diff --git a/.gitignore b/.gitignore
index f87b334..68972f9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -30,3 +30,8 @@ analysis_out/
 # the auto-labeler. Belongs with the imagery, which is intentionally not in git
 # (see benchmark/README.md) -- nothing in this repo reads it.
 benchmark/*/index.csv
+
+# Model weights never go in git -- durable homes are HF / lab storage (see
+# scripts/model_comparison/yolo_baseline/README.md, "Where the weights live").
+*.pt
+*.pth
diff --git a/docs/model_comparison.md b/docs/model_comparison.md
index 1047766..2822b46 100644
--- a/docs/model_comparison.md
+++ b/docs/model_comparison.md
@@ -26,6 +26,26 @@ Grounding DINO. Every split now carries the full roster; the remaining ❌s are
 GT-completeness correction (#55, done on two cities) and the null-recall pass on
 `manual_gold` (O(n²) in panos — see that section).
 
+**Supervised baseline in progress (issue #51).** The roster above is all zero-shot except
+RampNet. A supervised **YOLO** baseline — the architecture-vs-data ablation (does a generic
+detector trained on the RampNet dataset *also* beat the zero-shot field, or is RampNet's
+keypoint architecture doing the work?) — is training on Hyak but is **not yet in the results
+tables**: benchmark eval is pending. Live status is in #51; the training record (per-epoch
+curves, configs, provenance) is committed under
+[`scripts/model_comparison/yolo_baseline/`](../scripts/model_comparison/yolo_baseline/).
+Known gaps and caveats so far: the `y11x_tiles` config was dropped (GPU-saturated — epoch time
+exceeded the ckpt scheduling slice, so it never checkpointed), and **all five remaining configs
+hit the same training instability** — validation mAP peaks at epoch 1, collapses at epoch 3, and
+recovers as the learning rate decays, while training loss falls straight through with no NaN or
+AMP error. The collapse coincides exactly with the peak of the warmup LR ramp (`lr/pg0` 0.010 →
+0.029 over `warmup_epochs=3`), and **not** with batch size, input resolution, architecture, or the
+ckpt preemptions — `y11l_pano` and `y26_pano` share batch 4 / imgsz 1280 and both collapsed and
+recovered. Figures and the full ruling-out are in the training record; the stabilized rerun is
+tracked in #70 (whose original small-batch hypothesis this evidence refutes) and the caveat
+write-up in #72. The reported baseline will be the best-val checkpoint, selected on val, per the
+protocol in #71 — with the caveat that today's checkpoints are undertrained relative to a stable
+schedule and so give a **lower bound** on supervised-YOLO performance.
+
 Three classes of challenger, which fail differently and are worth keeping distinct:
 
 | class | models | output | tunable? |
diff --git a/scripts/model_comparison/yolo_baseline/README.md b/scripts/model_comparison/yolo_baseline/README.md
new file mode 100644
index 0000000..861fe6c
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/README.md
@@ -0,0 +1,175 @@
+# Supervised YOLO baseline — training record (issue #51)
+
+This directory is the **preserved record** of the supervised-YOLO baseline runs
+trained on the RampNet dataset on Hyak (klone). It exists so the experiment survives
+`/gscratch/scrubbed`'s ~21-day auto-purge and is reproducible/defensible for the paper.
+
+> **Status:** the runs are in-flight (issue #51); live status is tracked there and the
+> finding is summarized in `docs/model_comparison.md`. The training code lives at its
+> canonical paths — `scripts/model_comparison/run_yolo_train.slurm` and the repo-root
+> `hyak_yolo_runbook.sh` — merged in PR #76. (This record briefly carried as-run
+> snapshots of both; the only drift from the canonical copies was a display-only
+> pano-count bug, documented in #76's description, and the snapshots remain in this
+> branch's history.)
+
+**What's here (small, text, committed):**
+
+- `runs/<config>/results.csv` — per-epoch training curves (the primary evidence).
+- `runs/<config>/args.yaml` — the exact resolved config for each run.
+- `plot_training_curves.py` — regenerates every figure below from those CSVs alone
+  (CPU-only, no network, no checkpoints): `python plot_training_curves.py`.
+- `figures/*.png` — the generated diagnostics, committed so the finding is legible
+  without re-running anything.
+
+**What's deliberately NOT here** (see the repo `.gitignore` philosophy — curated record
+in git, bulk/binary/regenerable out):
+
+- **Model weights** (`best.pt`/`last.pt`) — binaries; they belong in durable external
+  storage, not git history. See "Where the weights live" below.
