import arm qwen3/head_k16_solo_o0_lamA0.5: anchors (every close), stats, results, log, readout

import/qwen3/head_k16_solo_o0_lamA0.5/manifest.json

@@ -13,6 +13,6 @@
  "base": "solo",
  "steps": 2000,
  "trunk_step": 245674,
- "shipped": "2026-10-08 23:21 local",
+ "shipped": "2026-10-08 23:23 local",
  "note": "every close's anchor ships (mids included); the final close's anchor is the run's s15_qwen_import.safetensors"
 }

import/qwen3/head_k16_solo_o0_lamA0.5/readout.md

@@ -0,0 +1,104 @@
+# The import arm's readout: Qwen3 0.6B base's states on her own bytes
+
+An import arm is a detachable arm on Beatrix's frozen trunk trained so that her states at Qwen3's token closings carry more of Qwen3 0.6B base's own states (the ones Anima's text path reads). Every number below is read on the dual-extraction read's held-out captions (512 of each COCO draw, never trained on; draw 2 is the repeat), one caption a row, at every token's closing byte, as whitened Procrustes alignment against Qwen's states (1 = the same geometry up to a rotation; two folds by caption). The pre-registered gates are in the plan; I3-I5 (the mood axis, the relay's cell, the pictures) are the instrument's read.
+
+## The runs
+
+- qimp_head_k16_solo_o0_lamA0.5: the trained-head form (a linear read head per block into Qwen's frame, trained with the arm; per-site MSE in Qwen's frame); the teacher Qwen3 0.6B base at depth 16; alone on the bare trunk; quiet lambda 2.0 off-domain, lambda_a 0.5 on the caption rows; blocks 0-31; 2000 of 2000 steps; seed offset 0
+
+## The gates (the last close of each run)
+
+### qimp_head_k16_solo_o0_lamA0.5
+
+- I1 the reach: NOT MET (the largest rise of her armed reading over her bare reading against Qwen's final state +0.038 at block 30 on draw 1, +0.036 at block 31 on draw 2; bar +0.1; the armed peak 0.572 at block 26; against Qwen's depth-16 state the rise is +0.085 at block 30)
+- I2 the context: NOT MET (with the token means removed: +0.031 at block 30 on draw 1, +0.032 at block 30 on draw 2; bar +0.1)
+- the costs in bits: MET (the held captions -0.03972 bpb, web +0.00013; bar +0.012 each)
+
+## The birth checks
+
+### qimp_head_k16_solo_o0_lamA0.5
+
+- the instrument check: her bare reading against Qwen's final state (this driver / e038's row): block 8 0.426 / 0.426, block 12 0.470 / 0.470, block 16 0.508 / 0.509, block 20 0.526 / 0.526, block 24 0.562 / 0.562, block 28 0.540 / 0.540; the largest difference 0.001
+- her bare reading against Qwen's depth-16 state: block 8 0.496, block 12 0.548, block 16 0.579, block 18 0.581, block 20 0.573, block 24 0.526, block 28 0.452, block 31 0.385
+- the heads at birth, held-out R2 per block (the linear ceiling of the form): 8: 0.351, 12: 0.397, 16: 0.432, 20: 0.458, 24: 0.438, 28: 0.385, 31: 0.334
+- the birth gates: web rows with the fresh arm 0.0e+00 bpb from before; the hand-run trunk against the model's forward 0.0e+00 armed / 0.0e+00 masked; 13,911 reference sites for the statistics
