Upload folder using huggingface_hub

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/README.md

@@ -0,0 +1,5 @@
+# Issue 33 Native Selected-Region Student Artifacts
+
+This folder contains compact artifacts for the issue #33 EN->PT native selected-region student run: spec, notes, reusable scripts, FLORES teacher hypothesis data when packaged, NTREX eval summaries, generated hypotheses, XCOMET summaries, logs, and the native region delta checkpoint when present.
+
+The checkpoint is the experiment-owned selected-region slice edit only. Upstream HY-MT weights, XCOMET weights, Hugging Face cache directories, API keys, and service tokens are intentionally excluded.

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/SHA256SUMS

@@ -0,0 +1,45 @@
+5377eebba36af3224b8578fbc9e2c2147e388c8616478df454af1046572ba380  ./README.md
+57311c9080321e1501f1aad273141477511760ec83f310066003d9fa5bf86407  ./SHA256SUMS
+8e495b612ae0b39782e6f01339f003802e6e6d66dd2df7420baa7a05135abfa1  ./checkpoints/students/native_k160000_r8/native_region_delta.pt
+f6250df9e1f5c8bcdad0cf064f8ad5ad58cd3c413dbbc52db8f2fdcc6b742449  ./data/flores_teacher_train_hyps.jsonl
+5e62ef48e601ce757fc16457bcecd111b061e10f375fd0be2ee0577c7a005219  ./dumps/fixed_k160000.jsonl
+dcb0e935e85c599e638149bfdbf6b38e869acdfb6254b9441c71e0f15b95499a  ./dumps/native_region_k160000_r8.jsonl
+376c838cfc0186fa06d57e65807773320ac112ec3251adaffe16145533595f40  ./dumps/no_mask.jsonl
+389c5670cef01f37142b42d39e1d56b915a60f872a3aa697eb55454014518d79  ./eval/base_k160000_masks.json
+3b5d282c6ff4c675d5f2cdf57d0a4b6d56b815d8abaa8df060cb5182b4bd6866  ./eval/native_region_k160000_r8.json
+5d23689341f142dc78ac159c290c4c85fe7137cb14e119bbc51d4513f1121ddf  ./eval/r8_k160000_masks.json
+0d6096293966e0c1513e1cbf1babf8977bae85e7b74a189e5f0b4f5597dd747a  ./logs/build_flores.log
+35216e1631ff96c59dadabdb99d6b0e5ccd378a82eb2396e4ae4885951dcbd05  ./logs/build_ntrex.log
+354f6f1634af03560bf6e8bfb41786dda72bad51f890c3759449c5bee332f949  ./logs/download_hy_lora_conditions.log
+806dbbf60a8cf81357e0d8b5a2c7909adb0493f812715b0c87684856a794cbd9  ./logs/eval_base_k160000.log
+a6bf9e90e5d27e6b0f59eeb811871322478e4bd4ded5969d8d26b87f3d590d84  ./logs/eval_native_region_k160000_r8.log
+e5b4e6a536edd0357eff261f94da92609397acfdd84332ed0ceddccbd9765331  ./logs/eval_r8_k160000.log
+28a0b3ca62680abcbfc47769b0645dcf155e5fa23bef6440fb68b3f6a641078b  ./logs/generate_flores_teacher_hyps.log
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  ./logs/package_issue33_hf_upload.log
+7fe93e156c539a317d2f9e9e878c9883a81ef500930e2e259b82f4fc73101ab3  ./logs/progress.log
+b4427f5f59f03921a3d665e719c1aa7db65c517a61a566c209d2d80cbe5f12f6  ./logs/runner_outer.log
+d36549469909fec423d61a31cbe096f61f5be0a1d753312a331de896c30ddcc0  ./logs/summarize_issue33_k160000_r8.log
+a8a83a601e7c8742baff8625bfb233b4bd5c51afd96b69f8ab3e09f8b677a3fa  ./logs/train_native_region_k160000_r8.log
+2d7e9e4a7d252bb58aa1284035060d4df60a116c83a01afd2d58df30edb4d7f6  ./logs/xcomet_base_fixed_k160000.log
+c1cb58da8e4a407ea43a7fea3f1261033002ce41852e457bbf1ae0991738f130  ./logs/xcomet_native_region_k160000_r8.log
+9cd8ec5b86a7c4e1b6a67356d6729dbe0ee2e9ff0460b19ecf4c9712e36b2ed8  ./logs/xcomet_r8_fixed_k160000.log
+78efa64b813d573843b1a1a42ead9dec9258410b4130075c70c626d5dbf61edd  ./manifest.json
+f47756d43f191446b39a0d334f4ea6fd9f6e284780bca20829b2c95d2e357f78  ./manifests/students/native_k160000_r8/config.json
+b7c852002b46ed04766af112a27ad124735e01c82b2026120a52b6918cc6b171  ./manifests/students/native_k160000_r8/native_region_delta_config.json
+a10193dc61826976e2abb9497f37009fca8e3de5011777422df9122e1a78d0f6  ./manifests/students/native_k160000_r8/train_summary.json
+999bdc1224133f6979c0a38e82dfa47f981804e36a15e589ec8a121deee45a32  ./notes/issue33_native_region_mvc.md
+1e8fd77f80848dd8b601b180abf895a96c28c20673c503805542bbf0661ab40e  ./scripts/evaluate_native_region_delta_translation.py
+2f1927c360063d7772025c8bcfc9dfe0cc4293598f5744311bd82f4813551c6a  ./scripts/issue33_native_region_runner.sh
+f4bc614fc0943abbff070f4769cc0f8a7c074541ddcffa071aa5579d1dc8a3de  ./scripts/package_issue33_hf_upload.sh
+5a3f35a35acb4167cc9c4f349d3638cd007ca68f9b149169d4db17c481a7c0f6  ./scripts/summarize_issue33_native_region.py
+277b881a7b8bf170928b85871c70680619c4c793d5e732f20a6babd03861d4ce  ./scripts/train_native_region_delta_translation.py
+f297051a8e09341abf6f1d87ad330a7240ccc2388ba4081487efe41e39a2131d  ./scripts/translation_native_region.py
+5b25d49517c46c4661447930322054ab8b1acba6451171d441dcbe8f035f7d71  ./spec/native_region_mvc_issue33.json
+468a00a7a476a39d5aeca9dfb3d133d8d4864644d9390e99f75d63468f5c3fd7  ./summaries/issue33_native_region_k160000_r8.json
+2dbb25a2d883c7faf1ecaeb6149cde74e3d59951951fc10d81a6d1062a8e4e18  ./summaries/issue33_native_region_k160000_r8.md
+fdbbbc45d92b5bc682f434547a585137f6df055519ee86a433ec5c08c41f26b7  ./xcomet/base_fixed_k160000.json
+0827072b81648d4b68586c6a7b2cff916fbcc4d826587a4b4532d4a4398bd8ad  ./xcomet/base_fixed_k160000.scored_pool.jsonl
+065770e05d353611f48a431945937df58f3ad1a21477d801d9d2bde67db2756d  ./xcomet/native_region_k160000_r8.json
+a9dc9c58f2cbe7394261147fc62c29114e3b53ec65b61d8614cab3838c62e82a  ./xcomet/native_region_k160000_r8.scored_pool.jsonl
+8965fdcfaa323feb9898494449a8c38b3b79b75f87654af12cbd44e0b2db229c  ./xcomet/r8_fixed_k160000.json
+4cc25e049f6b71ba945bf5843566db8926acd78dbb6c76a26ec97b09a4b01cd4  ./xcomet/r8_fixed_k160000.scored_pool.jsonl

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/checkpoints/students/native_k160000_r8/native_region_delta.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e495b612ae0b39782e6f01339f003802e6e6d66dd2df7420baa7a05135abfa1
+size 21725255

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/eval/base_k160000_masks.json

@@ -0,0 +1,30 @@
+{
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "adapter": null,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "n_examples": 1012,
+  "n_calib": 64,
+  "results": {
+    "no_mask": {
+      "scores": {
+        "chrFpp": 54.22844417616828,
+        "chrF": 56.92735106386091,
+        "BLEU": 25.81246822014709,
+        "n": 1012
+      },
+      "kept": -1
+    },
+    "fixed_k160000": {
+      "scores": {
+        "chrFpp": 46.99461052474228,
+        "chrF": 49.88832385757205,
+        "BLEU": 18.22647479890369,
+        "n": 1012
+      },
+      "kept": 160000,
+      "mask_path": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+      "elapsed_s": 50.31123614311218
+    }
+  }
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/eval/native_region_k160000_r8.json

@@ -0,0 +1,18 @@
+{
+  "model": "tencent/HY-MT1.5-1.8B",
+  "student_dir": "/root/runs/issue33_native_region/students/native_k160000_r8",
+  "mask": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+  "mask_kept": 160000,
+  "rank": 8,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "scores": {
+    "chrFpp": 50.97431297760433,
+    "chrF": 53.79762252779031,
+    "BLEU": 22.184172798821177,
+    "n": 1012
+  },
+  "dump_hyps": "/root/runs/issue33_native_region/dumps/native_region_k160000_r8.jsonl",
+  "elapsed_s": 89.59548020362854,
+  "method": "native_selected_region_delta"
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/eval/r8_k160000_masks.json

@@ -0,0 +1,30 @@
+{
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "adapter": "/root/runs/issue33_native_region/hy_lora_conditions/low_rank_lens/k160_r8/adapter",
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "n_examples": 1012,
+  "n_calib": 64,
+  "results": {
+    "no_mask": {
+      "scores": {
+        "chrFpp": 53.72683437781636,
+        "chrF": 56.546862635098364,
+        "BLEU": 24.989224851370132,
+        "n": 1012
+      },
+      "kept": -1
+    },
+    "fixed_k160000": {
+      "scores": {
+        "chrFpp": 51.21928365554815,
+        "chrF": 53.989194431609235,
+        "BLEU": 22.699147288564674,
+        "n": 1012
+      },
+      "kept": 160000,
+      "mask_path": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+      "elapsed_s": 54.19245409965515
+    }
+  }
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/build_flores.log

@@ -0,0 +1,3 @@
+
+
+wrote 1012 pairs -> /root/runs/flores_eval/flores_en2pt_devtest.jsonl

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/build_ntrex.log

@@ -0,0 +1,6 @@
+
+NTREX-128 columns sample: ['sna_Latn', 'est_Latn', 'glg_Latn', 'bem_Latn', 'nob_Latn', 'zul_Latn', 'hye_Armn', 'nep_Deva'] ...
+NTREX-128 size: 1997 rows
+  src column: eng_Latn
+  tgt column: por_Latn
+wrote 1012 pairs -> /root/runs/ntrex_eval/ntrex_en2pt.jsonl

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/download_hy_lora_conditions.log