+- **Slurm logs** (`*.out`/`*.err`) — 30 MB of progress-bar spam, already gitignored;
+  the few relevant lines are quoted below.
+- **Benchmark-city eval numbers** — these `results.csv` values are the **internal YOLO
+  val-split proxy** used to judge *training health*, **not** the reported baseline. The
+  reported numbers come from local `compare.py` benchmark eval on each `best.pt`.
+
+## Config grid
+
+Six configs = {YOLO11l, YOLO11x, YOLO26l} × {tiles @1024, pano @1280}. One dropped.
+
+| Config       | Model    | Input        | Base ckpt     | Batch |
+|--------------|----------|--------------|---------------|-------|
+| `y11l_tiles` | YOLO11l  | tiles @1024  | `yolo11l.pt`  | 6     |
+| `y11x_tiles` | YOLO11x  | tiles @1024  | `yolo11x.pt`  | 3→12  |
+| `y26_tiles`  | YOLO26l  | tiles @1024  | `yolo26l.pt`  | 6     |
+| `y11l_pano`  | YOLO11l  | pano @1280   | `yolo11l.pt`  | 4     |
+| `y11x_pano`  | YOLO11x  | pano @1280   | `yolo11x.pt`  | 2     |
+| `y26_pano`   | YOLO26l  | pano @1280   | `yolo26l.pt`  | 4     |
+
+Batches are the as-run values from each run's committed `args.yaml`. `y11x_tiles` was
+submitted at batch 3, then resubmitted at batch 12 trying to finish an epoch inside the
+ckpt scheduling slice (its `args.yaml` is the batch-12 attempt); neither completed one.
+
+## Status & findings (snapshot: 2026-07-28, training in progress)
+
+Val-split proxy mAP50 from `results.csv`. **Every config peaks at epoch 1, collapses at
+epoch 3, and recovers as the learning rate decays.** The instability is universal, not
+per-config — see the figures.
+
+| Config       | Epochs | best (ep) | Current mAP50 | Assessment |
+|--------------|--------|-----------|---------------|------------|
+| `y26_pano`   | 14     | 0.738 (1) | **0.659 ↑**   | ✅ fully round-tripped: 0.738 → 0.125 @ep6 → 0.659 @ep14, climbing |
+| `y11l_pano`  | 10     | 0.779 (1) | **0.183 ↑**   | 🟡 recovering — 7 epochs near 0, then 0.005 → 0.183 @ep10 |
+| `y26_tiles`  | 3      | 0.647 (1) | 0.280 ↓       | 🟡 in the dip, too early to call |
+| `y11l_tiles` | 3      | 0.655 (1) | 0.042 ↓       | 🟡 in the dip, too early to call |
+| `y11x_pano`  | 7      | 0.777 (1) | **0.000**     | ❌ five straight epochs at literal 0; no recovery yet |
+| `y11x_tiles` | 0      | —         | —             | ⏹ dropped 2026-07-27 (GPU-saturated: epoch ~10 h > ckpt slice) |
+
+### The instability: collapse tracks the warmup LR peak
+
+![learning curves](figures/fig1_learning_curves.png)
+
+The `lr/pg0` column ramps **~3× during warmup** — 0.0100 (ep1) → 0.0197 (ep2) → **0.0290
+(ep3)** — then decays linearly. `warmup_epochs=3.0`. Every config's validation collapse
+begins at that peak, and every recovery so far tracks the decay.
+
+What the evidence rules **in** and **out**:
+
+- **Not a crash.** No NaN/Inf, no AMP failure anywhere in the logs; training loss keeps
+  falling straight through the collapse while val `cls_loss` explodes 1.2 → 8.2
+  (`figures/fig3_loss_divergence.png`). A textbook optimization instability.
+- **Not the preemptions.** ckpt requeues are scattered across epochs 2–13 and do not line
+  up with the collapses (`figures/fig4_per_config.png`, purple vs grey lines). `resume=True`
+  restores the full training state; `y26_pano` climbed straight through five of them.
+- **Not small physical batch.** `y11l_pano` (batch 4, imgsz 1280) and `y26_pano`
+  (**batch 4, imgsz 1280** — identical) both collapsed, and both recovered. Batch spans
+  2/4/6 and imgsz 1024/1280 across the grid with no separation. This **refutes** the
+  BatchNorm/small-batch hypothesis this record previously carried.
