diff --git a/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/protocol.yaml b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/protocol.yaml
new file mode 100644
index 0000000..197d7c7
--- /dev/null
+++ b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/protocol.yaml
@@ -0,0 +1,145 @@
+protocol:
+  name: nanogpt_four_head_2026_08_27_ww_baseline
+  version: 1
+  description: Four-corpus-equivalent-epoch AdamW and MuonClip baselines with five paired seeds, four attention heads, and WeightWatcher raw and clip_xmax monitoring.
+
+dataset:
+  name: HuggingFaceFW/fineweb-edu
+  config: sample-10BT
+  split: train
+  revision: 593b3a867298afb8ce42625a270ef20ddcad28f9
+  tokenizer: gpt2
+  train_tokens: 80000000
+  val_tokens: 1000000
+  test_tokens: 1000000
+
+model:
+  vocab_size: 50257
+  block_size: 256
+  n_layer: 1
+  n_head: 4
+  n_embd: 128
+  dropout: 0.0
+  bias: false
+  tie_weights: true
+
+training:
+  seeds: [1337, 2027, 4099, 31415, 271828]
+  batch_size: 4
+  grad_accum_steps: 8
+  target_epochs: 4.0
+  epoch_interval: 0.25
+  eval_interval_steps: 500
+  eval_batches: 64
+  checkpoint_interval_steps: 500
+  grad_clip: 1.0
+
+optimizer_profiles:
+  sgd_momentum:
+    display_name: SGD + Nesterov momentum (not a campaign arm)
+    family: sgd
+    learning_rate: 0.05
+    min_learning_rate: 0.005
+    warmup_fraction: 0.10
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.90
+    dampening: 0.0
+    nesterov: true
+    weight_decay: 0.01
+
+  adamw:
+    display_name: AdamW
+    family: adamw
+    learning_rate: 0.0006
+    min_learning_rate: 0.00006
+    warmup_fraction: 0.01
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    weight_decay: 0.10
+
+  adam:
+    display_name: Adam (not a campaign arm)
+    family: adam
+    learning_rate: 0.0006
+    min_learning_rate: 0.00006
+    warmup_fraction: 0.01
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    weight_decay: 0.0
+
+  muon:
+    display_name: Muon + auxiliary AdamW (not a campaign arm)
+    family: muon
+    matrix_learning_rate: 0.02
+    matrix_min_learning_rate: 0.002
+    aux_learning_rate: 0.0003
+    aux_min_learning_rate: 0.00003
+    warmup_fraction: 0.05
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.95
+    nesterov: true
+    newton_schulz_steps: 5
+    muon_epsilon: 1.0e-7
+    matrix_weight_decay: 0.01
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    aux_weight_decay: 0.01
+
+  muon_clip:
+    display_name: MuonClip + auxiliary AdamW
+    family: muon_clip
+    learning_rate: 0.0002
+    min_learning_rate: 0.00002
+    warmup_fraction: 0.0512
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.95
+    nesterov: false
+    newton_schulz_steps: 5
+    muon_epsilon: 1.0e-7
+    weight_decay: 0.10
+    update_rms_scale: 0.20
+    qk_clip_threshold: 100.0
+    qk_clip_balance: 0.50
+    qk_diagnostics_interval: 500
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+
+evaluation:
+  train_probe_seed: 21001
+  validation_probe_seed: 22001
+  test_probe_seed: 23001
+  bleu_probe_seed: 24001
+  bleu_examples: 64
+  bleu_prompt_tokens: 64
+  bleu_continuation_tokens: 32
+  bleu_batch_size: 4
+
+weightwatcher:
+  enabled: true
+  ERG: true
+  randomize: true
+  strict: true
+  min_evals: 20
+  fix_fingers: clip_xmax
+  max_fingers: 10
+  require_raw_alpha: true
+
+runtime:
+  matmul_precision: highest
+  allow_tf32: false
+  cudnn_benchmark: false
+  mps_fallback: true
+  deterministic_algorithms: true
+  deterministic_warn_only: false
+  empty_mps_cache_after_weightwatcher: true
diff --git a/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/report.html b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/report.html
new file mode 100644
index 0000000..29bc914
--- /dev/null
+++ b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/four_head_5seed/report.html
@@ -0,0 +1,467 @@
+<!doctype html>
+<html>
+<head>
+<meta charset="utf-8">
+<title>Four-head nanoGPT five-seed analysis</title>
+<style>
+body { font-family: sans-serif; max-width: 1400px; margin: 30px auto; }
+img { width: 100%; border: 1px solid #ddd; margin-bottom: 25px; }
+table { border-collapse: collapse; font-size: 12px; }
+th, td { border: 1px solid #ccc; padding: 5px; text-align: right; }
+</style>
+</head>
+<body>
+<h1>Four-head nanoGPT: AdamW versus MuonClip</h1>
+<p>Five paired seeds: 1337, 2027, 4099, 31415, 271828.</p>
+
+<h2>Completion status</h2>
+<table border="1" class="dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>optimizer_label</th>
+      <th>seed</th>
+      <th>complete</th>
+      <th>steps</th>