@@ -0,0 +1,2 @@
+
+/root/runs/issue33_native_region/hy_lora_conditions

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/eval_base_k160000.log

@@ -0,0 +1,12 @@
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.22844417616828, 'chrF': 56.92735106386091, 'BLEU': 25.81246822014709, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/base_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 46.99461052474228, 'chrF': 49.88832385757205, 'BLEU': 18.22647479890369, 'n': 1012} (50.3s)
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/base_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue33_native_region/eval/base_k160000_masks.json

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/eval_native_region_k160000_r8.log

@@ -0,0 +1,19 @@
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+{
+  "model": "tencent/HY-MT1.5-1.8B",
+  "student_dir": "/root/runs/issue33_native_region/students/native_k160000_r8",
+  "mask": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+  "mask_kept": 160000,
+  "rank": 8,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "scores": {
+    "chrFpp": 50.97431297760433,
+    "chrF": 53.79762252779031,
+    "BLEU": 22.184172798821177,
+    "n": 1012
+  },
+  "dump_hyps": "/root/runs/issue33_native_region/dumps/native_region_k160000_r8.jsonl",
+  "elapsed_s": 89.59548020362854,
+  "method": "native_selected_region_delta"
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/eval_r8_k160000.log

@@ -0,0 +1,14 @@
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] adapter /root/runs/issue33_native_region/hy_lora_conditions/low_rank_lens/k160_r8/adapter
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 53.72683437781636, 'chrF': 56.546862635098364, 'BLEU': 24.989224851370132, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/r8_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 51.21928365554815, 'chrF': 53.989194431609235, 'BLEU': 22.699147288564674, 'n': 1012} (54.2s)
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/r8_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue33_native_region/eval/r8_k160000_masks.json

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/generate_flores_teacher_hyps.log

@@ -0,0 +1,5 @@
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+[load] 1012 pairs (slice [0,1012) of 1012) from /root/runs/flores_eval/flores_en2pt_devtest.jsonl
+[gen] load tencent/HY-MT1.5-1.8B
+[gen] 1012 in 47s (21.7 sent/s)
+[save] /root/runs/issue33_native_region/data/flores_teacher_train_hyps.jsonl  (1012 rows)

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/progress.log

@@ -0,0 +1,14 @@
+[2026-05-13T16:44:53+00:00] issue33 start model=tencent/HY-MT1.5-1.8B k=160000 ranks=8 xcomet_mode=direct
+[2026-05-13T16:44:53+00:00] download preserved mask/reference adapters from Occupying-Mars/hy-lora-conditions
+[2026-05-13T16:44:59+00:00] build FLORES EN->PT devtest jsonl
+[2026-05-13T16:45:04+00:00] build NTREX EN->PT held-out jsonl
+[2026-05-13T16:45:07+00:00] generate FLORES teacher-hyp train data
+[2026-05-13T16:46:08+00:00] evaluate base selected-region anchor
+[2026-05-13T16:47:56+00:00] evaluate best trained reference from issue #28 r8 for comparison
+[2026-05-13T16:49:48+00:00] train native selected-region student k=160000 rank=8
+[2026-05-13T16:50:54+00:00] evaluate native selected-region student k=160000 rank=8
+[2026-05-13T16:52:29+00:00] score base selected-region anchor
+[2026-05-13T17:00:55+00:00] score best trained reference from issue #28 r8
+[2026-05-13T17:08:01+00:00] score native selected-region student k=160000 rank=8
+[2026-05-13T17:14:58+00:00] summarize issue33 k=160000 rank=8
+[2026-05-13T17:14:58+00:00] package and upload issue33 artifacts

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/runner_outer.log

@@ -0,0 +1,60 @@
+[2026-05-13T16:44:53+00:00] issue33 start model=tencent/HY-MT1.5-1.8B k=160000 ranks=8 xcomet_mode=direct
+[2026-05-13T16:44:53+00:00] download preserved mask/reference adapters from Occupying-Mars/hy-lora-conditions
+
+/root/runs/issue33_native_region/hy_lora_conditions
+[2026-05-13T16:44:59+00:00] build FLORES EN->PT devtest jsonl
+
+
+wrote 1012 pairs -> /root/runs/flores_eval/flores_en2pt_devtest.jsonl
+[2026-05-13T16:45:04+00:00] build NTREX EN->PT held-out jsonl
+
+NTREX-128 columns sample: ['sna_Latn', 'est_Latn', 'glg_Latn', 'bem_Latn', 'nob_Latn', 'zul_Latn', 'hye_Armn', 'nep_Deva'] ...
+NTREX-128 size: 1997 rows
+  src column: eng_Latn
+  tgt column: por_Latn
+wrote 1012 pairs -> /root/runs/ntrex_eval/ntrex_en2pt.jsonl
+[2026-05-13T16:45:07+00:00] generate FLORES teacher-hyp train data
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+[load] 1012 pairs (slice [0,1012) of 1012) from /root/runs/flores_eval/flores_en2pt_devtest.jsonl
+[gen] load tencent/HY-MT1.5-1.8B
+[gen] 1012 in 47s (21.7 sent/s)
+[save] /root/runs/issue33_native_region/data/flores_teacher_train_hyps.jsonl  (1012 rows)
+[2026-05-13T16:46:08+00:00] evaluate base selected-region anchor
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.22844417616828, 'chrF': 56.92735106386091, 'BLEU': 25.81246822014709, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/base_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 46.99461052474228, 'chrF': 49.88832385757205, 'BLEU': 18.22647479890369, 'n': 1012} (50.3s)
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/base_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue33_native_region/eval/base_k160000_masks.json
+[2026-05-13T16:47:56+00:00] evaluate best trained reference from issue #28 r8 for comparison
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] adapter /root/runs/issue33_native_region/hy_lora_conditions/low_rank_lens/k160_r8/adapter
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 53.72683437781636, 'chrF': 56.546862635098364, 'BLEU': 24.989224851370132, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/r8_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 51.21928365554815, 'chrF': 53.989194431609235, 'BLEU': 22.699147288564674, 'n': 1012} (54.2s)
+     dumped 1012 hyps -> /root/runs/issue33_native_region/dumps/r8_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue33_native_region/eval/r8_k160000_masks.json
+[2026-05-13T16:49:48+00:00] train native selected-region student k=160000 rank=8
+[2026-05-13T16:50:54+00:00] evaluate native selected-region student k=160000 rank=8
+[2026-05-13T16:52:29+00:00] score base selected-region anchor
+[2026-05-13T17:00:55+00:00] score best trained reference from issue #28 r8
+[2026-05-13T17:08:01+00:00] score native selected-region student k=160000 rank=8
+[2026-05-13T17:14:58+00:00] summarize issue33 k=160000 rank=8
+{
+  "out_json": "/root/runs/issue33_native_region/summaries/issue33_native_region_k160000_r8.json",
+  "out_md": "/root/runs/issue33_native_region/summaries/issue33_native_region_k160000_r8.md"
+}
+[2026-05-13T17:14:58+00:00] package and upload issue33 artifacts

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/summarize_issue33_k160000_r8.log

@@ -0,0 +1,4 @@
+{
+  "out_json": "/root/runs/issue33_native_region/summaries/issue33_native_region_k160000_r8.json",
+  "out_md": "/root/runs/issue33_native_region/summaries/issue33_native_region_k160000_r8.md"
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/train_native_region_k160000_r8.log

@@ -0,0 +1,11 @@
+{"data_rows": 1012, "dropped_rows": 0}
+{"mask_kept": 160000, "mask": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz"}
+{"rank": 8, "trainable_parameters": 5412864, "zero_init_means_initial_eval_matches_base_selected_region": true}
+{"step": 1, "loss": 0.748046875, "masked_kl": 0.607421875, "ce": 0.7060546875, "lr": 3.3333333333333335e-05, "elapsed_s": 0.6507542133331299}
+{"step": 25, "loss": 0.6422627766927084, "masked_kl": 0.5158182779947916, "ce": 0.6314798990885416, "lr": 0.00018807712330634642, "elapsed_s": 12.182272672653198}
+{"step": 50, "loss": 0.46009765625, "masked_kl": 0.36353515625, "ce": 0.48236328125, "lr": 0.00014154150130018866, "elapsed_s": 23.924031734466553}
+{"step": 75, "loss": 0.3659765625, "masked_kl": 0.2878173828125, "ce": 0.3909765625, "lr": 7.810966761934053e-05, "elapsed_s": 35.229416608810425}
+{"step": 100, "loss": 0.367216796875, "masked_kl": 0.2874072265625, "ce": 0.3992724609375, "lr": 2.3581308246275103e-05, "elapsed_s": 45.786269187927246}
+{"step": 125, "loss": 0.3673876953125, "masked_kl": 0.286630859375, "ce": 0.4037890625, "lr": 1.3479116011769767e-07, "elapsed_s": 57.23423790931702}
+{"step": 127, "loss": 0.39501953125, "masked_kl": 0.311279296875, "ce": 0.4197998046875, "lr": 0.0, "elapsed_s": 58.13042879104614}
+{"done": true, "out_dir": "/root/runs/issue33_native_region/students/native_k160000_r8", "elapsed_s": 58.131454944610596}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/xcomet_base_fixed_k160000.log

@@ -0,0 +1,56 @@
+[load] 1012 rows from /root/runs/issue33_native_region/dumps/base_k160000/fixed_k160000.jsonl
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 128/1012 mean=0.5578 throughput=2.90 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 256/1012 mean=0.5899 throughput=3.01 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 384/1012 mean=0.6023 throughput=3.08 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 512/1012 mean=0.6041 throughput=3.08 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 640/1012 mean=0.6153 throughput=3.08 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 768/1012 mean=0.6158 throughput=3.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 896/1012 mean=0.6235 throughput=3.02 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 1012/1012 mean=0.6259 throughput=2.98 seg/s
+[save] scored pool -> /root/runs/issue33_native_region/xcomet/base_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue33_native_region/xcomet/base_fixed_k160000.json

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/xcomet_native_region_k160000_r8.log

@@ -0,0 +1,56 @@
+[load] 1012 rows from /root/runs/issue33_native_region/dumps/native_region_k160000_r8.jsonl
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 128/1012 mean=0.7877 throughput=3.04 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 256/1012 mean=0.7973 throughput=3.11 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 384/1012 mean=0.8007 throughput=3.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 512/1012 mean=0.8006 throughput=3.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 640/1012 mean=0.7974 throughput=3.11 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 768/1012 mean=0.7923 throughput=3.06 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 896/1012 mean=0.7925 throughput=3.07 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 1012/1012 mean=0.7951 throughput=3.04 seg/s
+[save] scored pool -> /root/runs/issue33_native_region/xcomet/native_region_k160000_r8.scored_pool.jsonl
+[save] summary -> /root/runs/issue33_native_region/xcomet/native_region_k160000_r8.json

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/logs/xcomet_r8_fixed_k160000.log