+- **Not architecture.** Both YOLO11 and YOLO26 collapse; both have now begun recovering.
+  Only the *rate* differs (`y26_pano` recovered by ep9; `y11l_pano` took until ep10).
+- **Not a data fault.** A broken dataset would depress train loss too, and would not
+  recover on an LR schedule.
+
+**Hypothesis (untested):** the effective peak LR at the end of warmup is too high for this
+task — 150k single-class, small-object images, fine-tuned from COCO weights. `optimizer=auto`
+resolves to **`MuSGD(lr=0.01, momentum=0.9)`** in every job (from the Slurm logs), Ultralytics'
+Muon-family optimizer. The clean test is a rerun with a lower peak LR / longer warmup, **not**
+a larger batch.
+
+### Failure signature: recall collapse, not false-positive flood
+
+![precision and recall](figures/fig2_precision_recall.png)
+
+The model **stops firing** rather than starting to guess: `y11l_pano` holds precision at
+0.94–1.00 while recall sits at 0.007–0.03, and `y11x_pano` emits no boxes at all (its
+precision reads 0 by the 0/0 convention, not from false positives). This matters for
+interpretation — the instability suppresses detections, so a checkpoint caught inside the
+dip understates recall catastrophically and would badly misrepresent the baseline.
+
+### What is reportable today
+
+Epoch 1–2 are **not** a "pretrained artifact" — an earlier version of this record said so
+and that was wrong. By epoch 1 the model has trained on all 150k images; the score is real,
+and it is high because the LR was still in the low part of the warmup ramp. Each run's
+`best.pt` therefore holds a **genuine, selectable checkpoint** under the best-val protocol.
+
+The honest caveat: these checkpoints are **undertrained** relative to a stable schedule, so
+any benchmark number from them is a **lower bound** on supervised-YOLO performance, and must
+be reported as such.
+
+## Provenance
+
+- **Training code:** `scripts/model_comparison/run_yolo_train.slurm` + the repo-root
+  `hyak_yolo_runbook.sh` (PR #76). (The Hyak runs used an rsync of the working tree;
+  the committed launcher is byte-identical to the as-run copy.)
+- **Dataset:** `projectsidewalk/rampnet-dataset` (HF), pulled onto the cluster via
+  `download_dataset.py`. Prep: `prepare_yolo_dataset.py` via `run_yolo_prep.slurm` —
+  `--box-size pitch --ramp-size-m 1.8 --bg-keep-frac 0.15` per that launcher's as-run
+  defaults (its header records that `fixed:0.03` was rejected as too small and that
+  `gps` fell back to pitch on ~85% of panos). Tiles rendered at 1024 (train imgsz
+  1024); pano at 2048×1024 (train imgsz 1280).
+  **Confirmed** from the run notes (2026-07-25, prep job 37677130, ~2h48m): box-size
+  `pitch` + ramp 1.8 m was chosen after eyeballing the overlay QA — `fixed:0.03` was
+  too small and `gps` fell back to pitch on 169/200 smoke panos. Output verified on
+  scratch: tiles 557,413 train / 161,002 val images; pano 150,063 / 42,875. (An
+  earlier draft of this record said `fixed:0.03` — that was the runbook's old
+  default, not what ran.)
+- **Toolchain:** Ultralytics 8.4.105 · Python 3.11.15 · torch 2.13.0+cu126.
+- **Hardware:** NVIDIA L40 (45 GB), 1 GPU/job, Hyak `ckpt-g2` (preemptable/requeue).
+- **Hyperparameters (resolved):** `epochs=60`, `patience=20`, `optimizer=auto`,
+  `lr0=0.01`, `lrf=0.01`, `momentum=0.937`, `weight_decay=0.0005`, `warmup_epochs=3.0`,
+  `close_mosaic=10`, `amp=true`, `seed=0`. Per-run detail in each `runs/<config>/args.yaml`.
+  **`optimizer=auto` resolved to `MuSGD(lr=0.01, momentum=0.9)`** in all six jobs, which
+  overrides the passed `lr0`/`momentum` (the Slurm logs say so explicitly). The realized
+  schedule peaks at `lr/pg0 = 0.029` at the end of warmup — recorded per-epoch in the
+  `lr/pg*` columns of every `results.csv`, and the subject of the instability above.
