code>falloff</code> and <code>step, soft τ</code> at the same epoch
+(34 seeds each) in gray. The hairlines are the other conditions.</p></figcaption></figure>
+
 That difference is the one that tracks the spill. At $\tau_\ell$ 0.1, α at the example answers at slice 1 averages 0.22, against 0.37 at 0.4. Across all 136 runs, the runs over the criterion are the ones with more of the anchor at slice 1, though at the same α some spill and some do not.
 
 <figure class="mini-themed-figure">
@@ -103,10 +136,114 @@ is the criterion.</p></figcaption></figure>
 
 This fits the account recipe-compare gave: the pool spills when it puts part of the pull on block 1, and the edit then takes something from the other ops there. A sharper slice pool takes more of the pull off slice 1 and puts it on slice 3, so block 1 holds less of the anchor and the edit has less to take. It fits without proving it; the screen shows the two moving together, and a restricted edit on the finished runs (block 1 alone) would test it.
 
+## Removal, spill, and task at epoch 400 (E4)
+
+Do the two finished conditions keep the anchor, stay within the spill criterion, and leave the task alone at the end of training, seed by seed, against the base and the two quiet recipes?
+
+They do. No finished run lost the anchor: the lowest removal is 0.83. Both conditions remove a little more than the base on most seeds (the paired column of the table), and a little more than `falloff` on the mean. Holding $\tau_\ell$ fixed gives the tightest removal of the five (the figure, left), with all eight seeds between 0.88 and 0.91.
+
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-finished-light.png" alt="Three panels with one column per condition. Left: removal sits between about 0.85 and 0.9 for every condition, with the two τ_ℓ 0.1 conditions highest and τ_ℓ 0.1, τ_ℓ fixed the tightest. Middle: spill sits near zero for every condition but the base, whose seeds reach 0.055; the two τ_ℓ 0.1 conditions have two seeds and one seed at the criterion line, none over it. Right: the task change sits near zero, the pool conditions a little higher than falloff and step, soft τ, all above the gate." width="644" height="290" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-finished" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-finished-dark.png" alt="Three panels with one column per condition. Left: removal sits between about 0.85 and 0.9 for every condition, with the two τ_ℓ 0.1 conditions highest and τ_ℓ 0.1, τ_ℓ fixed the tightest. Middle: spill sits near zero for every condition but the base, whose seeds reach 0.055; the two τ_ℓ 0.1 conditions have two seeds and one seed at the criterion line, none over it. Right: the task change sits near zero, the pool conditions a little higher than falloff and step, soft τ, all above the gate." width="644" height="290" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-finished" />
+<figcaption><p><strong>Removal, spill, and task at epoch 400, by condition.</strong> One column per condition, the
+8 seeds as small dots over a thin bar for their range, and the seed mean as the large mark.
+<strong>Left:</strong> removal at the full dose, net of the median over the other axes of the run. <strong>Middle:</strong> spill,
+the largest net drop on another op at any dose; the dashed line is the criterion, with the side over it
+hatched. <strong>Right:</strong> the task change, the mean clean score over ops less that of the control at the seed;
+the dashed line is the task gate on that mean, with the failing side hatched.</p></figcaption></figure>
+
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>removal ↑</th><th class=num>removal less base, paired</th><th class=num>spill ↓</th><th class=num>spill less base, paired</th><th class=num>over criterion ↓</th><th class=num>over, control median ↓</th><th class=num>task change ↑</th><th class=num>worst op change ↑</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>0.87 (0.77 to 0.91)</td><td class=num>—</td><td class=num>−0.002 (−0.009 to 0.014)</td><td class=num>—</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.007 (−0.032 to 0.005)</td><td class=num>−0.066 (<code>hue-hsv</code>, 708)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>0.86 (0.76 to 0.91)</td><td class=num>—</td><td class=num>−0.001 (−0.011 to 0.012)</td><td class=num>—</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.005 (−0.014 to 0.007)</td><td class=num>−0.036 (<code>hue-hsv</code>, 705)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>0.85 (0.81 to 0.91)</td><td class=num>—</td><td class=num>0.015 (−0.005 to 0.055)</td><td class=num>—</td><td class=num>2 of 8</td><td class=num>3 of 8</td><td class=num>0.001 (−0.003 to 0.008)</td><td class=num>−0.015 (<code>hue-hsv</code>, 708)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>0.89 (0.83 to 0.92)</td><td class=num>0.04 (−0.01 to 0.09); 6 up, 2 down</td><td class=num>0.003 (−0.011 to 0.020)</td><td class=num>−0.012 (−0.059 to 0.014); 3 up, 5 down</td><td class=num>0 of 8</td><td class=num>1 of 8</td><td class=num>−0.001 (−0.023 to 0.009)</td><td class=num>−0.037 (<code>darken</code>, 708)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>0.90 (0.88 to 0.91)</td><td class=num>0.05 (−0.01 to 0.08); 7 up, 1 down</td><td class=num>−0.001 (−0.011 to 0.020)</td><td class=num>−0.015 (−0.061 to 0.006); 3 up, 5 down</td><td class=num>0 of 8</td><td class=num>0 of 8</td><td class=num>−0.000 (−0.008 to 0.010)</td><td class=num>−0.023 (<code>hue-hsv</code>, 709)</td></tr></tbody></table><figcaption><p><strong>Removal, spill, and task at epoch 400</strong>, seed mean and range over 8 runs. Removal and