@@ -0,0 +1,56 @@
+[load] 1012 rows from /root/runs/issue33_native_region/dumps/r8_k160000/fixed_k160000.jsonl
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 128/1012 mean=0.8068 throughput=2.93 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 256/1012 mean=0.8169 throughput=2.99 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 384/1012 mean=0.8189 throughput=3.04 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 512/1012 mean=0.8223 throughput=3.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 640/1012 mean=0.8135 throughput=3.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 768/1012 mean=0.8111 throughput=2.97 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 896/1012 mean=0.8100 throughput=2.98 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1]
+[chunk] 1012/1012 mean=0.8104 throughput=2.97 seg/s
+[save] scored pool -> /root/runs/issue33_native_region/xcomet/r8_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue33_native_region/xcomet/r8_fixed_k160000.json

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/manifest.json

@@ -0,0 +1,42 @@
+{
+  "issue": 33,
+  "task": "Native selected-region EN->PT student training",
+  "run_root": "/root/runs/issue33_native_region",
+  "upload_prefix": "issue33_native_region_20260513T171458Z",
+  "summaries": {
+    "issue33_native_region_k160000_r8.json": {
+      "issue": 33,
+      "run_root": "/root/runs/issue33_native_region",
+      "k": 160000,
+      "rank": 8,
+      "rows": [
+        {
+          "condition": "base_selected_region_k160000",
+          "role": "initial selected-region baseline",
+          "xcomet": 0.6259405999205152,
+          "reference_xcomet": 0.6278288431,
+          "chrFpp": 46.99461052474228
+        },
+        {
+          "condition": "best_trained_reference_issue28_r8",
+          "role": "comparison reference, not teacher",
+          "xcomet": 0.8103810173308426,
+          "reference_xcomet": 0.8145537329,
+          "chrFpp": 51.21928365554815
+        },
+        {
+          "condition": "native_region_k160000_r8",
+          "role": "issue33 trained native selected-region student",
+          "xcomet": 0.7950550597565977,
+          "reference_xcomet": null,
+          "chrFpp": 50.97431297760433
+        }
+      ],
+      "native_delta_vs_base_selected": 0.16911445983608253,
+      "native_delta_vs_best_trained_reference": -0.01532595757424482,
+      "pass_floor_for_continuation": true
+    }
+  },
+  "file_count": 42,
+  "weights_policy": "Includes only experiment-native small checkpoints. Excludes upstream HY-MT/XCOMET weights, HF caches, API keys, and service tokens."
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/manifests/students/native_k160000_r8/config.json

@@ -0,0 +1,59 @@
+{
+  "issue": 33,
+  "script": "train_native_region_delta_translation.py",
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/issue33_native_region/data/flores_teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/issue33_native_region/students/native_k160000_r8",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 33,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "rank": 8,
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 4,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 25,
+    "num_workers": 0
+  },
+  "n_rows": 1012,
+  "dropped_rows": 0,
+  "n_layers": 32,
+  "hidden_size": 2048,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "rank": 8,
+  "trainable_parameters": 5412864,
+  "total_steps": 127,
+  "warmup_steps": 6,
+  "logs": [],
+  "method": {
+    "student": "selected native MLP slices",
+    "selected_slice_edits": [
+      "gate_proj rows",
+      "up_proj rows",
+      "down_proj columns"
+    ],
+    "nonselected_mlp_channels": "mean ablated at down_proj input",
+    "teacher_logits": "full unmasked base model",
+    "global_adapter": false,
+    "peft_lora": false,
+    "initial_state": "base selected-region masked circuit plus zero low-rank slice edits"
+  }
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/manifests/students/native_k160000_r8/native_region_delta_config.json

@@ -0,0 +1,14 @@
+{
+  "rank": 8,
+  "hidden_size": 2048,
+  "mask_path": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+  "means_path": "/root/runs/issue33_native_region/students/native_k160000_r8/means.pt",
+  "trainable_parameters": 5412864,
+  "parameterization": "selected native MLP gate/up/down low-rank slice edits",
+  "model": "tencent/HY-MT1.5-1.8B",
+  "target_language": "Portuguese",
+  "prompt_style": "hy_mt",
+  "mask_kept": 160000,
+  "n_layers": 32,
+  "d_ffn": 6144
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/manifests/students/native_k160000_r8/train_summary.json

@@ -0,0 +1,117 @@
+{
+  "issue": 33,
+  "script": "train_native_region_delta_translation.py",
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/issue33_native_region/data/flores_teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/issue33_native_region/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/issue33_native_region/students/native_k160000_r8",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 33,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "rank": 8,
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 4,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 25,
+    "num_workers": 0
+  },
+  "n_rows": 1012,
+  "dropped_rows": 0,
+  "n_layers": 32,
+  "hidden_size": 2048,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "rank": 8,
+  "trainable_parameters": 5412864,
+  "total_steps": 127,
+  "warmup_steps": 6,
+  "logs": [
+    {
+      "step": 1,
+      "loss": 0.748046875,
+      "masked_kl": 0.607421875,
+      "ce": 0.7060546875,
+      "lr": 3.3333333333333335e-05,
+      "elapsed_s": 0.6507542133331299
+    },
+    {
+      "step": 25,
+      "loss": 0.6422627766927084,
+      "masked_kl": 0.5158182779947916,
+      "ce": 0.6314798990885416,
+      "lr": 0.00018807712330634642,
+      "elapsed_s": 12.182272672653198
+    },
+    {
+      "step": 50,
+      "loss": 0.46009765625,
+      "masked_kl": 0.36353515625,
+      "ce": 0.48236328125,
+      "lr": 0.00014154150130018866,
+      "elapsed_s": 23.924031734466553
+    },
+    {
+      "step": 75,
+      "loss": 0.3659765625,
+      "masked_kl": 0.2878173828125,
+      "ce": 0.3909765625,
+      "lr": 7.810966761934053e-05,
+      "elapsed_s": 35.229416608810425
+    },
+    {
+      "step": 100,
+      "loss": 0.367216796875,
+      "masked_kl": 0.2874072265625,
+      "ce": 0.3992724609375,
+      "lr": 2.3581308246275103e-05,
+      "elapsed_s": 45.786269187927246
+    },
+    {
+      "step": 125,
+      "loss": 0.3673876953125,
+      "masked_kl": 0.286630859375,
+      "ce": 0.4037890625,
+      "lr": 1.3479116011769767e-07,
+      "elapsed_s": 57.23423790931702
+    },
+    {
+      "step": 127,
+      "loss": 0.39501953125,
+      "masked_kl": 0.311279296875,
+      "ce": 0.4197998046875,
+      "lr": 0.0,
+      "elapsed_s": 58.13042879104614
+    }
+  ],
+  "method": {
+    "student": "selected native MLP slices",
+    "selected_slice_edits": [
+      "gate_proj rows",
+      "up_proj rows",
+      "down_proj columns"
+    ],
+    "nonselected_mlp_channels": "mean ablated at down_proj input",
+    "teacher_logits": "full unmasked base model",
+    "global_adapter": false,
+    "peft_lora": false,
+    "initial_state": "base selected-region masked circuit plus zero low-rank slice edits"
+  },
+  "elapsed_s": 58.131454944610596
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/notes/issue33_native_region_mvc.md

@@ -0,0 +1,34 @@
+# Issue 33 Native Selected-Region MVC
+
+Goal: train the selected EN->PT MLP region itself into a stronger masked-region
+student. This run starts from the live base selected-region masked circuit, not a
+fresh replacement writer and not a global LoRA adapter.
+
+Pipeline:
+
+```text
+preserved ReLP k=160k mask
+  -> base masked-region anchor on NTREX
+  -> preserved best trained reference from issue #28 r8 for comparison
+  -> FLORES teacher-hyp rows from unmasked Tencent HY-MT
+  -> train rank-8 native selected-region slice edits
+  -> NTREX generation
+  -> XCOMET-XXL scoring
+  -> backup to TokenBender/synth-data-en-pt-circuit and TokenBender/circuit-discovery
+```
+
+The trainable object is a set of local low-rank edits over selected native MLP
+slices:
+
+- selected `gate_proj` output rows
+- selected `up_proj` output rows
+- selected `down_proj` input columns
+
+During the masked objective and evaluation, selected MLP intermediate channels
+stay live and non-selected MLP intermediate channels are replaced by their
+calibration mean at the `down_proj` input. The full unmasked base model supplies
+teacher logits for distillation. The best trained reference from issue #28 r8 is
+only a comparison point.
+
+First-pass continuation rule: continue beyond rank 8 only if the native student
+beats the reproduced base selected-region baseline by at least `+0.03` XCOMET.