+
+## The closes
+
+### qimp_head_k16_solo_o0_lamA0.5
+
+| close | step | I1 the reach (draw 1 / 2) | at block | I2 the context (draw 1 / 2) | armed peak vs final (draw 1) | bits captions / web | gate mean / widest |
+|---|---|---|---|---|---|---|---|
+| 0 | 0 | +0.000 / +0.000 | 31 / 31 | +0.000 / +0.000 | 0.562 at 24 | +0.00000 / +0.00000 | 0.047 / 0.047 |
+| 1 | 1000 | +0.024 / +0.025 | 31 / 31 | +0.023 / +0.021 | 0.572 at 25 | -0.03536 / +0.00001 | 0.197 / 0.267 |
+| 2 | 2000 | +0.038 / +0.036 | 30 / 31 | +0.031 / +0.032 | 0.572 at 26 | -0.03972 / +0.00013 | 0.462 / 0.663 |
+
+## Every block at the last close
+
+### qimp_head_k16_solo_o0_lamA0.5, draw1 (512 captions, 4,920 sites), close 2 at step 2000
+
+| block | bare vs final | armed vs final | bare vs depth 16 | armed vs depth 16 | context bare | context armed | mse | cos |
+|---|---|---|---|---|---|---|---|---|
+| 0 | 0.295 | 0.295 | 0.327 | 0.327 | 0.121 | 0.121 | 1.032 | 0.341 |
+| 4 | 0.388 | 0.389 | 0.443 | 0.447 | 0.229 | 0.235 | 0.894 | 0.486 |
+| 8 | 0.426 | 0.430 | 0.496 | 0.505 | 0.285 | 0.295 | 0.819 | 0.547 |
+| 12 | 0.470 | 0.472 | 0.548 | 0.559 | 0.331 | 0.338 | 0.766 | 0.588 |
+| 16 | 0.508 | 0.510 | 0.579 | 0.601 | 0.372 | 0.376 | 0.716 | 0.624 |
+| 18 | 0.514 | 0.517 | 0.581 | 0.606 | 0.386 | 0.387 | 0.691 | 0.641 |
+| 20 | 0.526 | 0.530 | 0.573 | 0.606 | 0.399 | 0.403 | 0.681 | 0.648 |
+| 22 | 0.549 | 0.552 | 0.550 | 0.591 | 0.441 | 0.443 | 0.683 | 0.647 |
+| 24 | 0.562 | 0.566 | 0.526 | 0.579 | 0.466 | 0.468 | 0.690 | 0.641 |
+| 26 | 0.556 | 0.572 | 0.487 | 0.558 | 0.464 | 0.476 | 0.708 | 0.629 |
+| 28 | 0.540 | 0.565 | 0.452 | 0.530 | 0.448 | 0.467 | 0.727 | 0.616 |
+| 29 | 0.527 | 0.557 | 0.436 | 0.519 | 0.433 | 0.458 | 0.736 | 0.609 |
+| 30 | 0.494 | 0.532 | 0.402 | 0.486 | 0.399 | 0.430 | 0.759 | 0.596 |
+| 31 | 0.475 | 0.513 | 0.385 | 0.466 | 0.385 | 0.414 | 0.767 | 0.591 |
+
+### qimp_head_k16_solo_o0_lamA0.5, draw2 (512 captions, 5,086 sites), close 2 at step 2000
+
+| block | bare vs final | armed vs final | bare vs depth 16 | armed vs depth 16 | context bare | context armed | mse | cos |
+|---|---|---|---|---|---|---|---|---|
+| 0 | 0.295 | 0.295 | 0.331 | 0.331 | 0.143 | 0.143 | 1.007 | 0.345 |
+| 4 | 0.392 | 0.395 | 0.450 | 0.453 | 0.242 | 0.247 | 0.873 | 0.488 |
+| 8 | 0.431 | 0.434 | 0.502 | 0.512 | 0.300 | 0.309 | 0.799 | 0.548 |
+| 12 | 0.471 | 0.474 | 0.555 | 0.566 | 0.344 | 0.353 | 0.746 | 0.589 |
+| 16 | 0.508 | 0.510 | 0.583 | 0.603 | 0.381 | 0.385 | 0.697 | 0.626 |
+| 18 | 0.513 | 0.517 | 0.583 | 0.610 | 0.397 | 0.402 | 0.671 | 0.644 |
+| 20 | 0.529 | 0.533 | 0.577 | 0.609 | 0.413 | 0.419 | 0.660 | 0.651 |
+| 22 | 0.551 | 0.557 | 0.558 | 0.597 | 0.453 | 0.458 | 0.661 | 0.652 |
+| 24 | 0.567 | 0.571 | 0.534 | 0.586 | 0.474 | 0.478 | 0.669 | 0.645 |
+| 26 | 0.564 | 0.579 | 0.497 | 0.564 | 0.472 | 0.486 | 0.687 | 0.633 |