+- **Slurm job IDs:** y11l_tiles 37745358 · y11x_tiles 37745359 (batch-3 original;
+  batch-12 resubmit 37809205, scancel'd 2026-07-27) · y26_tiles 37745360 · y11l_pano
+  37745361 · y11x_pano 37745362 · y26_pano 37745363. The job↔config map was read off
+  the `yolo_train_<jobid>.out` log headers during the 2026-07-27 preemption check.
+  Dataset-download job 37649635; prep job 37677130.
+- **Run dates:** 2026-07-26 → in progress as of 2026-07-28.
+
+## Where the weights live
+
+`best.pt` files are **not** in git. Durable homes:
+
+_TODO — stage to Hugging Face (`projectsidewalk/…`) or lab storage and record the URL here._
+
+Keep **every** run's `best.pt`, including the ones that collapsed. An earlier version of
+this record called those non-reportable epoch-1 artifacts and proposed letting them expire;
+that was wrong on both counts (see "What is reportable today"). Each is a real best-val
+checkpoint and the only supervised-YOLO baseline available until a stable schedule lands.
+
+## Reproducing
+
+On klone, via the repo-root runbook (each stage is idempotent; the multi-hour `data` /
+`prep` stages have compute-node sbatch wrappers, `run_yolo_data.slurm` /
+`run_yolo_prep.slurm` — login nodes reap heavy processes):
+
+```bash
+bash hyak_yolo_runbook.sh env      # lean venv on scratch (ultralytics + cu126 torch)
+bash hyak_yolo_runbook.sh data     # pull RampNet dataset from HF (hours)
+bash hyak_yolo_runbook.sh prep     # build tiles/ and pano/ YOLO datasets
+bash hyak_yolo_runbook.sh train    # sbatch the 6 configs concurrently
+bash hyak_yolo_runbook.sh collect  # rsync best.pt back; eval locally with compare.py
+```
diff --git a/scripts/model_comparison/yolo_baseline/figures/fig1_learning_curves.png b/scripts/model_comparison/yolo_baseline/figures/fig1_learning_curves.png
new file mode 100644
index 0000000..1501a33
Binary files /dev/null and b/scripts/model_comparison/yolo_baseline/figures/fig1_learning_curves.png differ
diff --git a/scripts/model_comparison/yolo_baseline/figures/fig2_precision_recall.png b/scripts/model_comparison/yolo_baseline/figures/fig2_precision_recall.png
new file mode 100644
index 0000000..2e13c2a
Binary files /dev/null and b/scripts/model_comparison/yolo_baseline/figures/fig2_precision_recall.png differ
diff --git a/scripts/model_comparison/yolo_baseline/figures/fig3_loss_divergence.png b/scripts/model_comparison/yolo_baseline/figures/fig3_loss_divergence.png
new file mode 100644
index 0000000..b774cf4
Binary files /dev/null and b/scripts/model_comparison/yolo_baseline/figures/fig3_loss_divergence.png differ
diff --git a/scripts/model_comparison/yolo_baseline/figures/fig4_per_config.png b/scripts/model_comparison/yolo_baseline/figures/fig4_per_config.png
new file mode 100644
index 0000000..51e94b4
Binary files /dev/null and b/scripts/model_comparison/yolo_baseline/figures/fig4_per_config.png differ
diff --git a/scripts/model_comparison/yolo_baseline/plot_training_curves.py b/scripts/model_comparison/yolo_baseline/plot_training_curves.py
new file mode 100644
index 0000000..ad3c541
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/plot_training_curves.py
@@ -0,0 +1,310 @@
+#!/usr/bin/env python3
+"""Training-diagnostic figures for the supervised-YOLO baseline (issue #51).
+
+Reads the committed per-epoch records in ``runs/<config>/results.csv`` (pulled from
+the Hyak run directories, which live on ``/gscratch/scrubbed`` and auto-purge) and
+writes the standard set of figures used to diagnose a detection training run:
+
+  fig1_learning_curves   val mAP50 / mAP50-95 vs epoch, with the LR schedule overlaid
+  fig2_precision_recall  val precision and recall vs epoch, separately
+  fig3_loss_divergence   train vs val loss per component, one panel per config
+  fig4_per_config        per-config mAP50 with LR and preemption/resume markers
+
+Everything here is CPU-only and reads nothing but the committed CSVs, so the
+figures are reproducible from a clean checkout:
+
+    python scripts/model_comparison/yolo_baseline/plot_training_curves.py
+
+Why the LR overlay is on the headline figure: every config's validation collapse
+begins at the epoch where the warmup ramp peaks, and recovery tracks the decay.