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/evaluate_native_region_delta_translation.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+"""Evaluate a native selected-region EN->PT student under its fixed mask."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from pathlib import Path
+
+import sacrebleu
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from translation_io import DEFAULT_PROMPT_STYLE, Pair, generate_translations, load_flores_devtest_any
+from translation_native_region import load_native_region_delta
+from translation_region_student import count_mask, load_mask_npz, text_config
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--model", required=True)
+    p.add_argument("--student-dir", required=True)
+    p.add_argument("--mask", required=True)
+    p.add_argument("--out", required=True)
+    p.add_argument("--dump-hyps", required=True)
+    p.add_argument("--device", default="cuda")
+    p.add_argument("--dtype", default="bfloat16", choices=["float32", "float16", "bfloat16"])
+    p.add_argument("--target-language", default="Portuguese")
+    p.add_argument("--prompt-style", default=DEFAULT_PROMPT_STYLE)
+    p.add_argument("--src-lang", default="eng_Latn")
+    p.add_argument("--tgt-lang", default="por_Latn")
+    p.add_argument("--input-jsonl", default=None)
+    p.add_argument("--max-examples", type=int, default=None)
+    p.add_argument("--batch-size", type=int, default=8)
+    p.add_argument("--max-new-tokens", type=int, default=384)
+    p.add_argument("--category", default="ntrex_test")
+    p.add_argument("--tag", default="heldout")
+    p.add_argument("--mask-name", default=None)
+    return p.parse_args()
+
+
+def dtype_from_name(name: str) -> torch.dtype:
+    return {
+        "float32": torch.float32,
+        "float16": torch.float16,
+        "bfloat16": torch.bfloat16,
+    }[name]
+
+
+def load_pairs(path: str | None, *, src_lang: str, tgt_lang: str,
+               max_examples: int | None) -> list[Pair]:
+    if path is None:
+        return load_flores_devtest_any(
+            src_lang=src_lang,
+            tgt_lang=tgt_lang,
+            max_examples=max_examples,
+        )
+    pairs: list[Pair] = []
+    with Path(path).open() as f:
+        for line in f:
+            if not line.strip():
+                continue
+            row = json.loads(line)
+            src = row.get("en") or row.get("src")
+            tgt = row.get("pt") or row.get("tgt")
+            if src and tgt:
+                pairs.append(Pair(src=src, tgt=tgt))
+                if max_examples is not None and len(pairs) >= max_examples:
+                    break
+    return pairs
+
+
+def main() -> None:
+    args = parse_args()
+    dtype = dtype_from_name(args.dtype)
+    out_path = Path(args.out)
+    dump_path = Path(args.dump_hyps)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    dump_path.parent.mkdir(parents=True, exist_ok=True)
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model,
+        dtype=dtype,
+        attn_implementation="eager",
+    ).to(args.device).eval()
+    model.config.use_cache = True
+    for param in model.parameters():
+        param.requires_grad_(False)
+
+    cfg = text_config(model)
+    n_layers = int(cfg.num_hidden_layers)
+    hidden_size = int(cfg.hidden_size)
+    d_ffn = int(cfg.intermediate_size)
+    mask = load_mask_npz(args.mask, n_layers, d_ffn)
+    kept = count_mask(mask)
+    means = torch.load(Path(args.student_dir) / "means.pt", map_location="cpu")
+    controller = load_native_region_delta(
+        args.student_dir,
+        mask=mask,
+        means=means,
+        hidden_size=hidden_size,
+        map_location="cpu",
+    ).to(args.device).eval()
+
+    pairs = load_pairs(
+        args.input_jsonl,
+        src_lang=args.src_lang,
+        tgt_lang=args.tgt_lang,
+        max_examples=args.max_examples,
+    )
+    sources = [pair.src for pair in pairs]
+    refs = [pair.tgt for pair in pairs]
+
+    t0 = time.time()
+    hooks = controller.install(model)
+    try:
+        hyps = generate_translations(
+            model,
+            tokenizer,
+            sources,
+            target_language=args.target_language,
+            prompt_style=args.prompt_style,
+            batch_size=args.batch_size,
+            max_new_tokens=args.max_new_tokens,
+            do_sample=False,
+            device=args.device,
+        )
+    finally:
+        for hook in hooks:
+            hook.remove()
+    elapsed_s = time.time() - t0
+
+    with dump_path.open("w") as f:
+        for i, (src, ref, hyp) in enumerate(zip(sources, refs, hyps)):
+            f.write(json.dumps({
+                "id": i,
+                "en": src,
+                "pt": ref,
+                "model_hyp": hyp,
+                "category": args.category,
+                "tag": args.tag,
+                "mask_name": args.mask_name or Path(args.mask).stem,
+                "student_dir": args.student_dir,
+                "rank": int(controller.rank),
+                "kept": kept,
+                "method": "native_selected_region_delta",
+            }, ensure_ascii=False) + "\n")
+
+    scores = {
+        "chrFpp": sacrebleu.corpus_chrf(hyps, [refs], word_order=2).score,
+        "chrF": sacrebleu.corpus_chrf(hyps, [refs], word_order=0).score,
+        "BLEU": sacrebleu.corpus_bleu(hyps, [refs]).score,
+        "n": len(hyps),
+    }
+    payload = {
+        "model": args.model,
+        "student_dir": args.student_dir,
+        "mask": args.mask,
+        "mask_kept": kept,
+        "rank": int(controller.rank),
+        "n_layers": n_layers,
+        "d_ffn": d_ffn,
+        "scores": scores,
+        "dump_hyps": str(dump_path),
+        "elapsed_s": elapsed_s,
+        "method": "native_selected_region_delta",
+    }
+    out_path.write_text(json.dumps(payload, indent=2, ensure_ascii=False) + "\n")
+    print(json.dumps(payload, indent=2, ensure_ascii=False), flush=True)
+
+
+if __name__ == "__main__":
+    main()

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/issue33_native_region_runner.sh

@@ -0,0 +1,284 @@
+#!/usr/bin/env bash
+# Issue #33 native selected-region EN->PT student runner.
+set -euo pipefail
+
+export PATH="/root/.local/bin:$PATH"
+REPO_DIR="${REPO_DIR:-/root/work/circuit-shotting}"
+cd "$REPO_DIR"
+source .venv/bin/activate
+
+MODEL="${MODEL:-tencent/HY-MT1.5-1.8B}"
+ARTIFACT_REPO="${ARTIFACT_REPO:-Occupying-Mars/hy-lora-conditions}"
+RUN_ROOT="${RUN_ROOT:-/root/runs/issue33_native_region}"
+ARTIFACT_DIR="${ARTIFACT_DIR:-$RUN_ROOT/hy_lora_conditions}"
+FLORES_JSONL="${FLORES_JSONL:-/root/runs/flores_eval/flores_en2pt_devtest.jsonl}"
+TRAIN_JSONL="${TRAIN_JSONL:-$RUN_ROOT/data/flores_teacher_train_hyps.jsonl}"
+NTREX_JSONL="${NTREX_JSONL:-/root/runs/ntrex_eval/ntrex_en2pt.jsonl}"
+K="${K:-160000}"
+SHORT_K="${SHORT_K:-$((K / 1000))}"
+MASK="${MASK:-$ARTIFACT_DIR/masks/base_attr/relp_k${K}.full.npz}"
+RANKS="${RANKS:-8}"
+GPU_LIST="${GPU_LIST:-0 1}"
+XCOMET_MODE="${XCOMET_MODE:-service}" # service | direct | skip
+XCOMET_MODEL="${XCOMET_MODEL:-Unbabel/XCOMET-XXL}"
+SCORE_AFTER="${SCORE_AFTER:-1}"
+UPLOAD_AFTER="${UPLOAD_AFTER:-1}"
+ISSUE33_USE_PLAIN_PYTHON="${ISSUE33_USE_PLAIN_PYTHON:-0}"
+
+mkdir -p "$RUN_ROOT"/{data,eval,dumps,xcomet,students,logs,summaries} "$ARTIFACT_DIR"
+
+mark() {
+  echo "[$(date -Iseconds)] $*" | tee -a "$RUN_ROOT/logs/progress.log"
+}
+
+run_python() {
+  if [[ "$ISSUE33_USE_PLAIN_PYTHON" == "1" || "$ISSUE33_USE_PLAIN_PYTHON" == "true" ]]; then
+    python "$@"
+  else
+    uv run python "$@"
+  fi
+}
+
+pick_gpu() {
+  local idx="$1"
+  python - "$GPU_LIST" "$idx" <<'PY'
+import sys
+gpus=[g for g in sys.argv[1].replace(",", " ").split() if g]
+if not gpus:
+    gpus=["0"]
+print(gpus[int(sys.argv[2]) % len(gpus)])
+PY
+}
+
+extra_train_args=()
+if [[ -n "${MAX_ROWS:-}" ]]; then extra_train_args+=(--max-rows "$MAX_ROWS"); fi
+if [[ -n "${MAX_STEPS:-}" ]]; then extra_train_args+=(--max-steps "$MAX_STEPS"); fi
+
+score_jsonl() {
+  local hyps_jsonl="$1"
+  local out_json="$2"
+  local request_id="$3"
+  local system_name="$4"
+  local log_path="$5"
+
+  [[ -f "$hyps_jsonl" ]] || {
+    echo "[issue33] missing hyps jsonl: $hyps_jsonl" >&2
+    return 1
+  }
+  [[ -f "$out_json" ]] && return 0
+  if [[ "$XCOMET_MODE" == "skip" ]]; then
+    mark "skip XCOMET for $system_name"
+    return 0
+  fi
+  if [[ "$XCOMET_MODE" == "service" ]]; then
+    run_python score_xcomet_service_client.py \
+      --hyps-jsonl "$hyps_jsonl" \
+      --out-json "$out_json" \
+      --request-id "$request_id" \
+      --system-name "$system_name" \
+      --batch-size "${XCOMET_BATCH_SIZE:-8}" \
+      --chunk-size "${XCOMET_CHUNK_SIZE:-128}" \
+      --timeout-s "${XCOMET_TIMEOUT_S:-7200}" \
+      > "$log_path" 2>&1
+    return 0
+  fi
+  if [[ "$XCOMET_MODE" == "direct" ]]; then
+    run_python score_xcomet_pool.py \
+      --hyps-jsonl "$hyps_jsonl" \
+      --out-jsonl "${out_json%.json}.scored_pool.jsonl" \
+      --summary-json "$out_json" \
+      --comet-model "$XCOMET_MODEL" \
+      --system-name "$system_name" \
+      --batch-size "${XCOMET_BATCH_SIZE:-8}" \
+      --chunk-size "${XCOMET_CHUNK_SIZE:-128}" \
+      > "$log_path" 2>&1
+    return 0
+  fi
+  echo "[issue33] unknown XCOMET_MODE=$XCOMET_MODE" >&2
+  return 1
+}
+
+mark "issue33 start model=$MODEL k=$K ranks=$RANKS xcomet_mode=$XCOMET_MODE"
+
+if [[ ! -f "$MASK" ]]; then
+  mark "download preserved mask/reference adapters from $ARTIFACT_REPO"
+  run_python - "$ARTIFACT_REPO" "$ARTIFACT_DIR" "$K" "$SHORT_K" <<'PY' \
+    2>&1 | tee "$RUN_ROOT/logs/download_hy_lora_conditions.log"
+import sys
+from huggingface_hub import snapshot_download
+
+repo, local_dir, k, short_k = sys.argv[1:5]
+patterns = [
+    "README.md",
+    "manifest.json",
+    f"low_rank_lens/k{short_k}_r8/adapter/*",
+    f"low_rank_lens/k{short_k}_r8/config.json",
+    f"low_rank_lens/k{short_k}_r8/train_summary.json",
+    f"masks/base_attr/relp_k{k}.full.npz",
+]
+print(snapshot_download(
+    repo_id=repo,
+    repo_type="model",
+    local_dir=local_dir,
+    allow_patterns=patterns,
+))
+PY
+fi
+
+if [[ ! -f "$MASK" ]]; then
+  echo "missing fixed mask: $MASK" >&2
+  exit 2
+fi
+
+if [[ ! -f "$FLORES_JSONL" ]]; then
+  mark "build FLORES EN->PT devtest jsonl"
+  run_python build_flores_en2pt_jsonl.py 2>&1 | tee "$RUN_ROOT/logs/build_flores.log"
+fi
+
+if [[ ! -f "$NTREX_JSONL" ]]; then
+  mark "build NTREX EN->PT held-out jsonl"
+  run_python build_ntrex_en2pt_jsonl.py 2>&1 | tee "$RUN_ROOT/logs/build_ntrex.log"
+fi
+
+if [[ ! -f "$TRAIN_JSONL" ]]; then
+  mark "generate FLORES teacher-hyp train data"
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" run_python gen_translations_only.py \
+    --jsonl "$FLORES_JSONL" \
+    --model "$MODEL" \
+    --out "$TRAIN_JSONL" \
+    --target-language Portuguese \
+    --batch-size "${TEACHER_BATCH_SIZE:-16}" \
+    --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/generate_flores_teacher_hyps.log"
+fi
+
+if [[ ! -f "$RUN_ROOT/eval/base_k${K}_masks.json" ]]; then
+  mark "evaluate base selected-region anchor"
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" run_python evaluate_translation_adapter_masks.py \
+    --base-model "$MODEL" \
+    --input-jsonl "$NTREX_JSONL" \
+    --mask "fixed_k${K}:$MASK" \
+    --out "$RUN_ROOT/eval/base_k${K}_masks.json" \
+    --target-language Portuguese \
+    --include-no-mask \
+    --dump-hyps-dir "$RUN_ROOT/dumps/base_k${K}" \
+    --dump-category ntrex_test \
+    --dump-tag heldout \
+    --batch-size "${EVAL_BATCH_SIZE:-16}" \
+    --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/eval_base_k${K}.log"
+fi
+
+if [[ ! -f "$RUN_ROOT/eval/r8_k${K}_masks.json" ]]; then
+  mark "evaluate best trained reference from issue #28 r8 for comparison"
+  ADAPTER="$ARTIFACT_DIR/low_rank_lens/k${SHORT_K}_r8/adapter"
+  [[ -f "$ADAPTER/adapter_model.safetensors" ]] || {
+    echo "missing reference adapter: $ADAPTER" >&2
+    exit 3
+  }
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 1)" run_python evaluate_translation_adapter_masks.py \
+    --base-model "$MODEL" \
+    --adapter "$ADAPTER" \
+    --input-jsonl "$NTREX_JSONL" \
+    --mask "fixed_k${K}:$MASK" \
+    --out "$RUN_ROOT/eval/r8_k${K}_masks.json" \
+    --target-language Portuguese \
+    --include-no-mask \
+    --dump-hyps-dir "$RUN_ROOT/dumps/r8_k${K}" \
+    --dump-category ntrex_test \
+    --dump-tag heldout \
+    --batch-size "${EVAL_BATCH_SIZE:-16}" \
+    --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/eval_r8_k${K}.log"
+fi
+
+for rank in $RANKS; do
+  student_dir="$RUN_ROOT/students/native_k${K}_r${rank}"
+  if [[ ! -f "$student_dir/train_summary.json" ]]; then
+    mark "train native selected-region student k=$K rank=$rank"
+    mkdir -p "$student_dir"
+    CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" run_python train_native_region_delta_translation.py \
+      --model "$MODEL" \
+      --jsonl "$TRAIN_JSONL" \
+      --fixed-mask "$MASK" \
+      --out-dir "$student_dir" \
+      --target-field model_hyp \
+      --target-language Portuguese \
+      --rank "$rank" \
+      --batch-size "${BATCH_SIZE:-1}" \
+      --grad-accum "${GRAD_ACCUM:-8}" \
+      --epochs "${EPOCHS:-1.0}" \
+      --lr "${LR:-2e-4}" \
+      --masked-kl-beta "${MASKED_KL_BETA:-1.0}" \
+      --ce-beta "${CE_BETA:-0.2}" \
+      --n-calib "${N_CALIB:-128}" \
+      --eval-every "${EVAL_EVERY:-25}" \
+      "${extra_train_args[@]}" \
+      > "$RUN_ROOT/logs/train_native_region_k${K}_r${rank}.log" 2>&1
+  fi
+
+  eval_json="$RUN_ROOT/eval/native_region_k${K}_r${rank}.json"
+  dump_jsonl="$RUN_ROOT/dumps/native_region_k${K}_r${rank}.jsonl"
+  if [[ ! -f "$eval_json" ]]; then
+    mark "evaluate native selected-region student k=$K rank=$rank"
+    CUDA_VISIBLE_DEVICES="$(pick_gpu 1)" run_python evaluate_native_region_delta_translation.py \
+      --model "$MODEL" \
+      --student-dir "$student_dir" \
+      --mask "$MASK" \
+      --input-jsonl "$NTREX_JSONL" \
+      --out "$eval_json" \
+      --dump-hyps "$dump_jsonl" \
+      --target-language Portuguese \
+      --batch-size "${EVAL_BATCH_SIZE:-8}" \
+      --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+      --mask-name "native_region_k${K}_r${rank}" \
+      > "$RUN_ROOT/logs/eval_native_region_k${K}_r${rank}.log" 2>&1
+  fi
+done
+
+if [[ "$SCORE_AFTER" == "1" || "$SCORE_AFTER" == "true" ]]; then
+  mark "score base selected-region anchor"
+  score_jsonl \
+    "$RUN_ROOT/dumps/base_k${K}/fixed_k${K}.jsonl" \
+    "$RUN_ROOT/xcomet/base_fixed_k${K}.json" \
+    "issue33_base_fixed_k${K}" \
+    "issue33_base_fixed_k${K}" \
+    "$RUN_ROOT/logs/xcomet_base_fixed_k${K}.log"
+
+  mark "score best trained reference from issue #28 r8"
+  score_jsonl \
+    "$RUN_ROOT/dumps/r8_k${K}/fixed_k${K}.jsonl" \
+    "$RUN_ROOT/xcomet/r8_fixed_k${K}.json" \
+    "issue33_r8_fixed_k${K}" \
+    "issue33_r8_fixed_k${K}" \
+    "$RUN_ROOT/logs/xcomet_r8_fixed_k${K}.log"
+
+  for rank in $RANKS; do
+    mark "score native selected-region student k=$K rank=$rank"
+    score_jsonl \
+      "$RUN_ROOT/dumps/native_region_k${K}_r${rank}.jsonl" \
+      "$RUN_ROOT/xcomet/native_region_k${K}_r${rank}.json" \
+      "issue33_native_region_k${K}_r${rank}" \
+      "issue33_native_region_k${K}_r${rank}" \
+      "$RUN_ROOT/logs/xcomet_native_region_k${K}_r${rank}.log"
+  done
+fi
+
+for rank in $RANKS; do
+  mark "summarize issue33 k=$K rank=$rank"
+  run_python summarize_issue33_native_region.py \
+    --run-root "$RUN_ROOT" \
+    --k "$K" \
+    --rank "$rank" \
+    --out-json "$RUN_ROOT/summaries/issue33_native_region_k${K}_r${rank}.json" \
+    --out-md "$RUN_ROOT/summaries/issue33_native_region_k${K}_r${rank}.md" \
+    2>&1 | tee "$RUN_ROOT/logs/summarize_issue33_k${K}_r${rank}.log"
+done
+
+if [[ "$UPLOAD_AFTER" == "1" || "$UPLOAD_AFTER" == "true" ]]; then
+  mark "package and upload issue33 artifacts"
+  RUN_ROOT="$RUN_ROOT" REPO_DIR="$REPO_DIR" scripts/package_issue33_hf_upload.sh \
+    2>&1 | tee "$RUN_ROOT/logs/package_issue33_hf_upload.log"
+fi
+
+mark "issue33 done"

