# The Sousa run's own log — pass lines and step lines, quoted

**This is not the log file.** The Sousa trainer's log lives on the training box at
`music/out/sousa-shakedown/20260830T004943Z-sousa0/train.log`, and that box was not
reachable from the lane that assembled this kit. What is below is the log's own text,
**quoted verbatim** from the arc's receipts pack — `MATERIALS.md` §1.3, "Sousa epoch
lines, VERBATIM", which cites `train.log:115-149` and `train.log:96-113` — and the
article's Sousa numbers were read from exactly this text.

The receipts pack it was quoted from:
`estate/bench/ace-composers-2026-08/MATERIALS.md`, built 2026-08-31 (UTC),
sha256 `9cb722fb315995de1a385e2f9682980fe759403915fd5e6c323da275ce2fe614`.

The article says the kit carries per-pass **and** per-step figures. Both are here:
the `Epoch n/10, Step m, Loss:` lines are the per-step figures the trainer printed
between passes; the `[OK] Epoch n/10 in Ns, Loss:` lines are the per-pass means the
article's "training error" column quotes.

## The pass and step lines (`train.log:115-149`)

```
Epoch 1/10, Step 10, Loss: 1.5218
Epoch 1/10, Step 20, Loss: 0.8999
[OK] Epoch 1/10 in 86.6s, Loss: 1.2449
Epoch 2/10, Step 30, Loss: 0.8622
Epoch 2/10, Step 40, Loss: 0.9497
[OK] Epoch 2/10 in 86.6s, Loss: 0.9547
Epoch 3/10, Step 50, Loss: 0.9343
Epoch 3/10, Step 60, Loss: 0.8528
[OK] Epoch 3/10 in 87.3s, Loss: 0.9210
Epoch 4/10, Step 70, Loss: 0.9456
Epoch 4/10, Step 80, Loss: 0.7540
[OK] Epoch 4/10 in 87.7s, Loss: 0.8943
Epoch 5/10, Step 90, Loss: 0.9674
Epoch 5/10, Step 100, Loss: 0.7845
[OK] Epoch 5/10 in 87.8s, Loss: 0.8768
[OK] Checkpoint saved at epoch 5
Epoch 6/10, Step 110, Loss: 0.8498
Epoch 6/10, Step 120, Loss: 0.9318
[OK] Epoch 6/10 in 87.8s, Loss: 0.8690
Epoch 7/10, Step 130, Loss: 0.7757
Epoch 7/10, Step 140, Loss: 0.8762
[OK] Epoch 7/10 in 87.8s, Loss: 0.8863
Epoch 8/10, Step 150, Loss: 0.8837
Epoch 8/10, Step 160, Loss: 0.8782
[OK] Epoch 8/10 in 87.8s, Loss: 0.8797
Epoch 9/10, Step 170, Loss: 0.9002
Epoch 9/10, Step 180, Loss: 0.7671
[OK] Epoch 9/10 in 87.9s, Loss: 0.8745
Epoch 10/10, Step 190, Loss: 0.8337
Epoch 10/10, Step 200, Loss: 1.0968
Epoch 10/10, Step 210, Loss: 0.9319
[OK] Epoch 10/10 in 87.8s, Loss: 0.8839
[OK] Checkpoint saved at epoch 10
[OK] Training complete! LoRA saved to …/20260830T004943Z-sousa0-r64/final
```

## The trainer's own summary block (`train.log:96-113`)

```
============================================================
  Training Complete
============================================================
  Total time .......... 14m 36s
  Epochs .............. 0 / 10
  Total steps ......... 210
  Loss ................ 1.5218 -> 1.1180
  Best loss ........... 0.7540
  Peak VRAM ........... 5.7 GiB

  Saved Checkpoints:
    Epoch    5  Loss: 0.8768
    Epoch   10  Loss: 0.8839
```

## Three things to read carefully in this block

- **`Epochs 0 / 10` is a trainer defect**, not a failed run. The epoch counter never
  populates; the same line appears in all four runs of the arc.
- **`Loss 1.5218 -> 1.1180` and `Best loss 0.7540` are STEP losses**, not pass means.
  The article's table quotes the pass means (1.2449 → 0.8839) and says so. Two
  definitions of "loss" sit in the same log; both are printed above.
- **The confession the article prints comes straight from the last two lines.** Pass
  five's checkpoint carries the better loss (0.8768) than pass ten's (0.8839), and the
  adapter that shipped — and that every downstream render used — is the pass-ten file.