+      <th>final_test_loss</th>
+      <th>final_test_bleu</th>
+      <th>run_dir</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>AdamW</td>
+      <td>1337</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>6.549376</td>
+      <td>0.247733</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/adamw/seed_1337</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>AdamW</td>
+      <td>2027</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.385771</td>
+      <td>0.690209</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/adamw/seed_2027</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>AdamW</td>
+      <td>4099</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>6.572981</td>
+      <td>0.432271</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/adamw/seed_4099</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>AdamW</td>
+      <td>31415</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.903410</td>
+      <td>0.416945</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/adamw/seed_31415</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>AdamW</td>
+      <td>271828</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.660902</td>
+      <td>0.513238</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/adamw/seed_271828</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip</td>
+      <td>1337</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.856352</td>
+      <td>0.444428</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/muon_clip/seed_1337</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip</td>
+      <td>2027</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.632226</td>
+      <td>0.653022</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/muon_clip/seed_2027</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip</td>
+      <td>4099</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>6.210440</td>
+      <td>0.240927</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/muon_clip/seed_4099</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip</td>
+      <td>31415</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>5.751830</td>
+      <td>0.404720</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/muon_clip/seed_31415</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip</td>
+      <td>271828</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>6.512031</td>
+      <td>0.429166</td>
+      <td>/tmp/rg-nanogpt-four-head-20260827/results/muon_clip/seed_271828</td>
+    </tr>
+  </tbody>
+</table>
+
+<h2>Held-out test summary</h2>
+<table border="1" class="dataframe">
+  <thead>
+    <tr>
+      <th>optimizer</th>
+      <th>checkpoint</th>
+      <th colspan="3" halign="left">test_loss</th>
+      <th colspan="3" halign="left">test_perplexity</th>
+      <th colspan="3" halign="left">test_accuracy</th>
+      <th colspan="3" halign="left">test_top5_accuracy</th>
+      <th colspan="3" halign="left">test_bleu</th>
+      <th colspan="3" halign="left">test_continuation_token_accuracy</th>
+    </tr>
+    <tr>
+      <th></th>
+      <th></th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>count</th>
+      <th>mean</th>
+      <th>std</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>5</td>
+      <td>6.014488</td>
+      <td>0.531664</td>
+      <td>5</td>
+      <td>457.255696</td>
+      <td>234.132277</td>
+      <td>5</td>
+      <td>0.163257</td>
+      <td>0.034601</td>
+      <td>5</td>
+      <td>0.318628</td>
+      <td>0.051642</td>
+      <td>5</td>
+      <td>0.460079</td>
+      <td>0.160865</td>
+      <td>5</td>
+      <td>0.022461</td>
+      <td>0.003347</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>5</td>
+      <td>5.700518</td>
+      <td>0.361662</td>
+      <td>5</td>
+      <td>317.137042</td>
+      <td>136.137852</td>
+      <td>5</td>
+      <td>0.184805</td>
+      <td>0.024026</td>
+      <td>5</td>
+      <td>0.350369</td>
+      <td>0.036404</td>
+      <td>5</td>
+      <td>0.547807</td>
+      <td>0.135450</td>
+      <td>5</td>
+      <td>0.024316</td>
+      <td>0.003425</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>5</td>
+      <td>5.992575</td>
+      <td>0.361824</td>
+      <td>5</td>
+      <td>422.921627</td>
+      <td>162.779180</td>
+      <td>5</td>
+      <td>0.171960</td>
+      <td>0.019319</td>
+      <td>5</td>
+      <td>0.330957</td>
+      <td>0.030304</td>
+      <td>5</td>
+      <td>0.434453</td>
+      <td>0.146830</td>
+      <td>5</td>
+      <td>0.024805</td>