+spill are net of the median drop over the other 63 coordinate axes of the run, and "over, control median"
+counts the runs over the criterion with the median taken in the control at the seed instead. The paired
+columns are the condition less the base at the same seed, with the count of seeds on each side of zero. The
+task change is the mean over ops of the clean score less that of the control at the seed; the worst op change
+is the largest loss on one op at any seed. The rows for <code>falloff</code>, <code>step, soft τ</code> and the base are those of
+recipe-compare.</p></figcaption></figure>
+
+On spill (middle), neither condition has a run over the criterion, where the base has 2. The change shows best on the two seeds the base spills on most (the table by seed, below): at 711 both conditions sit below zero, and at 707 both come down to the line.
+
+That leaves one or two seeds per condition at the criterion, each within a thousandth of it: 705 and 707 under `τ_ℓ 0.1, by 40`, and 707 under `τ_ℓ 0.1, τ_ℓ fixed`. Taking the median over the other axes of the control at the seed, in place of those of the run, puts one run of `τ_ℓ 0.1, by 40` over the line (705), and none of `τ_ℓ 0.1, τ_ℓ fixed`.
+
+On the task (right), both pass the gate on the seed mean, and no op at any seed goes past the per-op line, so the task change stays near zero, as recipe-compare found for the pool. One run of `τ_ℓ 0.1, by 40` (708) is a little under the line for the mean over ops, at the seed where `falloff` is lowest too.
+
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>measure</th><th class=num>704</th><th class=num>705</th><th class=num>706</th><th class=num>707</th><th class=num>708</th><th class=num>709</th><th class=num>710</th><th class=num>711</th></tr></thead><tbody><tr><td><code>τ_ℓ 0.4, by 40</code></td><td class=num>removal</td><td class=num>0.83</td><td class=num>0.91</td><td class=num>0.82</td><td class=num>0.87</td><td class=num>0.83</td><td class=num>0.87</td><td class=num>0.84</td><td class=num>0.81</td></tr><tr><td></td><td class=num>spill</td><td class=num>−0.003</td><td class=num>0.020</td><td class=num>−0.005</td><td class=num><strong>0.027</strong></td><td class=num>0.012</td><td class=num>−0.003</td><td class=num>0.016</td><td class=num><strong>0.055</strong></td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>value-hsv</code></td><td class=num><code>darken</code></td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>removal</td><td class=num>0.92</td><td class=num>0.90</td><td class=num>0.87</td><td class=num>0.89</td><td class=num>0.83</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.90</td></tr><tr><td></td><td class=num>spill</td><td class=num>0.010</td><td class=num>0.020</td><td class=num>−0.004</td><td class=num>0.019</td><td class=num>−0.002</td><td class=num>−0.003</td><td class=num>−0.011</td><td class=num>−0.004</td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>hue-hsv</code></td><td class=num><code>mix</code></td><td class=num><code>darken</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>removal</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.89</td><td class=num>0.90</td><td class=num>0.91</td><td class=num>0.91</td><td class=num>0.90</td><td class=num>0.88</td></tr><tr><td></td><td class=num>spill</td><td class=num>−0.002</td><td class=num>−0.005</td><td class=num>0.001</td><td class=num>0.020</td><td class=num>−0.006</td><td class=num>0.003</td><td class=num>−0.011</td><td class=num>−0.006</td></tr><tr><td></td><td class=num>op</td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>mix</code></td><td class=num><code>lighten</code></td><td class=num><code>mix</code></td></tr></tbody></table><figcaption><p><strong>Removal and spill by seed</strong>, as in the table above, with the op the spill is largest on. Bold marks a run
+over the criterion.</p></figcaption></figure>
+
+## Spill along training (E5)
+
+[Spill-timing](/docs/m2/spill-timing/report.py) found two kinds of spill under the base: an early kind, which rises while the pool sharpens and already shows at epoch 64, and a late kind, which grows after epoch 100 on seeds that learn `value-hsv` late. Which kind is left on the seeds that come near the criterion at epoch 400?