+| 28 | 0.546 | 0.570 | 0.466 | 0.541 | 0.458 | 0.481 | 0.707 | 0.619 |
+| 29 | 0.535 | 0.562 | 0.454 | 0.531 | 0.447 | 0.470 | 0.716 | 0.613 |
+| 30 | 0.503 | 0.536 | 0.420 | 0.499 | 0.415 | 0.447 | 0.739 | 0.599 |
+| 31 | 0.476 | 0.511 | 0.399 | 0.474 | 0.399 | 0.429 | 0.748 | 0.594 |
+
+The member's share in bits per byte (all on minus the member masked): web +0.00013, captions -0.03972
+
+## The training curves
+
+### qimp_head_k16_solo_o0_lamA0.5
+
+| step | task loss | mse | cos | InfoNCE | quiet on the caption rows | quiet off-domain |
+|---|---|---|---|---|---|---|
+| 1 | 0.6398 | 0.640 | 0.603 | None | 0.00000 | 0.00000 |
+| 50 | 0.6273 | 0.627 | 0.610 | None | 0.00026 | 0.00021 |
+| 250 | 0.6036 | 0.604 | 0.628 | None | 0.00063 | 0.00027 |
+| 500 | 0.5969 | 0.597 | 0.632 | None | 0.00314 | 0.00027 |
+| 1000 | 0.5965 | 0.597 | 0.636 | None | 0.01011 | 0.00034 |
+| 1500 | 0.5849 | 0.585 | 0.644 | None | 0.01594 | 0.00041 |
+| 2000 | 0.5689 | 0.569 | 0.656 | None | 0.02013 | 0.00045 |
+| 1 | 0.6398 | 0.640 | 0.603 | None | 0.00000 | 0.00000 |
+| 50 | 0.6273 | 0.627 | 0.610 | None | 0.00026 | 0.00021 |
+| 250 | 0.6036 | 0.604 | 0.628 | None | 0.00063 | 0.00027 |
+| 500 | 0.5969 | 0.597 | 0.632 | None | 0.00314 | 0.00027 |
+| 1000 | 0.5965 | 0.597 | 0.636 | None | 0.01011 | 0.00034 |
+| 1500 | 0.5849 | 0.585 | 0.644 | None | 0.01594 | 0.00041 |
+| 2000 | 0.5689 | 0.569 | 0.656 | None | 0.02013 | 0.00045 |
+
+## The recipe
+
+- per step: her armed model reads the caption rows (her own bytes): at each Qwen3 token's closing byte her standardized LayerNorm'd block output through a trained linear head matches Qwen3 0.6B base's standardized depth-16 state, per-site MSE averaged over the served blocks; quiet off-domain: lambda x KL(member masked || armed) on one web row + one partner row; quiet on the caption rows: lambda_a x KL(teacher || armed)
+- lambda 2.0 off-domain, lambda_a 0.5 on the caption rows; pure Adam lr 0.001 wd 0; 2 x 4096-byte rows a step; the arm LAWFUL_16x8 at every block ({'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}); the trunk mini-beatrix-3 at step 245,674; alephllm 0.10.7, amoe-lora 0.2.11, torch 2.8.0+cu128
+- the teacher: Qwen3 0.6B base from Anima's text-encoder file (qwen_3_06b_base.safetensors, sha256 cd2a512003e2f9f3...), its residual at depth 16; the sites: every Qwen3 token's closing byte on her own bytes (tokenizer.json sha256 be75606093db2094..., vocab 151,669); documents = one caption-like field of a caption-pack row (caption, short, long, tags, medium), sources drawn by {'dff': 0.161, 'scf': 0.161, 'danbooru': 0.164, 'cc12m': 0.406, 'coco': 0.108}
+- the test set: the first 512 captions of each draw of coco_caps_2x2048.json (every caption of the file kept out of training); the instrument's reference row (e038): {'8': 0.426, '12': 0.47, '16': 0.509, '20': 0.526, '24': 0.562, '28': 0.54}