+      <td>0.003866</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>5</td>
+      <td>5.862132</td>
+      <td>0.216198</td>
+      <td>5</td>
+      <td>358.435895</td>
+      <td>83.430356</td>
+      <td>5</td>
+      <td>0.177106</td>
+      <td>0.014019</td>
+      <td>5</td>
+      <td>0.340280</td>
+      <td>0.020654</td>
+      <td>5</td>
+      <td>0.430188</td>
+      <td>0.153888</td>
+      <td>5</td>
+      <td>0.023730</td>
+      <td>0.002643</td>
+    </tr>
+  </tbody>
+</table>
+
+<h2>Paired MuonClip-minus-AdamW differences</h2>
+<table border="1" class="dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>checkpoint</th>
+      <th>metric</th>
+      <th>n</th>
+      <th>mean_muon_clip_minus_adamw</th>
+      <th>sd</th>
+      <th>ci95_lower</th>
+      <th>ci95_upper</th>
+      <th>paired_t</th>
+      <th>paired_p_two_sided</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>final</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>-0.021913</td>
+      <td>0.595284</td>
+      <td>-0.761056</td>
+      <td>0.717230</td>
+      <td>-0.082311</td>
+      <td>0.938354</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>-34.334069</td>
+      <td>282.269870</td>
+      <td>-384.818405</td>
+      <td>316.150267</td>
+      <td>-0.271986</td>
+      <td>0.799095</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>0.008704</td>
+      <td>0.034766</td>
+      <td>-0.034464</td>
+      <td>0.051871</td>
+      <td>0.559803</td>
+      <td>0.605479</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>0.012329</td>
+      <td>0.052791</td>
+      <td>-0.053219</td>
+      <td>0.077877</td>
+      <td>0.522228</td>
+      <td>0.629098</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>-0.025627</td>
+      <td>0.141967</td>
+      <td>-0.201901</td>
+      <td>0.150648</td>
+      <td>-0.403636</td>
+      <td>0.707127</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>0.002344</td>
+      <td>0.006813</td>
+      <td>-0.006116</td>
+      <td>0.010803</td>
+      <td>0.769207</td>
+      <td>0.484674</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>0.161614</td>
+      <td>0.153532</td>
+      <td>-0.029021</td>
+      <td>0.352249</td>
+      <td>2.353777</td>
+      <td>0.078192</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>41.298853</td>
+      <td>56.168592</td>
+      <td>-28.443666</td>
+      <td>111.041372</td>
+      <td>1.644105</td>
+      <td>0.175500</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>-0.007700</td>
+      <td>0.011894</td>
+      <td>-0.022468</td>
+      <td>0.007069</td>
+      <td>-1.447512</td>
+      <td>0.221313</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>-0.010089</td>
+      <td>0.016709</td>
+      <td>-0.030836</td>
+      <td>0.010658</td>
+      <td>-1.350188</td>
+      <td>0.248296</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>-0.117620</td>
+      <td>0.141903</td>
+      <td>-0.293815</td>
+      <td>0.058576</td>
+      <td>-1.853421</td>
+      <td>0.137444</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>-0.000586</td>
+      <td>0.003477</td>
+      <td>-0.004903</td>
+      <td>0.003731</td>
+      <td>-0.376845</td>
+      <td>0.725428</td>
+    </tr>
+  </tbody>
+</table>
+
+<h2>Epoch-4 spectral summary</h2>
+<table border="1" class="dataframe">
+  <thead>
+    <tr>
+      <th>optimizer</th>
+      <th>matrix_type</th>
+      <th colspan="2" halign="left">alpha_raw</th>
+      <th colspan="2" halign="left">alpha_clip_xmax</th>
+      <th colspan="2" halign="left">ERG_gap</th>
+      <th colspan="2" halign="left">num_traps</th>
+    </tr>
+    <tr>
+      <th></th>
+      <th></th>
+      <th>mean</th>
+      <th>std</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>mean</th>
+      <th>std</th>
+      <th>mean</th>
+      <th>std</th>
+    </tr>
+  </thead>
+  <tbody>
+  </tbody>
+</table>
+
+<h2>adamw_D</h2><img src="plots/adamw_D.png">
+<h2>adamw_ERG_gap</h2><img src="plots/adamw_ERG_gap.png">
+<h2>adamw_alpha_clip_xmax</h2><img src="plots/adamw_alpha_clip_xmax.png">
+<h2>adamw_alpha_clip_xmax_zoomed</h2><img src="plots/adamw_alpha_clip_xmax_zoomed.png">
+<h2>adamw_alpha_raw</h2><img src="plots/adamw_alpha_raw.png">
+<h2>adamw_alpha_raw_zoomed</h2><img src="plots/adamw_alpha_raw_zoomed.png">
+<h2>adamw_num_traps</h2><img src="plots/adamw_num_traps.png">
+<h2>adamw_rand_distance</h2><img src="plots/adamw_rand_distance.png">
+<h2>grad_norm_pre_clip</h2><img src="plots/grad_norm_pre_clip.png">
+<h2>muon_clip_D</h2><img src="plots/muon_clip_D.png">
+<h2>muon_clip_ERG_gap</h2><img src="plots/muon_clip_ERG_gap.png">