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/package_issue33_hf_upload.sh

@@ -0,0 +1,106 @@
+#!/usr/bin/env bash
+# Package and upload issue #33 native selected-region student artifacts.
+set -euo pipefail
+
+RUN_ROOT="${RUN_ROOT:-/root/runs/issue33_native_region}"
+REPO_DIR="${REPO_DIR:-/root/work/circuit-shotting}"
+HF_REPO="${HF_REPO:-TokenBender/synth-data-en-pt-circuit}"
+CIRCUIT_HF_REPO="${CIRCUIT_HF_REPO:-TokenBender/circuit-discovery}"
+UPLOAD_ROOT="${UPLOAD_ROOT:-$RUN_ROOT/hf_upload}"
+STAMP="${STAMP:-$(date -u +%Y%m%dT%H%M%SZ)}"
+UPLOAD_PREFIX="${UPLOAD_PREFIX:-issue33_native_region_$STAMP}"
+UPLOAD_CHECKPOINTS="${UPLOAD_CHECKPOINTS:-1}"
+UPLOAD_DIR="$UPLOAD_ROOT/$UPLOAD_PREFIX"
+ISSUE33_USE_PLAIN_PYTHON="${ISSUE33_USE_PLAIN_PYTHON:-0}"
+
+run_hf() {
+  if [[ "$ISSUE33_USE_PLAIN_PYTHON" == "1" || "$ISSUE33_USE_PLAIN_PYTHON" == "true" ]]; then
+    hf "$@"
+  else
+    uv run hf "$@"
+  fi
+}
+
+mkdir -p "$UPLOAD_DIR"/{spec,notes,scripts,data,eval,dumps,xcomet,summaries,logs,manifests}
+
+cp "$REPO_DIR/configs/native_region_mvc_issue33.json" "$UPLOAD_DIR/spec/"
+cp "$REPO_DIR/notes/issue33_native_region_mvc.md" "$UPLOAD_DIR/notes/"
+cp "$REPO_DIR/issue33_native_region_runner.sh" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/train_native_region_delta_translation.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/evaluate_native_region_delta_translation.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/translation_native_region.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/summarize_issue33_native_region.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/scripts/package_issue33_hf_upload.sh" "$UPLOAD_DIR/scripts/"
+
+cp "$RUN_ROOT"/data/*.jsonl "$UPLOAD_DIR/data/" 2>/dev/null || true
+cp "$RUN_ROOT"/eval/*.json "$UPLOAD_DIR/eval/" 2>/dev/null || true
+find "$RUN_ROOT/dumps" -type f -name '*.jsonl' -exec cp {} "$UPLOAD_DIR/dumps/" \; 2>/dev/null || true
+cp "$RUN_ROOT"/xcomet/*.json "$UPLOAD_DIR/xcomet/" 2>/dev/null || true
+cp "$RUN_ROOT"/xcomet/*.jsonl "$UPLOAD_DIR/xcomet/" 2>/dev/null || true
+cp "$RUN_ROOT"/summaries/* "$UPLOAD_DIR/summaries/" 2>/dev/null || true
+cp "$RUN_ROOT"/logs/*.log "$UPLOAD_DIR/logs/" 2>/dev/null || true
+
+if [[ -d "$RUN_ROOT/students" ]]; then
+  find "$RUN_ROOT/students" \( -name 'train_summary.json' -o -name 'config.json' -o -name 'native_region_delta_config.json' \) \
+    | while read -r path; do
+        rel="${path#$RUN_ROOT/}"
+        mkdir -p "$UPLOAD_DIR/manifests/$(dirname "$rel")"
+        cp "$path" "$UPLOAD_DIR/manifests/$rel"
+      done
+fi
+
+if [[ "$UPLOAD_CHECKPOINTS" == "1" || "$UPLOAD_CHECKPOINTS" == "true" ]]; then
+  if [[ -d "$RUN_ROOT/students" ]]; then
+    find "$RUN_ROOT/students" -name 'native_region_delta.pt' \
+      | while read -r path; do
+          rel="${path#$RUN_ROOT/}"
+          mkdir -p "$UPLOAD_DIR/checkpoints/$(dirname "$rel")"
+          cp "$path" "$UPLOAD_DIR/checkpoints/$rel"
+        done
+  fi
+fi
+
+python - "$UPLOAD_DIR" "$RUN_ROOT" "$UPLOAD_PREFIX" <<'PY'
+import json
+import os
+import sys
+from pathlib import Path
+
+upload = Path(sys.argv[1])
+run_root = Path(sys.argv[2])
+prefix = sys.argv[3]
+
+def maybe_json(path: Path):
+    return json.loads(path.read_text()) if path.exists() else None
+
+summaries = {
+    path.name: maybe_json(path)
+    for path in sorted((run_root / "summaries").glob("issue33_native_region_*.json"))
+}
+manifest = {
+    "issue": 33,
+    "task": "Native selected-region EN->PT student training",
+    "run_root": str(run_root),
+    "upload_prefix": prefix,
+    "summaries": summaries,
+    "file_count": sum(len(files) for _, _, files in os.walk(upload)),
+    "weights_policy": "Includes only experiment-native small checkpoints. Excludes upstream HY-MT/XCOMET weights, HF caches, API keys, and service tokens.",
+}
+(upload / "manifest.json").write_text(json.dumps(manifest, indent=2, ensure_ascii=False) + "\n")
+(upload / "README.md").write_text(
+    "# Issue 33 Native Selected-Region Student Artifacts\n\n"
+    "This folder contains compact artifacts for the issue #33 EN->PT native "
+    "selected-region student run: spec, notes, reusable scripts, FLORES teacher "
+    "hypothesis data when packaged, NTREX eval summaries, generated hypotheses, "
+    "XCOMET summaries, logs, and the native region delta checkpoint when present.\n\n"
+    "The checkpoint is the experiment-owned selected-region slice edit only. "
+    "Upstream HY-MT weights, XCOMET weights, Hugging Face cache directories, API "
+    "keys, and service tokens are intentionally excluded.\n"
+)
+print(json.dumps({"upload_dir": str(upload), "prefix": prefix, "files": manifest["file_count"]}, indent=2))
+PY
+
+(cd "$UPLOAD_DIR" && find . -type f -print0 | sort -z | xargs -0 sha256sum > SHA256SUMS)
+run_hf upload "$HF_REPO" "$UPLOAD_DIR" "$UPLOAD_PREFIX" --repo-type dataset
+run_hf upload "$CIRCUIT_HF_REPO" "$UPLOAD_DIR" "circuit-shotting/artifacts/issue33/$UPLOAD_PREFIX" --repo-type dataset
+echo "$UPLOAD_PREFIX"

