btx-e003 2s ladder C5/C4/C6 ledger

mini-beatrix-2s/reports/arms/btx_e003_C456_ledger.json

@@ -34384,7 +34384,8 @@
     "detach_verify": true,
     "peak_vram_gb": 5.26,
     "finished": "2026-09-20 01:52"
-   }
+   },
+   "error": "tools\\btx_2s_ladder_e_i.py\", line 549, in <module>\n    train_frame_member(FR, f\"btx-e003-2s-EI-frame-{S}\", {\"chain\": d5_abst, \"stop\": chat_abst, \"neutral\": fw_abst}, DOM)\n  File \"E:\\mirel\\claude\\history\\dockets\\2026-09-05_bytelex_plan\\tools\\btx_2s_ladder_e_i.py\", line 342, in train_frame_member\n    la = LAMBDA * kl_abstain(m, handle, rows[int(rng.integers(0, len(rows)))]); (la / 2).backward(); ta += float(la) / 2\n                                                                                ^^^^^^^^^^^^^^^^^^^\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\_tensor.py\", line 631, in backward\n    torch.autograd.backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\__init__.py\", line 381, in backward\n    _engine_run_backward(\n  File \"E:\\mirel\\claude\\.venv\\Lib\\site-packages\\torch\\autograd\\graph.py\", line 869, in _engine_run_backward\n    return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\ntorch.AcceleratorError: CUDA error: unknown error\nSearch for `cudaErrorUnknown' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.\nCUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.\nFor debugging consider passing CUDA_LAUNCH_BLOCKING=1\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\n\n"
   },
   "E-D0": {
    "started": "2026-09-10 11:28",
@@ -63062,7 +63063,118 @@
      "c4b_no_abstention_masked_chain": 0.013
     }
    },
-   "finished": "2026-09-16 22:23"
+   "finished": "2026-09-16 22:23",
+   "started-lawful": "2026-09-20 02:08",
+   "ECO-staged-sL0": {
+    "arm": "ECO-staged-sL0",
+    "cell": "ECO",
+    "form": "staged co-training (item 10)",
+    "torch_seed": [
+     1849955378,
+     2679450906
+    ],
+    "seed_key": "btx-e003-2s-ECO-staged-sL0-chain",
+    "description": "STAGED: chain-sel-sL0 LOADED trainable (inner) + fresh quiet stop arm (outer) under one Adam; per-member abstention with the partner AS IS",
+    "recipe": "pure Adam wd0 lr0.001 1600 steps 2x4 (even chunks: chain mbR+end + stop CE; odd: per-member KL(bare||armed) lambda 1.0, bare = THIS member masked partner AS IS; chain off-domain = chat+fineweb, stop off-domain = d5+fineweb)",
+    "lambda": 1.0,
+    "train_mode": {
+     "batch": 2,
+     "accum": 4,
+     "hub_ckpt": 0,
+     "bf16_replication": false
+    },
+    "inner_anchor": "btx-e003-2s-chain-sel-sL0.b2a4@step61422.anchor.pt",
+    "stop_corpus": {
+     "rows_total": 7884,
+     "rows_kept": 4818,
+     "dropped_gt_512B": 3066,
+     "max_row_bytes_kept": 512,
+     "source": "geolip.alephllm.chat_sft.load_chat_corpus(context 2048, max_row_bytes 900) + [10,10] (overnight_arms.corpus_stop, the stopnl-s1600 recipe); rows > 512 B dropped = this runner's VRAM rule (FLAGGED deviation from the 900-byte cap)"
+    },
+    "trunk_sha256_before": "95a4606ff34dbcc0d907540303796bf080010ca8708c013bd18ac9b34107ecf5",
+    "trainable_params": 17116200,
+    "pre_train_chain_read": {
+     "n": 150,
+     "strict": 0.84,
+     "chain": 0.84,
+     "final": 0.9667,
+     "echo_rate": 0.0,
+     "loop_rate": 0.0,
+     "stop_rate": 1.0,
+     "termination_pass": 0.9733,
+     "n_answers": 4.913,
+     "stop_strings": {
+      "\n\n": 150
+     },
+     "max_new": 240,
+     "sample": "So Wren is quen. So Wren is harl. So Wren is torv. So Wren is prin. So Wren is mund.",
+     "gauge_s": 57.8,
+     "frame": "text",
+     "template": "qa newline pair: CHAT_HEADER + 'User: ' + q + '\\n' + 'Beatrix: ' + reply + '\\n\\n' (amoe_bridge.py:30-34)",
+     "stops": [
+      "\nUser:",
+      "\nBeatrix:",
+      "User:",
+      "Beatrix:",
+      "\n\n"
+     ],
+     "turn_end": "newline pair [10,10]"
+    },
+    "gate_apparatus": {
+     "passed": true,
+     "first": 0.025996,
+     "last": 6e-06,
+     "rel_drop": 0.9998
+    },
+    "gate_own_target": {
+     "passed": true,
+     "first": 0.356784,
+     "last": 0.000979,
+     "rel_drop": 0.9973,
+     "mode": "batch 2 x accum 4",
+     "steps": 50,
+     "lr": 0.01,
+     "min_rel_drop": 0.2,
+     "both_masked_reference_loss": 0.6632698960602283,
+     "step1_loss": 0.3567843196215108
+    },
+    "train_s": 6542,
+    "s_per_step": 4.089,
+    "curve": [
+     [
+      250,
+      0.09509
+     ],
+     [
+      500,
+      0.04158
+     ],
+     [
+      750,
+      0.06945
+     ],
+     [
+      1000,
+      0.08083
+     ],
+     [
+      1250,
+      0.06814
+     ],
+     [
+      1500,
+      0.0954
+     ],
+     [
+      1600,
+      0.07697
+     ]
+    ],
+    "peak_vram_train_gb": 6.34,
+    "anchor": "E:\\mirel\\data\\btx_e003\\arms_2s\\btx-e003-2s-ECO-staged-sL0-chain.b2a4@step61422.anchor.pt",
+    "anchor_stop": "E:\\mirel\\data\\btx_e003\\arms_2s\\btx-e003-2s-ECO-staged-sL0-stop.b2a4@step61422.anchor.pt",
+    "provisional": true
+   }
   }
  },
  "_env": {