+
+The late kind. Under the base, three seeds go over the criterion by epoch 64, and 711 does so already at epoch 32. Under the two finished conditions no seed goes over it by then, and up to epoch 128 the largest spill is 0.015 under `τ_ℓ 0.1, by 40` and 0.003 under `τ_ℓ 0.1, τ_ℓ fixed`. After that, two seeds under `τ_ℓ 0.1, by 40` (705 and 707) and one under `τ_ℓ 0.1, τ_ℓ fixed` (707) climb to about the criterion by the end.
+
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-along-light.png" alt="Three panels of spill against epoch, from 16 to 400 on a log scale, one line per seed. Base: three seeds rise over the criterion by epoch 64 (711 already at 32), and 707 and 711 stay over it to the end; two more seeds climb to near the criterion after epoch 128. τ_ℓ 0.1, by 40, and τ_ℓ 0.1, τ_ℓ fixed: every seed stays under 0.016 up to epoch 128; after that two seeds (705 and 707) and one seed (707) rise to about the criterion by epoch 400." width="644" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-along" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-along-dark.png" alt="Three panels of spill against epoch, from 16 to 400 on a log scale, one line per seed. Base: three seeds rise over the criterion by epoch 64 (711 already at 32), and 707 and 711 stay over it to the end; two more seeds climb to near the criterion after epoch 128. τ_ℓ 0.1, by 40, and τ_ℓ 0.1, τ_ℓ fixed: every seed stays under 0.016 up to epoch 128; after that two seeds (705 and 707) and one seed (707) rise to about the criterion by epoch 400." width="644" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-along" />
+<figcaption><p><strong>Spill along training.</strong> The spill of each run (one line per seed) at the checkpoints kept at epochs
+16, 32, 64, 128, 256 and at the end, scored as at the stop, against the
+median of eight reference edits rather than the 63 axes of E4, so a run at the line there can fall on
+either side of it here. The base is spill-timing's rerun of soft-pool, scored at every record.
+The dashed line is the criterion and the dotted line the stop of stage 1. The seeds named are the ones
+that go over the criterion or come within a quarter of it.</p></figcaption></figure>
+
+So the screen at epoch 64 called the early kind correctly: the conditions it found quiet stayed quiet through the part of training it covers.
+
+The late kind here is about as large as on the late seeds of the base (705 and 710 come to about 0.02 there too). It lands on different ops from those spill-timing saw, though: on `hue-hsv` and `darken` at 705 and 707 under `τ_ℓ 0.1, by 40`, and on `mix` at 707 under `τ_ℓ 0.1, τ_ℓ fixed`, where the late kind in the base was on `value-hsv`. Two or three runs are too few to say whether that is a different kind of late spill or the same one landing elsewhere.
+
+## Where the spill enters (E6)
+
+Recipe-compare traced the spill of the base to the edit at block 1, and E3 found that a sharper slice pool puts less of the anchor at slice 1. Does the edit at block 1 alone now cost the other ops less, and does the rest of what the edit does stay as it was?