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/summarize_issue33_native_region.py

@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+"""Summarize issue #33 native selected-region student artifacts."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+
+REFERENCE = {
+    "base_fixed_k160000": 0.6278288431,
+    "best_trained_reference_issue28_r8": 0.8145537329,
+}
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--run-root", type=Path, required=True)
+    p.add_argument("--k", type=int, default=160000)
+    p.add_argument("--rank", type=int, default=8)
+    p.add_argument("--out-json", type=Path, required=True)
+    p.add_argument("--out-md", type=Path, required=True)
+    return p.parse_args()
+
+
+def load_json(path: Path) -> dict[str, Any] | None:
+    if not path.exists():
+        return None
+    return json.loads(path.read_text())
+
+
+def xcomet(path: Path) -> float | None:
+    payload = load_json(path)
+    if not payload:
+        return None
+    for key in ("system_score", "system_xcomet_xxl"):
+        val = payload.get(key)
+        if isinstance(val, (int, float)):
+            return float(val)
+    summary = payload.get("summary")
+    if isinstance(summary, dict):
+        val = summary.get("system_score")
+        if isinstance(val, (int, float)):
+            return float(val)
+    return None
+
+
+def eval_metric(path: Path, block: str, key: str) -> float | None:
+    payload = load_json(path)
+    if not payload:
+        return None
+    row = (payload.get("results") or {}).get(block) or {}
+    score = (row.get("scores") or {}).get(key)
+    return float(score) if isinstance(score, (int, float)) else None
+
+
+def native_metric(path: Path, key: str) -> float | None:
+    payload = load_json(path)
+    if not payload:
+        return None
+    score = (payload.get("scores") or {}).get(key)
+    return float(score) if isinstance(score, (int, float)) else None
+
+
+def fmt(value: float | None) -> str:
+    return "" if value is None else f"{value:.6f}"
+
+
+def main() -> None:
+    args = parse_args()
+    root = args.run_root
+    native_name = f"native_region_k{args.k}_r{args.rank}"
+    rows = [
+        {
+            "condition": f"base_selected_region_k{args.k}",
+            "role": "initial selected-region baseline",
+            "xcomet": xcomet(root / "xcomet" / f"base_fixed_k{args.k}.json"),
+            "reference_xcomet": REFERENCE.get(f"base_fixed_k{args.k}"),
+            "chrFpp": eval_metric(root / "eval" / f"base_k{args.k}_masks.json",
+                                  f"fixed_k{args.k}", "chrFpp"),
+        },
+        {
+            "condition": "best_trained_reference_issue28_r8",
+            "role": "comparison reference, not teacher",
+            "xcomet": xcomet(root / "xcomet" / f"r8_fixed_k{args.k}.json"),
+            "reference_xcomet": REFERENCE["best_trained_reference_issue28_r8"],
+            "chrFpp": eval_metric(root / "eval" / f"r8_k{args.k}_masks.json",
+                                  f"fixed_k{args.k}", "chrFpp"),
+        },
+        {
+            "condition": native_name,
+            "role": "issue33 trained native selected-region student",
+            "xcomet": xcomet(root / "xcomet" / f"{native_name}.json"),
+            "reference_xcomet": None,
+            "chrFpp": native_metric(root / "eval" / f"{native_name}.json", "chrFpp"),
+        },
+    ]
+    base = rows[0]["xcomet"]
+    native = rows[-1]["xcomet"]
+    reference = rows[1]["xcomet"] or rows[1]["reference_xcomet"]
+    summary = {
+        "issue": 33,
+        "run_root": str(root),
+        "k": args.k,
+        "rank": args.rank,
+        "rows": rows,
+        "native_delta_vs_base_selected": None if base is None or native is None else native - base,
+        "native_delta_vs_best_trained_reference": (
+            None if reference is None or native is None else native - reference
+        ),
+        "pass_floor_for_continuation": (
+            None if base is None or native is None else native >= base + 0.03
+        ),
+    }
+    args.out_json.parent.mkdir(parents=True, exist_ok=True)
+    args.out_json.write_text(json.dumps(summary, indent=2, ensure_ascii=False) + "\n")
+
+    lines = [
+        "# Issue 33 Native Selected-Region Student Summary",
+        "",
+        f"- Run root: `{root}`",
+        f"- Budget: `k={args.k}`, rank `{args.rank}`",
+        "- Teacher for training: full unmasked base logits.",
+        "- Best trained reference from issue #28 r8 is comparison only.",
+        "",
+        "| condition | role | XCOMET | reference XCOMET | chrF++ |",
+        "|---|---|---:|---:|---:|",
+    ]
+    for row in rows:
+        lines.append(
+            "| `{}` | {} | {} | {} | {} |".format(
+                row["condition"],
+                row["role"],
+                fmt(row["xcomet"]),
+                fmt(row["reference_xcomet"]),
+                fmt(row["chrFpp"]),
+            )
+        )
+    lines.extend([
+        "",
+        f"Native delta vs base selected region: `{fmt(summary['native_delta_vs_base_selected'])}`.",
+        "Continuation floor (+0.03 XCOMET over base selected region): "
+        f"`{summary['pass_floor_for_continuation']}`.",
+    ])
+    args.out_md.parent.mkdir(parents=True, exist_ok=True)
+    args.out_md.write_text("\n".join(lines) + "\n")
+    print(json.dumps({"out_json": str(args.out_json), "out_md": str(args.out_md)}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/train_native_region_delta_translation.py