+<h2>muon_clip_alpha_clip_xmax</h2><img src="plots/muon_clip_alpha_clip_xmax.png">
+<h2>muon_clip_alpha_clip_xmax_zoomed</h2><img src="plots/muon_clip_alpha_clip_xmax_zoomed.png">
+<h2>muon_clip_alpha_raw</h2><img src="plots/muon_clip_alpha_raw.png">
+<h2>muon_clip_alpha_raw_zoomed</h2><img src="plots/muon_clip_alpha_raw_zoomed.png">
+<h2>muon_clip_num_traps</h2><img src="plots/muon_clip_num_traps.png">
+<h2>muon_clip_rand_distance</h2><img src="plots/muon_clip_rand_distance.png">
+<h2>train_loss</h2><img src="plots/train_loss.png">
+<h2>update_to_weight_ratio</h2><img src="plots/update_to_weight_ratio.png">
+<h2>val_accuracy</h2><img src="plots/val_accuracy.png">
+<h2>val_generalization_gap</h2><img src="plots/val_generalization_gap.png">
+<h2>val_loss</h2><img src="plots/val_loss.png">
+<h2>val_perplexity</h2><img src="plots/val_perplexity.png">
+<h2>weight_norm</h2><img src="plots/weight_norm.png">
+</body>
+</html>
diff --git a/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/protocol.yaml b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/protocol.yaml
new file mode 100644
index 0000000..7fd3c59
--- /dev/null
+++ b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/protocol.yaml
@@ -0,0 +1,145 @@
+protocol:
+  name: nanogpt_one_head_2026_08_21_ww_baseline
+  version: 3
+  description: Four-corpus-equivalent-epoch AdamW and MuonClip source-backed baselines with five paired seeds, one-call WeightWatcher clip_xmax/raw alpha monitoring, and seventeen permanent analysis states.
+
+dataset:
+  name: HuggingFaceFW/fineweb-edu
+  config: sample-10BT
+  split: train
+  revision: 593b3a867298afb8ce42625a270ef20ddcad28f9
+  tokenizer: gpt2
+  train_tokens: 80000000
+  val_tokens: 1000000
+  test_tokens: 1000000
+
+model:
+  vocab_size: 50257
+  block_size: 256
+  n_layer: 1
+  n_head: 1
+  n_embd: 128
+  dropout: 0.0
+  bias: false
+  tie_weights: true
+
+training:
+  seeds: [1337, 2027, 4099, 31415, 271828]
+  batch_size: 4
+  grad_accum_steps: 8
+  target_epochs: 4.0
+  epoch_interval: 0.25
+  eval_interval_steps: 500
+  eval_batches: 64
+  checkpoint_interval_steps: 500
+  grad_clip: 1.0
+
+optimizer_profiles:
+  sgd_momentum:
+    display_name: SGD + Nesterov momentum (not a campaign arm)
+    family: sgd
+    learning_rate: 0.05
+    min_learning_rate: 0.005
+    warmup_fraction: 0.10
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.90
+    dampening: 0.0
+    nesterov: true
+    weight_decay: 0.01
+
+  adamw:
+    display_name: AdamW
+    family: adamw
+    learning_rate: 0.0006
+    min_learning_rate: 0.00006
+    warmup_fraction: 0.01
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    weight_decay: 0.10
+
+  adam:
+    display_name: Adam (not a campaign arm)
+    family: adam
+    learning_rate: 0.0006
+    min_learning_rate: 0.00006
+    warmup_fraction: 0.01
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    weight_decay: 0.0
+
+  muon:
+    display_name: Muon + auxiliary AdamW (not a campaign arm)
+    family: muon
+    matrix_learning_rate: 0.02
+    matrix_min_learning_rate: 0.002
+    aux_learning_rate: 0.0003
+    aux_min_learning_rate: 0.00003
+    warmup_fraction: 0.05
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.95
+    nesterov: true
+    newton_schulz_steps: 5
+    muon_epsilon: 1.0e-7
+    matrix_weight_decay: 0.01
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+    aux_weight_decay: 0.01
+
+  muon_clip:
+    display_name: MuonClip + auxiliary AdamW
+    family: muon_clip
+    learning_rate: 0.0002
+    min_learning_rate: 0.00002
+    warmup_fraction: 0.0512
+    lr_schedule_epochs: 1.0
+    schedule: warmup_cosine
+    momentum: 0.95
+    nesterov: false
+    newton_schulz_steps: 5
+    muon_epsilon: 1.0e-7
+    weight_decay: 0.10
+    update_rms_scale: 0.20
+    qk_clip_threshold: 100.0
+    qk_clip_balance: 0.50
+    qk_diagnostics_interval: 500
+    beta1: 0.90
+    beta2: 0.95
+    epsilon: 1.0e-8
+
+evaluation:
+  train_probe_seed: 21001
+  validation_probe_seed: 22001
+  test_probe_seed: 23001
+  bleu_probe_seed: 24001
+  bleu_examples: 64
+  bleu_prompt_tokens: 64
+  bleu_continuation_tokens: 32
+  bleu_batch_size: 4
+
+weightwatcher:
+  enabled: true
+  ERG: true
+  randomize: true
+  strict: true
+  min_evals: 20
+  fix_fingers: clip_xmax
+  max_fingers: 10
+  require_raw_alpha: true
+
+runtime:
+  matmul_precision: highest
+  allow_tf32: false
+  cudnn_benchmark: false
+  mps_fallback: true
+  deterministic_algorithms: true
+  deterministic_warn_only: false
+  empty_mps_cache_after_weightwatcher: true