+
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>peak α, seeds</th><th class=num>removal, block 1 alone ↑</th><th class=num>removal, blocks 2 to 4 ↑</th><th class=num>spill, embedding alone ↓</th><th class=num>spill, block 1 alone ↓</th><th class=num>spill, blocks 1 to 4 ↓</th><th class=num>spill, every slice ↓</th><th class=num>lightness on e₁, embedding table</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>slice 4: 8</td><td class=num>0.71 (0.38 to 0.95)</td><td class=num>0.81 (0.11 to 0.94)</td><td class=num>−0.000 (−0.003 to 0.002); 0 over</td><td class=num>0.005 (−0.004 to 0.022); 1 over</td><td class=num>0.004 (−0.004 to 0.018); 0 over</td><td class=num>0.001 (−0.013 to 0.018); 0 over</td><td class=num>−0.41 (−0.64 to −0.20)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>slice 4: 8</td><td class=num>0.69 (0.22 to 0.95)</td><td class=num>0.78 (0.06 to 0.91)</td><td class=num>0.000 (−0.003 to 0.004); 0 over</td><td class=num>0.016 (−0.003 to 0.031); 3 over</td><td class=num>0.011 (−0.017 to 0.038); 2 over</td><td class=num>0.006 (−0.020 to 0.029); 2 over</td><td class=num>−0.38 (−0.63 to −0.04)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>slice 3: 8</td><td class=num>0.77 (0.71 to 0.93)</td><td class=num>0.80 (0.04 to 0.93)</td><td class=num>0.003 (−0.002 to 0.010); 0 over</td><td class=num>0.027 (−0.008 to 0.090); 4 over</td><td class=num>0.021 (−0.007 to 0.079); 4 over</td><td class=num>0.021 (−0.006 to 0.071); 4 over</td><td class=num>−0.61 (−0.65 to −0.57)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>slice 3: 8</td><td class=num>0.68 (0.25 to 0.96)</td><td class=num>0.82 (0.09 to 0.94)</td><td class=num>0.001 (−0.002 to 0.004); 0 over</td><td class=num>0.019 (−0.003 to 0.043); 3 over</td><td class=num>0.014 (−0.007 to 0.038); 2 over</td><td class=num>0.010 (−0.005 to 0.030); 3 over</td><td class=num>−0.36 (−0.53 to −0.05)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>slice 3: 8</td><td class=num>0.75 (0.52 to 0.95)</td><td class=num>0.82 (0.17 to 0.96)</td><td class=num>−0.001 (−0.003 to 0.002); 0 over</td><td class=num>0.011 (−0.020 to 0.044); 2 over</td><td class=num>0.005 (−0.017 to 0.033); 1 over</td><td class=num>0.002 (−0.027 to 0.014); 0 over</td><td class=num>−0.41 (−0.59 to −0.15)</td></tr><tr><td>control</td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num></td><td class=num>−0.02 (−0.33 to 0.62)</td></tr></tbody></table><figcaption><p><strong>Where the evidence sits, and where the edit finds it.</strong> The slice at which α at the example answers of
+<code>difference</code> contexts peaks, with the count of seeds. Removal and spill under the full edit restricted to a
+set of slices, at every position, net of the same edit on the control at the seed (the median baseline exists
+only for the edit at every slice), with the count of runs over the criterion; seed mean and range. The last
+column is the correlation, over the color embeddings, of the e₁ component with lightness (the lean of
+embedding-lean). The first three rows are those of recipe-compare.</p></figcaption></figure>
+
+The evidence settles where it did. Under both conditions α at the example answers peaks at slice 3 on every seed, as under the base. The edit at blocks 2 to 4 removes most of the op on every seed but 710, which falls short under every recipe. The edit at block 4 alone removes nothing. The edit at the embedding alone spills nothing, and neither does the edit at blocks 2 to 4 (this one and the edit at block 4 are not in the table).
+
+The edit at block 1 alone costs the other ops less than under the base on the mean, and about what it costs them under `step, soft τ`: the mean is between the step and the base under `τ_ℓ 0.1, by 40` and below the step under `τ_ℓ 0.1, τ_ℓ fixed`.
+
+Seed by seed the two conditions differ. `τ_ℓ 0.1, τ_ℓ fixed` is below the base on six seeds of eight, where `τ_ℓ 0.1, by 40` is below it on three, and the lower mean under `τ_ℓ 0.1, by 40` comes from 711, the seed the base spills on most at block 1. So holding $\tau_\ell$ fixed closes block 1 about as far as leaving it out of the pull does, and sharpening to 0.1 by epoch 40 does so on the mean, carried by one seed.