@@ -0,0 +1,290 @@
+#!/usr/bin/env python3
+"""Train selected native EN->PT MLP slices under a fixed region mask."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import random
+import time
+from pathlib import Path
+
+import torch
+from torch.utils.data import DataLoader
+from transformers import AutoModelForCausalLM, AutoTokenizer, get_cosine_schedule_with_warmup
+
+from translation_io import DEFAULT_PROMPT_STYLE
+from translation_native_region import NativeRegionDeltaController, save_native_region_delta
+from translation_region_student import (
+    TranslationJsonlDataset,
+    answer_ce_loss,
+    build_mean_cache,
+    collate_translation,
+    count_mask,
+    kl_loss,
+    load_mask_npz,
+    load_translation_rows,
+    move_batch,
+    text_config,
+)
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--model", required=True)
+    p.add_argument("--jsonl", required=True)
+    p.add_argument("--fixed-mask", required=True)
+    p.add_argument("--out-dir", required=True)
+    p.add_argument("--target-field", default="model_hyp")
+    p.add_argument("--target-language", default="Portuguese")
+    p.add_argument("--prompt-style", default=DEFAULT_PROMPT_STYLE)
+    p.add_argument("--device", default="cuda")
+    p.add_argument("--dtype", default="bfloat16", choices=["float32", "float16", "bfloat16"])
+    p.add_argument("--seed", type=int, default=33)
+    p.add_argument("--max-rows", type=int, default=None)
+    p.add_argument("--max-seq-length", type=int, default=1024)
+    p.add_argument("--n-calib", type=int, default=128)
+    p.add_argument("--mean-on", default="full", choices=["prompt", "full"])
+    p.add_argument("--rank", type=int, required=True)
+    p.add_argument("--epochs", type=float, default=1.0)
+    p.add_argument("--max-steps", type=int, default=None)
+    p.add_argument("--batch-size", type=int, default=1)
+    p.add_argument("--grad-accum", type=int, default=8)
+    p.add_argument("--lr", type=float, default=2e-4)
+    p.add_argument("--weight-decay", type=float, default=0.0)
+    p.add_argument("--warmup-ratio", type=float, default=0.05)
+    p.add_argument("--max-grad-norm", type=float, default=1.0)
+    p.add_argument("--masked-kl-beta", type=float, default=1.0)
+    p.add_argument("--ce-beta", type=float, default=0.2)
+    p.add_argument("--kl-temperature", type=float, default=1.0)
+    p.add_argument("--kl-on", default="answer", choices=["prompt", "answer", "all"])
+    p.add_argument("--eval-every", type=int, default=25)
+    p.add_argument("--num-workers", type=int, default=0)
+    return p.parse_args()
+
+
+def dtype_from_name(name: str) -> torch.dtype:
+    return {
+        "float32": torch.float32,
+        "float16": torch.float16,
+        "bfloat16": torch.bfloat16,
+    }[name]
+
+
+def main() -> None:
+    args = parse_args()
+    out_dir = Path(args.out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    dtype = dtype_from_name(args.dtype)
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    rows = load_translation_rows(args.jsonl, target_field=args.target_field, max_rows=args.max_rows)
+    random.shuffle(rows)
+    dataset = TranslationJsonlDataset(
+        rows,
+        tokenizer,
+        target_language=args.target_language,
+        prompt_style=args.prompt_style,
+        max_seq_length=args.max_seq_length,
+        kl_on=args.kl_on,
+    )
+    if len(dataset) == 0:
+        raise ValueError("no usable training rows")
+    loader = DataLoader(
+        dataset,
+        batch_size=args.batch_size,
+        shuffle=True,
+        num_workers=args.num_workers,
+        collate_fn=lambda xs: collate_translation(xs, tokenizer.pad_token_id),
+    )
+    print(json.dumps({"data_rows": len(dataset), "dropped_rows": dataset.dropped}), flush=True)
+
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model,
+        dtype=dtype,
+        attn_implementation="eager",
+    ).to(args.device).eval()
+    model.config.use_cache = False
+    for param in model.parameters():
+        param.requires_grad_(False)
+
+    cfg = text_config(model)
+    n_layers = int(cfg.num_hidden_layers)
+    hidden_size = int(cfg.hidden_size)
+    d_ffn = int(cfg.intermediate_size)
+    mask = load_mask_npz(args.fixed_mask, n_layers, d_ffn)
+    kept = count_mask(mask)
+    print(json.dumps({"mask_kept": kept, "mask": args.fixed_mask}), flush=True)
+
+    means = build_mean_cache(
+        model,
+        dataset.rows,
+        n_calib=args.n_calib,
+        mean_on=args.mean_on,
+        device=args.device,
+        dtype=dtype,
+        n_layers=n_layers,
+        d_ffn=d_ffn,
+    )
+    means_path = out_dir / "means.pt"
+    torch.save(means, means_path)
+
+    controller = NativeRegionDeltaController(
+        mask=mask,
+        means=means,
+        hidden_size=hidden_size,
+        rank=args.rank,
+    ).to(args.device)
+    controller.train()
+    trainable_params = controller.trainable_parameter_count()
+    print(json.dumps({
+        "rank": args.rank,
+        "trainable_parameters": trainable_params,
+        "zero_init_means_initial_eval_matches_base_selected_region": True,
+    }), flush=True)
+
+    total_batches = math.ceil(len(loader) * args.epochs)
+    if args.max_steps is not None:
+        total_batches = min(total_batches, args.max_steps * args.grad_accum)
+    total_steps = max(math.ceil(total_batches / args.grad_accum), 1)
+    warmup_steps = max(int(total_steps * args.warmup_ratio), 0)
+    optimizer = torch.optim.AdamW(
+        [p for p in controller.parameters() if p.requires_grad],
+        lr=args.lr,
+        weight_decay=args.weight_decay,
+    )
+    scheduler = get_cosine_schedule_with_warmup(
+        optimizer,
+        num_warmup_steps=warmup_steps,
+        num_training_steps=total_steps,
+    )
+
+    summary = {
+        "issue": 33,
+        "script": Path(__file__).name,
+        "args": vars(args),
+        "n_rows": len(dataset),
+        "dropped_rows": dataset.dropped,
+        "n_layers": n_layers,
+        "hidden_size": hidden_size,
+        "d_ffn": d_ffn,
+        "mask_kept": kept,
+        "rank": args.rank,
+        "trainable_parameters": trainable_params,
+        "total_steps": total_steps,
+        "warmup_steps": warmup_steps,
+        "logs": [],
+        "method": {
+            "student": "selected native MLP slices",
+            "selected_slice_edits": ["gate_proj rows", "up_proj rows", "down_proj columns"],
+            "nonselected_mlp_channels": "mean ablated at down_proj input",
+            "teacher_logits": "full unmasked base model",
+            "global_adapter": False,
+            "peft_lora": False,
+            "initial_state": "base selected-region masked circuit plus zero low-rank slice edits",
+        },
+    }
+    (out_dir / "config.json").write_text(json.dumps(summary, indent=2, ensure_ascii=False) + "\n")
+
+    start = time.time()
+    global_step = 0
+    seen_batches = 0
+    running = {"loss": 0.0, "masked_kl": 0.0, "ce": 0.0, "n": 0}
+    optimizer.zero_grad(set_to_none=True)
+
+    while global_step < total_steps:
+        for raw_batch in loader:
+            if global_step >= total_steps:
+                break
+            batch = move_batch(raw_batch, args.device)
+
+            with torch.no_grad():
+                teacher_logits = model(
+                    input_ids=batch["input_ids"],
+                    attention_mask=batch["attention_mask"],
+                    use_cache=False,
+                ).logits
+
+            hooks = controller.install(model)
+            try:
+                student_logits = model(
+                    input_ids=batch["input_ids"],
+                    attention_mask=batch["attention_mask"],
+                    use_cache=False,
+                ).logits
+            finally:
+                for hook in hooks:
+                    hook.remove()
+
+            masked_kl = kl_loss(
+                student_logits,
+                teacher_logits,
+                batch["kl_logit_mask"],
+                temperature=args.kl_temperature,
+            )
+            ce = answer_ce_loss(student_logits, batch["labels"])
+            loss = args.masked_kl_beta * masked_kl + args.ce_beta * ce
+            (loss / args.grad_accum).backward()
+
+            seen_batches += 1
+            running["loss"] += float(loss.detach().cpu())
+            running["masked_kl"] += float(masked_kl.detach().cpu())
+            running["ce"] += float(ce.detach().cpu())
+            running["n"] += 1
+            if seen_batches % args.grad_accum != 0:
+                continue
+
+            torch.nn.utils.clip_grad_norm_(controller.parameters(), args.max_grad_norm)
+            optimizer.step()
+            scheduler.step()
+            optimizer.zero_grad(set_to_none=True)
+            global_step += 1
+
+            if global_step == 1 or global_step % args.eval_every == 0 or global_step == total_steps:
+                denom = max(running["n"], 1)
+                row = {
+                    "step": global_step,
+                    "loss": running["loss"] / denom,
+                    "masked_kl": running["masked_kl"] / denom,
+                    "ce": running["ce"] / denom,
+                    "lr": scheduler.get_last_lr()[0],
+                    "elapsed_s": time.time() - start,
+                }
+                summary["logs"].append(row)
+                (out_dir / "train_summary.json").write_text(
+                    json.dumps(summary, indent=2, ensure_ascii=False) + "\n"
+                )
+                print(json.dumps(row), flush=True)
+                running = {"loss": 0.0, "masked_kl": 0.0, "ce": 0.0, "n": 0}
+
+    controller.eval()
+    summary["elapsed_s"] = time.time() - start
+    save_native_region_delta(
+        out_dir,
+        controller,
+        mask_path=str(args.fixed_mask),
+        means_path=str(means_path),
+        extra={
+            "model": args.model,
+            "target_language": args.target_language,
+            "prompt_style": args.prompt_style,
+            "mask_kept": kept,
+            "n_layers": n_layers,
+            "d_ffn": d_ffn,
+        },
+    )
+    (out_dir / "train_summary.json").write_text(
+        json.dumps(summary, indent=2, ensure_ascii=False) + "\n"
+    )
+    tokenizer.save_pretrained(out_dir / "tokenizer")
+    print(json.dumps({"done": True, "out_dir": str(out_dir), "elapsed_s": summary["elapsed_s"]}), flush=True)
+
+
+if __name__ == "__main__":
+    main()

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/scripts/translation_native_region.py

@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+"""Native selected-MLP-region student helpers for EN->PT experiments.
+
+This path starts from the live base selected-region circuit, not from a fresh
+replacement writer. Non-selected MLP intermediate channels are mean-ablated at
+the down_proj input. Selected channels stay live and receive trainable low-rank
+slice edits to gate_proj rows, up_proj rows, and down_proj columns.
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+import torch
+from torch import nn
+
+from translation_region_student import decoder_root
+
+
+class LowRankLinearDelta(nn.Module):
+    """Zero-initialized low-rank additive map."""
+
+    def __init__(self, in_features: int, out_features: int, rank: int):
+        super().__init__()
+        self.down = nn.Linear(in_features, rank, bias=False)
+        self.up = nn.Linear(rank, out_features, bias=False)
+        nn.init.normal_(self.down.weight, mean=0.0, std=0.02)
+        nn.init.zeros_(self.up.weight)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.up(self.down(x.to(torch.float32)))
+
+
+class NativeRegionDeltaController(nn.Module):
+    """Low-rank edits constrained to selected native MLP slices."""
+
+    def __init__(self, *, mask: dict[int, torch.Tensor], means: dict[int, torch.Tensor],
+                 hidden_size: int, rank: int):
+        super().__init__()
+        self.mask = {int(k): v.detach().cpu().bool() for k, v in mask.items()}
+        self.means = {int(k): v.detach().cpu() for k, v in means.items()}
+        self.hidden_size = int(hidden_size)
+        self.rank = int(rank)
+        self.selected_indices: dict[int, torch.Tensor] = {}
+        self.gate_deltas = nn.ModuleDict()
+        self.up_deltas = nn.ModuleDict()
+        self.down_deltas = nn.ModuleDict()
+
+        for layer_idx, layer_mask in sorted(self.mask.items()):
+            selected = torch.nonzero(layer_mask, as_tuple=False).flatten().cpu()
+            self.selected_indices[layer_idx] = selected
+            if selected.numel() == 0:
+                continue
+            key = str(layer_idx)
+            width = int(selected.numel())
+            self.gate_deltas[key] = LowRankLinearDelta(self.hidden_size, width, self.rank)
+            self.up_deltas[key] = LowRankLinearDelta(self.hidden_size, width, self.rank)
+            self.down_deltas[key] = LowRankLinearDelta(width, self.hidden_size, self.rank)
+
+    def trainable_parameter_count(self) -> int:
+        return sum(p.numel() for p in self.parameters() if p.requires_grad)
+
+    def install(self, model):
+        hooks = []
+        root = decoder_root(model)
+        selected_act_cache: dict[int, torch.Tensor] = {}
+
+        for layer_idx, layer in enumerate(root.layers):
+            selected_cpu = self.selected_indices[layer_idx]
+            if selected_cpu.numel() > 0:
+                hooks.append(layer.mlp.gate_proj.register_forward_hook(
+                    self._make_gate_or_up_hook(layer_idx, selected_cpu, self.gate_deltas)
+                ))
+                hooks.append(layer.mlp.up_proj.register_forward_hook(
+                    self._make_gate_or_up_hook(layer_idx, selected_cpu, self.up_deltas)
+                ))
+
+            hooks.append(layer.mlp.down_proj.register_forward_pre_hook(
+                self._make_down_pre_hook(layer_idx, selected_cpu, selected_act_cache)
+            ))
+            if selected_cpu.numel() > 0:
+                hooks.append(layer.mlp.down_proj.register_forward_hook(
+                    self._make_down_hook(layer_idx, selected_act_cache)
+                ))
+        return hooks
+
+    def _make_gate_or_up_hook(self, layer_idx: int, selected_cpu: torch.Tensor,
+                              modules: nn.ModuleDict):
+        def hook_fn(module, hook_args, output):
+            hidden = hook_args[0]
+            selected = selected_cpu.to(output.device)
+            delta = modules[str(layer_idx)](hidden).to(dtype=output.dtype)
+            patched = output.clone()
+            patched[..., selected] = patched[..., selected] + delta
+            return patched
+        return hook_fn
+
+    def _make_down_pre_hook(self, layer_idx: int, selected_cpu: torch.Tensor,
+                            selected_act_cache: dict[int, torch.Tensor]):
+        def hook_fn(module, hook_args):
+            act = hook_args[0]
+            mean = self.means[layer_idx].to(device=act.device, dtype=act.dtype).view(1, 1, -1)
+            keep = self.mask[layer_idx].to(device=act.device, dtype=act.dtype).view(1, 1, -1)
+            patched = act * keep + mean * (1.0 - keep)
+            if selected_cpu.numel() > 0:
+                selected = selected_cpu.to(act.device)
+                selected_act_cache[layer_idx] = act[..., selected]
+            return (patched,) + hook_args[1:]
+        return hook_fn
+
+    def _make_down_hook(self, layer_idx: int, selected_act_cache: dict[int, torch.Tensor]):
+        def hook_fn(module, hook_args, output):
+            selected_act = selected_act_cache.pop(layer_idx, None)
+            if selected_act is None:
+                return output
+            delta = self.down_deltas[str(layer_idx)](selected_act).to(dtype=output.dtype)
+            return output + delta
+        return hook_fn
+
+
+def save_native_region_delta(out_dir: str | Path, controller: NativeRegionDeltaController,
+                             *, mask_path: str, means_path: str,
+                             extra: dict[str, Any]) -> None:
+    out = Path(out_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    torch.save(controller.state_dict(), out / "native_region_delta.pt")
+    config = {
+        "rank": controller.rank,
+        "hidden_size": controller.hidden_size,
+        "mask_path": mask_path,
+        "means_path": means_path,
+        "trainable_parameters": controller.trainable_parameter_count(),
+        "parameterization": "selected native MLP gate/up/down low-rank slice edits",
+        **extra,
+    }
+    with (out / "native_region_delta_config.json").open("w") as f:
+        json.dump(config, f, indent=2, ensure_ascii=False)
+
+
+def load_native_region_delta(student_dir: str | Path, *, mask: dict[int, torch.Tensor],
+                             means: dict[int, torch.Tensor], hidden_size: int,
+                             map_location: str = "cpu") -> NativeRegionDeltaController:
+    root = Path(student_dir)
+    cfg = json.loads((root / "native_region_delta_config.json").read_text())
+    controller = NativeRegionDeltaController(
+        mask=mask,
+        means=means,
+        hidden_size=hidden_size,
+        rank=int(cfg["rank"]),
+    )
+    state = torch.load(root / "native_region_delta.pt", map_location=map_location)
+    controller.load_state_dict(state)
+    return controller