+See README.md for the reading.
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import pandas as pd
+
+HERE = Path(__file__).resolve().parent
+
+# Display order and styling. Geometry picks the linestyle, model family the colour,
+# so "is this a pano or a tiles run?" and "is this v11 or v26?" are both readable
+# without consulting the legend.
+CONFIGS = [
+    # name,          label,                   colour,    linestyle
+    ("y26_pano",   "YOLO26  pano  (b4, 1280)",  "#1b6ca8", "-"),
+    ("y11l_pano",  "YOLO11l pano  (b4, 1280)",  "#c1121f", "-"),
+    ("y11x_pano",  "YOLO11x pano  (b2, 1280)",  "#e07a5f", "-"),
+    ("y26_tiles",  "YOLO26  tiles (b6, 1024)",  "#457b9d", "--"),
+    ("y11l_tiles", "YOLO11l tiles (b6, 1024)",  "#8b1e3f", "--"),
+]
+
+MAP50 = "metrics/mAP50(B)"
+MAP5095 = "metrics/mAP50-95(B)"
+PRECISION = "metrics/precision(B)"
+RECALL = "metrics/recall(B)"
+LR = "lr/pg0"
+
+
+def load_runs(runs_dir: Path) -> dict[str, pd.DataFrame]:
+    """Load every config that has a results.csv, preserving CONFIGS order.
+
+    A config with no results.csv is skipped with a note rather than an error --
+    y11x_tiles was dropped before it finished epoch 1 and has args.yaml only, and
+    that absence is itself part of the record.
+    """
+    runs: dict[str, pd.DataFrame] = {}
+    for name, *_ in CONFIGS:
+        path = runs_dir / name / "results.csv"
+        if not path.exists():
+            print(f"  skip {name}: no results.csv (dropped before epoch 1?)")
+            continue
+        df = pd.read_csv(path)
+        df.columns = [c.strip() for c in df.columns]
+        runs[name] = df
+    return runs
+
+
+def resume_epochs(df: pd.DataFrame) -> list[int]:
+    """Epochs at which the training process restarted.
+
+    Ultralytics writes ``time`` as seconds elapsed in the *current* process, so it
+    resets on every requeue. A decrease therefore marks a ckpt preemption + resume.
+    """
+    t = df["time"].to_numpy()
+    return [int(df["epoch"].iloc[i]) for i in range(1, len(t)) if t[i] < t[i - 1]]
+
+
+def warmup_peak_epoch(df: pd.DataFrame) -> int | None:
+    """Epoch where the LR schedule peaks (end of the warmup ramp)."""
+    if LR not in df.columns or df[LR].isna().all():
+        return None
+    return int(df["epoch"].iloc[int(df[LR].to_numpy().argmax())])
+
+
+def _style(ax, xlabel="epoch", ylabel=None, title=None):
+    ax.grid(alpha=0.25, linewidth=0.6)
+    ax.set_xlabel(xlabel)
+    if ylabel:
+        ax.set_ylabel(ylabel)
+    if title:
+        ax.set_title(title, fontsize=11)
+    ax.spines[["top", "right"]].set_visible(False)
+
+
+def fig_learning_curves(runs, out: Path) -> None:
+    fig, axes = plt.subplots(1, 2, figsize=(13, 4.8))
+
+    for metric, ax, label in ((MAP50, axes[0], "val mAP@50"),
+                              (MAP5095, axes[1], "val mAP@50-95")):
+        for name, lbl, colour, ls in CONFIGS:
+            if name not in runs:
+                continue
+            df = runs[name]
+            ax.plot(df["epoch"], df[metric], color=colour, linestyle=ls,
+                    marker="o", markersize=3.5, linewidth=1.8, label=lbl, zorder=3)
+        _style(ax, ylabel=label, title=label + " vs epoch")
+        ax.set_ylim(bottom=-0.02)
+
+    # LR overlay on the headline panel only -- the point is the alignment, and
+    # repeating it on both panels just adds ink.