+
+<figure class="mini-themed-figure">
+<img class="mini-themed-img-light" src="_assets/soft-pool-scout-block1-light.png" alt="Two panels, one point per run at each seed from 704 to 711, for the base and the two finished conditions. Left, the edit at block 1 alone: the base is highest at 707, 708 and 711 (0.04 to 0.09); τ_ℓ 0.1, by 40 sits near the base or above it on most seeds, between about 0.01 and 0.04, and well below it at 711; τ_ℓ 0.1, τ_ℓ fixed sits below the base on most seeds, with one high point at 709. Right, the edit at every slice: the base reaches 0.07 at 711; by 40 is over the criterion at 704, 705 and 707; τ_ℓ fixed stays under it everywhere." width="626" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-block1" />
+<img class="mini-themed-img-dark" src="_assets/soft-pool-scout-block1-dark.png" alt="Two panels, one point per run at each seed from 704 to 711, for the base and the two finished conditions. Left, the edit at block 1 alone: the base is highest at 707, 708 and 711 (0.04 to 0.09); τ_ℓ 0.1, by 40 sits near the base or above it on most seeds, between about 0.01 and 0.04, and well below it at 711; τ_ℓ 0.1, τ_ℓ fixed sits below the base on most seeds, with one high point at 709. Right, the edit at every slice: the base reaches 0.07 at 711; by 40 is over the criterion at 704, 705 and 707; τ_ℓ fixed stays under it everywhere." width="626" height="260" style="max-width: 100%; height: auto;" data-asset-name="soft-pool-scout-block1" />
+<figcaption><p><strong>Spill under the edit at block 1 alone, and at every slice, seed by seed.</strong> The largest drop on another
+op under the full edit made at those slices only, net of the same edit on the control at the seed, for
+the base (hollow) and the two finished conditions, offset a little at each seed. The dashed line is the
+criterion, with the side over it hatched.</p></figcaption></figure>
+
+As recipe-compare noted, the block-1 edit on its own overstates what the full edit costs: under `τ_ℓ 0.1, τ_ℓ fixed` it goes over the criterion on two seeds whose full edit stays within it.
+<!-- REVIEW: added the paired-by-seed reading of the block-1 edit. On the mean both conditions sit below the base, but `τ_ℓ 0.1, by 40` is above the base on five seeds of eight (SLICE_SPILL[n]["block 1"] against the base's, seed by seed), and without 711 its mean is above the base's, so the mean alone overstated it. The same holds for the control-netted edit at every slice, below (above the base on five seeds). Verify: the per-seed block-1 and every-slice spill of the base and the two leaders. -->
+
+With the edit at every slice, net of the control rather than of the median, `τ_ℓ 0.1, by 40` has three runs over the criterion and `τ_ℓ 0.1, τ_ℓ fixed` none. Under this baseline `step, soft τ` has two over, so these counts are best compared with one another, and on them holding $\tau_\ell$ fixed is the quieter of the two.
+
+The embedding table moves with the slice pool as well. Under the base the e₁ components of the color embeddings follow lightness closely on every seed (the lean of [embedding-lean](/docs/m2/embedding-lean/report.py)). Under both finished conditions the lean is back at about the level of the step, and varies from seed to seed as it does there, which fits a pull that reaches slice 1 less.
+
+<figure class="report-figure"><table class="report-table dense"><thead><tr><th>condition</th><th class=num>largest write over bound ↓</th><th class=num>sites over 1.25×, all runs ↓</th><th class=num>bound at example answers, emb and slice 1 ↓</th><th class=num>share removed ↑</th><th class=num>landing ↑</th></tr></thead><tbody><tr><td><code>falloff</code></td><td class=num>1.08 (slice 1, 0.3°)</td><td class=num>28 (≤ 4°)</td><td class=num>5°, 32°</td><td class=num>0.85 (0.79 to 0.90)</td><td class=num>0.52 (0.49 to 0.56)</td></tr><tr><td><code>step, soft τ</code></td><td class=num>1.15 (slice 1, 3.6°)</td><td class=num>23 (≤ 6°)</td><td class=num>6°, 32°</td><td class=num>0.83 (0.77 to 0.90)</td><td class=num>0.49 (0.47 to 0.53)</td></tr><tr><td><code>τ_ℓ 0.4, by 40</code> (base)</td><td class=num>1.33 (slice 1, 4.9°)</td><td class=num>48 (≤ 5°)</td><td class=num>14°, 43°</td><td class=num>0.83 (0.79 to 0.88)</td><td class=num>0.51 (0.47 to 0.55)</td></tr><tr><td><code>τ_ℓ 0.1, by 40</code></td><td class=num>1.03 (slice 1, 4.9°)</td><td class=num>13 (≤ 5°)</td><td class=num>4°, 27°</td><td class=num>0.86 (0.79 to 0.91)</td><td class=num>0.50 (0.45 to 0.55)</td></tr><tr><td><code>τ_ℓ 0.1, τ_ℓ fixed</code></td><td class=num>1.06 (slice 1, 4.5°)</td><td class=num>22 (≤ 4°)</td><td class=num>5°, 29°</td><td class=num>0.87 (0.78 to 0.91)</td><td class=num>0.52 (0.47 to 0.54)</td></tr></tbody></table><figcaption><p><strong>The write, the grading and the landing.</strong> The largest ratio of write to bound in angle over the slices past
+the embedding and every position, from the seed-mean quantiles as in edit-effects, with where it falls and
+the bound there. The sites over 1.25× count slice and position pairs over all 8 runs, each
+run taken on its own, with the largest bound among them. The bound at the example answers is over those that
+do not fit <code>difference</code>. The share removed is over every <code>difference</code> context, net of the control at the
+seed, and the landing is the share of the way to the target null in total variation; seed mean and range.