diff --git a/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/report.html b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/report.html
new file mode 100644
index 0000000..dbc25b9
--- /dev/null
+++ b/baseline/experiments/nanogpt_one_head_2026_08_21_baseline/results/one_head_5seed/report.html
@@ -0,0 +1,877 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>One-head nanoGPT baseline — 2026-08-21</title>
+<style>
+body { max-width: 1500px; margin: 0 auto; padding: 2rem; font-family: system-ui, sans-serif; color: #1f2937; }
+h1, h2, h3 { color: #111827; }
+.note { background: #f3f4f6; border-left: 4px solid #4b5563; padding: 0.8rem 1rem; }
+table { border-collapse: collapse; width: 100%; font-size: 0.86rem; margin-bottom: 1.5rem; }
+th, td { border: 1px solid #d1d5db; padding: 0.35rem 0.5rem; text-align: right; }
+th { background: #f3f4f6; position: sticky; top: 0; }
+td:first-child, th:first-child { text-align: left; }
+img { display: block; width: 100%; height: auto; margin: 0.75rem 0 2rem; border: 1px solid #e5e7eb; }
+code { background: #f3f4f6; padding: 0.1rem 0.25rem; }
+</style>
+</head>
+<body>
+<h1>One-head nanoGPT baseline — 2026-08-21</h1>
+<p class="note">Exact campaign: AdamW and MuonClip; seeds 1337, 2027, 4099, 31415, and 271828. The run is the replicate. Test quantities never enter the saturation calculation. A 95% interval with n=5 uses Student-t critical value 2.77645.</p>
+<h2>Campaign validation</h2>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>seed</th>
+      <th>complete</th>
+      <th>optimizer_steps</th>
+      <th>train_epochs</th>
+      <th>best_validation_loss</th>
+      <th>final_test_loss</th>
+      <th>accelerator</th>
+      <th>torch_version</th>
+      <th>git_commit</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>1337</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.39583</td>
+      <td>7.01695</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>2027</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>5.89142</td>
+      <td>5.93026</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>4099</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.36731</td>
+      <td>16.63604</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>31415</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>5.81971</td>
+      <td>6.34314</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>271828</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.67032</td>
+      <td>7.01902</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>1337</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.05294</td>
+      <td>6.93159</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>2027</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>5.85370</td>
+      <td>5.84801</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>4099</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>5.99140</td>
+      <td>7.25584</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>31415</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.16925</td>
+      <td>10.06358</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>271828</td>
+      <td>True</td>
+      <td>39063</td>
+      <td>4.00005</td>
+      <td>6.22522</td>
+      <td>6.20787</td>
+      <td>mps</td>
+      <td>2.9.1</td>
+      <td>54419d969ad17afaa5c983e0b8a5443b83dfef98</td>
+    </tr>
+  </tbody>
+</table>
+<h2>Test performance</h2>
+<p>Accuracy is next-token top-1 accuracy on a fixed held-out probe, not a classification or sequence-level accuracy. Top-5 accuracy is separate. Perplexity is exp(mean token NLL), and bits/token is NLL/log(2). BLEU is a secondary fixed greedy held-out continuation diagnostic—not translation BLEU—and is accompanied by continuation token accuracy and exact match.</p>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>checkpoint</th>
+      <th>metric</th>
+      <th>n</th>
+      <th>mean</th>
+      <th>sd</th>
+      <th>ci95_low</th>
+      <th>ci95_high</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>8.58908</td>
+      <td>4.52227</td>
+      <td>2.97394</td>
+      <td>14.20422</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>3.358e+06</td>
+      <td>7.506e+06</td>
+      <td>-5.963e+06</td>
+      <td>1.268e+07</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_bits_per_token</td>
+      <td>5</td>
+      <td>12.39143</td>
+      <td>6.52426</td>
+      <td>4.29049</td>
+      <td>20.49236</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>0.11591</td>
+      <td>0.06100</td>
+      <td>0.04017</td>