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/spec/native_region_mvc_issue33.json

@@ -0,0 +1,58 @@
+{
+  "issue": 33,
+  "name": "native_selected_region_student_enpt",
+  "objective": "Improve the selected EN->PT MLP region itself, starting from its base masked-region recovery state, while keeping the rest of the model frozen and non-selected MLP channels mean-ablated.",
+  "model": {
+    "base_and_teacher": "tencent/HY-MT1.5-1.8B",
+    "teacher_for_training": "full unmasked base logits on FLORES teacher-hyp rows",
+    "comparison_reference": "best trained reference from issue #28 r8; comparison only, not the teacher"
+  },
+  "data": {
+    "train_source": "FLORES EN->PT devtest JSONL translated by unmasked HY-MT into model_hyp targets",
+    "eval_source": "NTREX EN->PT held-out JSONL",
+    "target_field": "model_hyp",
+    "judge": "XCOMET-XXL service when available"
+  },
+  "mask": {
+    "source": "preserved ReLP base attribution mask from Occupying-Mars/hy-lora-conditions",
+    "budget_mlp_channels": 160000,
+    "path_in_artifacts": "masks/base_attr/relp_k160000.full.npz"
+  },
+  "student": {
+    "initial_state": "base selected-region masked circuit with zero low-rank slice edits",
+    "trainable_region": "selected native MLP slices only",
+    "trainable_slices": [
+      "gate_proj selected output rows",
+      "up_proj selected output rows",
+      "down_proj selected input columns"
+    ],
+    "nonselected_mlp_channels": "mean ablated at down_proj input during masked objective and evaluation",
+    "outside_region": "frozen base model carrier",
+    "global_adapter": false,
+    "peft_lora": false,
+    "first_rank": 8
+  },
+  "loss": {
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "kl_temperature": 1.0,
+    "kl_on": "answer"
+  },
+  "anchors": {
+    "base_selected_region_xcomet": 0.6278288431,
+    "best_trained_reference_issue28_r8_xcomet": 0.8145537329,
+    "continue_if_native_student_beats_base_selected_region_by_xcomet": 0.03
+  },
+  "run": {
+    "pod_class": "2x NVIDIA RTX PRO 6000 Blackwell Server Edition",
+    "runner": "issue33_native_region_runner.sh",
+    "artifact_dataset_repo": "TokenBender/synth-data-en-pt-circuit",
+    "artifact_backup_repo": "TokenBender/circuit-discovery/circuit-shotting/artifacts/issue33"
+  },
+  "stop_conditions": [
+    "fixed mask or preserved r8 artifacts cannot be restored",
+    "base selected-region anchor is not reproduced closely enough to trust the run",
+    "native rank-8 student does not beat the selected-region baseline by at least 0.03 XCOMET",
+    "artifact upload fails after a completed run"
+  ]
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/summaries/issue33_native_region_k160000_r8.json

@@ -0,0 +1,32 @@
+{
+  "issue": 33,
+  "run_root": "/root/runs/issue33_native_region",
+  "k": 160000,
+  "rank": 8,
+  "rows": [
+    {
+      "condition": "base_selected_region_k160000",
+      "role": "initial selected-region baseline",
+      "xcomet": 0.6259405999205152,
+      "reference_xcomet": 0.6278288431,
+      "chrFpp": 46.99461052474228
+    },
+    {
+      "condition": "best_trained_reference_issue28_r8",
+      "role": "comparison reference, not teacher",
+      "xcomet": 0.8103810173308426,
+      "reference_xcomet": 0.8145537329,
+      "chrFpp": 51.21928365554815
+    },
+    {
+      "condition": "native_region_k160000_r8",
+      "role": "issue33 trained native selected-region student",
+      "xcomet": 0.7950550597565977,
+      "reference_xcomet": null,
+      "chrFpp": 50.97431297760433
+    }
+  ],
+  "native_delta_vs_base_selected": 0.16911445983608253,
+  "native_delta_vs_best_trained_reference": -0.01532595757424482,
+  "pass_floor_for_continuation": true
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/summaries/issue33_native_region_k160000_r8.md

@@ -0,0 +1,15 @@
+# Issue 33 Native Selected-Region Student Summary
+
+- Run root: `/root/runs/issue33_native_region`
+- Budget: `k=160000`, rank `8`
+- Teacher for training: full unmasked base logits.
+- Best trained reference from issue #28 r8 is comparison only.
+
+| condition | role | XCOMET | reference XCOMET | chrF++ |
+|---|---|---:|---:|---:|
+| `base_selected_region_k160000` | initial selected-region baseline | 0.625941 | 0.627829 | 46.994611 |
+| `best_trained_reference_issue28_r8` | comparison reference, not teacher | 0.810381 | 0.814554 | 51.219284 |
+| `native_region_k160000_r8` | issue33 trained native selected-region student | 0.795055 |  | 50.974313 |
+
+Native delta vs base selected region: `0.169114`.
+Continuation floor (+0.03 XCOMET over base selected region): `True`.

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/xcomet/base_fixed_k160000.json

@@ -0,0 +1,25 @@
+{
+  "hyps_jsonl": "/root/runs/issue33_native_region/dumps/base_k160000/fixed_k160000.jsonl",
+  "out_jsonl": "/root/runs/issue33_native_region/xcomet/base_fixed_k160000.scored_pool.jsonl",
+  "comet_model": "Unbabel/XCOMET-XXL",
+  "checkpoint_path": "/root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt",
+  "n": 1012,
+  "threshold": 0.99,
+  "system_score": 0.6259405999205152,
+  "passed": 38,
+  "elapsed_seconds": 339.8826160430908,
+  "batch_size": 8,
+  "chunk_size": 128,
+  "by_category_tag": [
+    {
+      "category": "ntrex_test",
+      "tag": "heldout",
+      "n": 1012,
+      "mean": 0.6259405999205152,
+      "passed": 38
+    }
+  ],
+  "input_categories": {
+    "ntrex_test": 1012
+  }
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/xcomet/native_region_k160000_r8.json

@@ -0,0 +1,25 @@
+{
+  "hyps_jsonl": "/root/runs/issue33_native_region/dumps/native_region_k160000_r8.jsonl",
+  "out_jsonl": "/root/runs/issue33_native_region/xcomet/native_region_k160000_r8.scored_pool.jsonl",
+  "comet_model": "Unbabel/XCOMET-XXL",
+  "checkpoint_path": "/root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt",
+  "n": 1012,
+  "threshold": 0.99,
+  "system_score": 0.7950550597565977,
+  "passed": 162,
+  "elapsed_seconds": 332.72063279151917,
+  "batch_size": 8,
+  "chunk_size": 128,
+  "by_category_tag": [
+    {
+      "category": "ntrex_test",
+      "tag": "heldout",
+      "n": 1012,
+      "mean": 0.7950550597565977,
+      "passed": 162
+    }
+  ],
+  "input_categories": {
+    "ntrex_test": 1012
+  }
+}

circuit-shotting/artifacts/issue33/issue33_native_region_20260513T171458Z/xcomet/r8_fixed_k160000.json

@@ -0,0 +1,25 @@
+{
+  "hyps_jsonl": "/root/runs/issue33_native_region/dumps/r8_k160000/fixed_k160000.jsonl",
+  "out_jsonl": "/root/runs/issue33_native_region/xcomet/r8_fixed_k160000.scored_pool.jsonl",
+  "comet_model": "Unbabel/XCOMET-XXL",
+  "checkpoint_path": "/root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt",
+  "n": 1012,
+  "threshold": 0.99,
+  "system_score": 0.8103810173308426,
+  "passed": 161,
+  "elapsed_seconds": 340.4577021598816,
+  "batch_size": 8,
+  "chunk_size": 128,
+  "by_category_tag": [
+    {
+      "category": "ntrex_test",
+      "tag": "heldout",
+      "n": 1012,
+      "mean": 0.8103810173308426,
+      "passed": 161
+    }
+  ],
+  "input_categories": {
+    "ntrex_test": 1012
+  }
+}