+The first three rows are those of recipe-compare.</p></figcaption></figure>
+
+What the edit does downstream barely changes, and where it changes it moves toward the step. The write stays close to the bound past the embedding, with fewer sites over the tolerance than under the base, and the share removed and the landing are about the same as under every recipe.
+
+Recipe-compare found the bound at the example answers of the other ops about 3× as wide at the embedding under the base as under `falloff`. Under both finished conditions it is back at the width of `falloff` at the embedding, and a little narrower at slice 1. So the other ops sit as far from e₁ at the first two slices as under the step.
+
 ## Discussion
 
-TODO after stage 2.
+With $\tau_\ell$ at 0.1, the pool over slices 1 to 3 keeps the anchor, removes a little more than the recipe of record, costs the task nothing, and stays within the spill criterion on every seed. The early spill that put two runs of the base over the criterion is gone.
 
-/// admonition | Open decision
-Which conditions go to stage 2. The plan finishes two or three leaders, ranked by no lost anchors and the lowest mean spill, preferring the fewest changes from the base when two are close. The four conditions at $\tau_\ell$ 0.1 are close together. A proposal: `τ_ℓ 0.1, by 40` (one change from the base) and `τ_ℓ 0.1, τ_ℓ fixed` (the simplest schedule, with only $\tau_p$ moving), and `τ_ℓ 0.2, by 40` as a middle step that would show whether 0.1 is needed or 0.2 is enough once the late kind is counted. To check: whether the finished runs keep their removal, since a sharper slice pool concentrates the pull.
-///
+So are most of the signs that block 1 is open to the edit. The wider bound at the first two slices and the lean in the embedding table are gone under both conditions, and the larger cost of the block-1 edit is gone under the condition with $\tau_\ell$ fixed.
+
+Of the two, holding $\tau_\ell$ at 0.1 from the start is the cleaner. It removes the most and the most evenly, has no run over the criterion under either median, and none under the control-netted edit at every slice. It also has one schedule fewer than the base, since only $\tau_p$ sharpens.
+
+The other condition is close behind on the median baseline of the D2.2 design, where the gap rests on two or three runs near the line. On the edits netted against the control the gap is larger: there the other condition sits at about the level of the base on most seeds, and only its mean is lower.
+
+What is left is the late kind of spill, which brings one or two seeds per condition to about the criterion by the end. It is about as large as under the base on its late seeds, so it does not seem to come from the sharper slice pool, and nothing in this study acts on it. When it appears, the term still changing is the anti weight, partway through its anneal, so if anything in the recipe acts on the late kind, that anneal is the likelier place.
+
+Before the pool with $\tau_\ell$ fixed at 0.1 could replace `falloff` as the recipe of record, it would need two things the step has: a count of lost anchors at forty seeds, which a loss-rate screen would give, and a set of fresh seeds, which a confirmation at seeds 712 to 719 with their controls would give. Since the pool chooses the slice within a range set by the architecture, it would also be the easier recipe to carry to another model, as recipe-compare argued.
+<!-- REVIEW: was "replace `step, soft τ`"; the report's own Scope and method and the D2.2 design doc name `falloff` as the recipe of record, and the two things named (forty seeds, fresh seeds) are what `falloff` has. Verify: docs/m2/d2.2/design.md, the recipe block. -->