Pin the assistant-axis commit, sort safetensors headers, name the code repo

README.md

@@ -85,7 +85,8 @@ alignment), and a pairwise judge with position swaps.
 
 **Assistant-axis control.** The per-role control vectors come from the extraction pipeline of
 Lu et al. (2026), [safety-research/assistant-axis](https://github.com/safety-research/assistant-axis)
-(MIT), run on OLMo-3-7B-Instruct: five role system prompts by 240 questions per role, each
+(MIT), at commit `a98961956072224eaf244eb289d6c01700b63795`, run on OLMo-3-7B-Instruct: five
+role system prompts by 240 questions per role, each
 rollout scored 0-3 for role-play by that pipeline's judge. A role's vector is the mean layer-16
 activation over its rollouts scored 3 (`type: pos_3`); `default` is the mean over the
 unprompted rollouts (`type: mean`). Their mean L2 norm is 9.68, so equal coefficients do not
@@ -116,9 +117,12 @@ are so the published numbers can be checked against exactly what produced them.
   is the one shipped, and the one the steering and geometry code used.
 - **Pairwise subset.** The 39 pairwise roles are the first 39 in alphabetical order
   (aberration through chimera), at alpha 2.5 with 50-pair vectors.
-- **Not recorded in the source runs:** the OLMo-3-7B-Instruct revision, the number of pairs
-  each role's judge accepted before reaching the 50-pair target, and the exact upstream
-  commit of the assistant-axis pipeline.
+- **Assistant-axis commit.** The source runs did not record which upstream commit they used.
+  The commit above is upstream's latest; its Python code is identical to every commit from
+  `05eca5bc` (15 January 2026) on, and the vectors were made before 22 March 2026, so only a
+  clone from 13-15 January would have run different code.
+- **Not recorded in the source runs:** the OLMo-3-7B-Instruct revision and the number of pairs
+  each role's judge accepted before reaching the 50-pair target.
 
 ## Intended use
 

manifest.json

@@ -1,4 +1,5 @@
 {
+  "code_repository": "HCAI-Lab-GT/casting-call-vectors",
   "facts": {
     "alignment_questions": {
       "duplicated_question_entries": [
@@ -32,7 +33,13 @@
       "types": [
         "mean",
         "pos_3"
-      ]
+      ],
+      "upstream": {
+        "commit": "a98961956072224eaf244eb289d6c01700b63795",
+        "python_code_unchanged_since": "05eca5bcc845ab323e11c85fd653d77c8725b289",
+        "recorded_by_source_run": false,
+        "repository": "https://github.com/safety-research/assistant-axis"
+      }
     },
     "castvectors": {
       "created_at": {
@@ -19654,7 +19661,7 @@
       "bytes": 4526248,
       "checked": "bit-exact vs JSON and pt_vectors",
       "path": "vectors/castvectors_olmo3-7b-instruct_L16_n50.safetensors",
-      "sha256": "35fb0d67de8bf1764838de7459918b6e72e3177e33e21fa5b930544730c5bb83",
+      "sha256": "eae7c5f68a6cbd41b0c75c1b85ee2c87d45852d0833a0b141d44081b5cf6f132",
       "sources": [
         {
           "blob": "940b69557f94f2b3707ad80617c9e8ca347ed5a4",
@@ -21862,7 +21869,7 @@
       "bytes": 3950552,
       "checked": "bit-exact vs JSON",
       "path": "vectors/trait_vectors_olmo3-7b-instruct_L16_n40.safetensors",
-      "sha256": "30e545d3f77cc937114a159c180ead6bdfbca9b870e804c9fe5fc38970a10b97",
+      "sha256": "9a0baf3b1e9c8b2632779b48df639825aed907f6c25453ad93ea472a8a55d06c",
       "sources": [
         {
           "blob": "daec6ddce0995a157c2da6fce8fcd63d9c04ddfb",
@@ -23106,6 +23113,6 @@
     "zealot",
     "zeitgeist"
   ],
-  "source_repository": "HCAI-Lab-GT/casting-call-vectors",
+  "source_repository": "HCAI-Lab-GT/casting-call-vectors-private",
   "source_revision": "4b54a4bc211f57f6c49d4824327c014549e9d815"
 }

vectors/castvectors_olmo3-7b-instruct_L16_n50.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:35fb0d67de8bf1764838de7459918b6e72e3177e33e21fa5b930544730c5bb83
+oid sha256:eae7c5f68a6cbd41b0c75c1b85ee2c87d45852d0833a0b141d44081b5cf6f132
 size 4526248

vectors/trait_vectors_olmo3-7b-instruct_L16_n40.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:30e545d3f77cc937114a159c180ead6bdfbca9b870e804c9fe5fc38970a10b97
+oid sha256:9a0baf3b1e9c8b2632779b48df639825aed907f6c25453ad93ea472a8a55d06c
 size 3950552