+      <td>0.19165</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>0.23844</td>
+      <td>0.10930</td>
+      <td>0.10273</td>
+      <td>0.37415</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>0.18454</td>
+      <td>0.13298</td>
+      <td>0.01943</td>
+      <td>0.34965</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>0.02510</td>
+      <td>0.00666</td>
+      <td>0.01682</td>
+      <td>0.03337</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>final</td>
+      <td>test_continuation_exact_match</td>
+      <td>5</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>6.20400</td>
+      <td>0.37294</td>
+      <td>5.74093</td>
+      <td>6.66707</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>522.29780</td>
+      <td>187.54543</td>
+      <td>289.42938</td>
+      <td>755.16621</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_bits_per_token</td>
+      <td>5</td>
+      <td>8.95049</td>
+      <td>0.53804</td>
+      <td>8.28242</td>
+      <td>9.61856</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>0.15501</td>
+      <td>0.02432</td>
+      <td>0.12481</td>
+      <td>0.18520</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>0.30306</td>
+      <td>0.03658</td>
+      <td>0.25764</td>
+      <td>0.34848</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>0.36203</td>
+      <td>0.14138</td>
+      <td>0.18648</td>
+      <td>0.53757</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>0.01982</td>
+      <td>0.00246</td>
+      <td>0.01677</td>
+      <td>0.02287</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>validation_selected</td>
+      <td>test_continuation_exact_match</td>
+      <td>5</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>7.26138</td>
+      <td>1.66348</td>
+      <td>5.19590</td>
+      <td>9.32686</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>5351</td>
+      <td>1.014e+04</td>
+      <td>-7238</td>
+      <td>1.794e+04</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_bits_per_token</td>
+      <td>5</td>
+      <td>10.47596</td>
+      <td>2.39989</td>
+      <td>7.49610</td>
+      <td>13.45582</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>0.11743</td>
+      <td>0.06626</td>
+      <td>0.03517</td>
+      <td>0.19970</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>0.23661</td>
+      <td>0.12378</td>
+      <td>0.08291</td>
+      <td>0.39030</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>0.26458</td>
+      <td>0.17452</td>
+      <td>0.04789</td>
+      <td>0.48126</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>0.01914</td>
+      <td>0.01178</td>
+      <td>0.00451</td>
+      <td>0.03377</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>final</td>
+      <td>test_continuation_exact_match</td>
+      <td>5</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_loss</td>
+      <td>5</td>
+      <td>6.02950</td>
+      <td>0.15409</td>
+      <td>5.83816</td>
+      <td>6.22083</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_perplexity</td>
+      <td>5</td>
+      <td>419.41701</td>
+      <td>63.37973</td>
+      <td>340.72068</td>
+      <td>498.11335</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_bits_per_token</td>
+      <td>5</td>
+      <td>8.69872</td>
+      <td>0.22231</td>
+      <td>8.42269</td>
+      <td>8.97476</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_accuracy</td>
+      <td>5</td>
+      <td>0.16896</td>
+      <td>0.00885</td>
+      <td>0.15797</td>
+      <td>0.17995</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_top5_accuracy</td>
+      <td>5</td>
+      <td>0.32617</td>
+      <td>0.01344</td>
+      <td>0.30949</td>
+      <td>0.34286</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_bleu</td>
+      <td>5</td>
+      <td>0.42092</td>
+      <td>0.13115</td>
+      <td>0.25807</td>
+      <td>0.58376</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_continuation_token_accuracy</td>
+      <td>5</td>
+      <td>0.02002</td>
+      <td>0.00401</td>
+      <td>0.01504</td>
+      <td>0.02500</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>validation_selected</td>
+      <td>test_continuation_exact_match</td>
+      <td>5</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+      <td>0.00000</td>
+    </tr>
+  </tbody>
+</table>
+<h2>Paired seeded differences</h2>
+<p>Every difference is <code>optimizer_b - optimizer_a</code>. Positive and negative values must be interpreted according to whether higher or lower is preferable for the metric.</p>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>checkpoint</th>
+      <th>metric</th>
+      <th>contrast</th>
+      <th>n</th>
+      <th>mean</th>
+      <th>sd</th>
+      <th>ci95_low</th>
+      <th>ci95_high</th>
+      <th>valid_exact_n5</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>final</td>
+      <td>test_accuracy</td>
+      <td>muon_clip minus adamw</td>
+      <td>5</td>
+      <td>0.00153</td>
+      <td>0.09016</td>
+      <td>-0.11042</td>
+      <td>0.11348</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>final</td>
+      <td>test_loss</td>
+      <td>muon_clip minus adamw</td>
+      <td>5</td>
+      <td>-1.32770</td>
+      <td>4.83962</td>
+      <td>-7.33689</td>
+      <td>4.68148</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_accuracy</td>
+      <td>muon_clip minus adamw</td>
+      <td>5</td>
+      <td>0.01395</td>
+      <td>0.02164</td>
+      <td>-0.01291</td>
+      <td>0.04082</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>validation_selected</td>
+      <td>test_loss</td>
+      <td>muon_clip minus adamw</td>
+      <td>5</td>
+      <td>-0.17451</td>
+      <td>0.34166</td>
+      <td>-0.59874</td>
+      <td>0.24972</td>
+      <td>True</td>
+    </tr>
+  </tbody>
+</table>
+<h2>Validation-only saturation diagnostic</h2>
+<p>Plateau means |one-epoch validation-NLL improvement| ≤ 0.010 nat/token for two consecutive intervals. Runs are not stopped by this diagnostic.</p>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>seed</th>
+      <th>plateau_detected</th>
+      <th>plateau_assessment_end_epoch</th>
+      <th>degradation_detected</th>
+      <th>first_degradation_end_epoch</th>
+      <th>best_validation_epoch</th>
+      <th>best_validation_loss</th>
+      <th>final_validation_loss</th>
+      <th>test_metrics_used</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>1337</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>2.74995</td>
+      <td>6.39583</td>
+      <td>7.04255</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>2027</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>1.75002</td>
+      <td>5.89142</td>
+      <td>5.96488</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>4099</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>3.24997</td>
+      <td>6.36731</td>
+      <td>16.62434</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>31415</td>
+      <td>False</td>
+      <td></td>
+      <td>True</td>
+      <td>3.00000</td>
+      <td>3.74999</td>
+      <td>5.81971</td>
+      <td>6.36395</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>271828</td>
+      <td>False</td>
+      <td></td>
+      <td>True</td>
+      <td>3.00000</td>
+      <td>1.25000</td>
+      <td>6.70168</td>
+      <td>7.05252</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>1337</td>
+      <td>False</td>
+      <td></td>
+      <td>True</td>
+      <td>4.00000</td>
+      <td>3.24997</td>
+      <td>6.05294</td>
+      <td>6.95514</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>2027</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>3.50003</td>
+      <td>5.85370</td>
+      <td>5.89004</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>4099</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>2.74995</td>
+      <td>5.99140</td>
+      <td>7.28535</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>31415</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>3.50003</td>
+      <td>6.16925</td>
+      <td>10.07311</td>
+      <td>False</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>271828</td>
+      <td>False</td>
+      <td></td>
+      <td>False</td>
+      <td></td>
+      <td>3.50003</td>
+      <td>6.22522</td>
+      <td>6.23400</td>
+      <td>False</td>
+    </tr>
+  </tbody>
+</table>
+<h2>Final WeightWatcher alpha</h2>
+<p>The six layer values are reduced to one median inside each seeded run before any across-seed mean or confidence interval is computed.</p>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>epoch</th>
+      <th>metric</th>
+      <th>n</th>
+      <th>mean</th>
+      <th>sd</th>
+      <th>ci95_low</th>
+      <th>ci95_high</th>
+      <th>valid_exact_n5</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>adamw</td>
+      <td>4.00005</td>
+      <td>alpha_raw_six_matrix_median</td>
+      <td>5</td>
+      <td>1.74557</td>
+      <td>0.17481</td>
+      <td>1.52851</td>
+      <td>1.96263</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>4.00005</td>
+      <td>alpha_clip_xmax_six_matrix_median</td>
+      <td>5</td>
+      <td>1.66405</td>
+      <td>0.03628</td>
+      <td>1.61901</td>
+      <td>1.70910</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>adamw</td>
+      <td>4.00005</td>
+      <td>alpha_clip_minus_raw_six_matrix_median</td>
+      <td>5</td>
+      <td>-0.08151</td>
+      <td>0.16059</td>
+      <td>-0.28091</td>
+      <td>0.11788</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>4.00005</td>
+      <td>alpha_raw_six_matrix_median</td>
+      <td>5</td>
+      <td>3.97721</td>
+      <td>0.34028</td>
+      <td>3.55469</td>
+      <td>4.39972</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>4.00005</td>
+      <td>alpha_clip_xmax_six_matrix_median</td>
+      <td>5</td>
+      <td>2.97151</td>
+      <td>0.42783</td>
+      <td>2.44029</td>
+      <td>3.50272</td>
+      <td>True</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>4.00005</td>
+      <td>alpha_clip_minus_raw_six_matrix_median</td>
+      <td>5</td>
+      <td>-1.00570</td>
+      <td>0.72123</td>
+      <td>-1.90123</td>
+      <td>-0.11018</td>
+      <td>True</td>
+    </tr>
+  </tbody>
+</table>
+<h2>MuonClip QK diagnostics</h2>
+<table class="dataframe dataframe">
+  <thead>
+    <tr style="text-align: right;">
+      <th>optimizer</th>
+      <th>optimizer_label</th>
+      <th>seed</th>
+      <th>diagnostic_rows</th>
+      <th>last_step</th>
+      <th>active_fraction_weighted</th>
+      <th>max_logit_observed</th>
+      <th>min_gamma_observed</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip + auxiliary AdamW</td>
+      <td>1337</td>
+      <td>79</td>
+      <td>3.906e+04</td>
+      <td>0.00000</td>
+      <td>11.74977</td>
+      <td>1.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip + auxiliary AdamW</td>
+      <td>2027</td>
+      <td>79</td>
+      <td>3.906e+04</td>
+      <td>0.00000</td>
+      <td>9.05290</td>
+      <td>1.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip + auxiliary AdamW</td>
+      <td>4099</td>
+      <td>79</td>
+      <td>3.906e+04</td>
+      <td>0.00000</td>
+      <td>10.37812</td>
+      <td>1.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip + auxiliary AdamW</td>
+      <td>31415</td>
+      <td>79</td>
+      <td>3.906e+04</td>
+      <td>0.00000</td>
+      <td>8.79687</td>
+      <td>1.00000</td>
+    </tr>
+    <tr>
+      <td>muon_clip</td>
+      <td>MuonClip + auxiliary AdamW</td>
+      <td>271828</td>
+      <td>79</td>
+      <td>3.906e+04</td>
+      <td>0.00000</td>
+      <td>4.69516</td>
+      <td>1.00000</td>
+    </tr>
+  </tbody>
+</table>
+<h2>Plots</h2>
+<h3>AdamW</h3><img src="plots/adamw_performance.png" alt="AdamW performance trajectories"><img src="plots/adamw_alpha_raw_vs_clip_xmax.png" alt="AdamW raw versus clipped alpha"><img src="plots/adamw_erg_gap_num_traps.png" alt="AdamW ERG gap and correlation traps"><h3>MuonClip + auxiliary AdamW</h3><img src="plots/muon_clip_performance.png" alt="MuonClip + auxiliary AdamW performance trajectories"><img src="plots/muon_clip_alpha_raw_vs_clip_xmax.png" alt="MuonClip + auxiliary AdamW raw versus clipped alpha"><img src="plots/muon_clip_erg_gap_num_traps.png" alt="MuonClip + auxiliary AdamW ERG gap and correlation traps">
+<h2>Checkpoint integrity</h2>
+<p>210 checkpoint files were indexed by byte size and SHA-256. See <a href="checkpoint_sha256.csv">checkpoint_sha256.csv</a>.</p>
+<h2>Machine-readable artifacts</h2>
+<ul><li><a href="campaign_runs.csv">campaign_runs.csv</a></li>
+<li><a href="metrics_all.csv">metrics_all.csv</a></li>
+<li><a href="epoch_metrics_all.csv">epoch_metrics_all.csv</a></li>
+<li><a href="spectral_layers_all.csv">spectral_layers_all.csv</a></li>
+<li><a href="spectral_summary_all.csv">spectral_summary_all.csv</a></li>
+<li><a href="test_results_all.csv">test_results_all.csv</a></li>
+<li><a href="qk_diagnostics_all.csv">qk_diagnostics_all.csv</a></li>
+<li><a href="qk_summary.csv">qk_summary.csv</a></li>
+<li><a href="performance_summary.csv">performance_summary.csv</a></li>
+<li><a href="paired_seed_differences.csv">paired_seed_differences.csv</a></li>
+<li><a href="alpha_run_medians.csv">alpha_run_medians.csv</a></li>
+<li><a href="alpha_across_seed_summary.csv">alpha_across_seed_summary.csv</a></li>
+<li><a href="saturation_diagnostics.csv">saturation_diagnostics.csv</a></li>
+<li><a href="saturation_integer_epoch_validation.csv">saturation_integer_epoch_validation.csv</a></li>
+<li><a href="saturation_across_seed_summary.csv">saturation_across_seed_summary.csv</a></li>
+<li><a href="checkpoint_sha256.csv">checkpoint_sha256.csv</a></li></ul>
+
+</body>
+</html>