+    ref = next(iter(runs.values()))
+    twin = axes[0].twinx()
+    twin.plot(ref["epoch"], ref[LR], color="#666666", linestyle=":", linewidth=1.6,
+              label="learning rate", zorder=1)
+    twin.set_ylabel("learning rate (lr/pg0)", color="#666666")
+    twin.tick_params(axis="y", colors="#666666")
+    twin.spines[["top"]].set_visible(False)
+
+    peak = warmup_peak_epoch(ref)
+    if peak is not None:
+        for ax in axes:
+            ax.axvline(peak, color="#666666", linestyle="-", linewidth=1.0, alpha=0.5, zorder=1)
+        axes[0].annotate(
+            f"warmup peak\n(epoch {peak}, lr={ref[LR].max():.3f})",
+            xy=(peak, axes[0].get_ylim()[1] * 0.92), xytext=(peak + 0.5, axes[0].get_ylim()[1] * 0.80),
+            fontsize=8.5, color="#444444",
+            arrowprops=dict(arrowstyle="->", color="#888888", linewidth=0.9),
+        )
+
+    axes[0].legend(fontsize=8, loc="center right", framealpha=0.92)
+    fig.suptitle(
+        "Supervised YOLO baseline (#51): every config peaks before the warmup LR peak, "
+        "collapses at it, and recovers as the LR decays",
+        fontsize=11.5, y=1.0,
+    )
+    fig.tight_layout()
+    fig.savefig(out, dpi=200, bbox_inches="tight")
+    plt.close(fig)
+    print(f"  wrote {out.name}")
+
+
+def fig_precision_recall(runs, out: Path) -> None:
+    fig, axes = plt.subplots(1, 2, figsize=(13, 4.8))
+
+    for metric, ax, label in ((PRECISION, axes[0], "val precision"),
+                              (RECALL, axes[1], "val recall")):
+        for name, lbl, colour, ls in CONFIGS:
+            if name not in runs:
+                continue
+            df = runs[name]
+            ax.plot(df["epoch"], df[metric], color=colour, linestyle=ls,
+                    marker="o", markersize=3.5, linewidth=1.8, label=lbl)
+        _style(ax, ylabel=label, title=label + " vs epoch")
+        ax.set_ylim(-0.03, 1.05)
+
+    axes[0].legend(fontsize=8, loc="lower left", framealpha=0.92)
+    # Precision reads 0 where a model emits *no* boxes at all (y11x_pano from epoch 3),
+    # which is the 0/0 convention, not a flood of false positives. y11l_pano is the
+    # legible case: precision holds at 0.94-1.00 while recall sits near 0.01.
+    fig.suptitle(
+        "The failure signature is a RECALL collapse: the model stops firing rather than "
+        "starting to guess\n(precision holds high, or reads 0 where nothing is emitted at all)",
+        fontsize=11.5, y=1.02,
+    )
+    fig.tight_layout()
+    fig.savefig(out, dpi=200, bbox_inches="tight")
+    plt.close(fig)
+    print(f"  wrote {out.name}")
+
+
+def fig_loss_divergence(runs, out: Path) -> None:
+    """Train vs val loss per config -- the generalization-gap view.
+
+    YOLO26 logs an l1_loss where YOLO11 logs dfl_loss, so the third component is
+    resolved per config rather than assumed.
+    """
+    names = [n for n, *_ in CONFIGS if n in runs]
+    fig, axes = plt.subplots(2, len(names), figsize=(3.3 * len(names), 6.6), squeeze=False)
+
+    for col, name in enumerate(names):
+        df = runs[name]
+        lbl = next(c[1] for c in CONFIGS if c[0] == name)
+        for row, comp in enumerate(("box_loss", "cls_loss")):
+            ax = axes[row][col]
+            ax.plot(df["epoch"], df[f"train/{comp}"], color="#2a9d8f", linewidth=1.8,
+                    marker="o", markersize=3, label="train")
+            ax.plot(df["epoch"], df[f"val/{comp}"], color="#c1121f", linewidth=1.8,
+                    linestyle="--", marker="s", markersize=3, label="val")
+            peak = warmup_peak_epoch(df)
+            if peak is not None:
+                ax.axvline(peak, color="#888888", linewidth=1.0, alpha=0.5)
+            _style(ax, ylabel=comp if col == 0 else None,
+                   title=lbl.split("(")[0].strip() if row == 0 else None)
+            if row == 0 and col == 0:
+                ax.legend(fontsize=8, loc="upper left")
+
+    fig.suptitle(
+        "Train vs validation loss: training loss keeps falling while validation "
+        "cls_loss explodes -- an optimization instability, not a crash or a data fault",
+        fontsize=11.5, y=1.0,
+    )
+    fig.tight_layout()
+    fig.savefig(out, dpi=200, bbox_inches="tight")
+    plt.close(fig)
+    print(f"  wrote {out.name}")
+
+
+def fig_per_config(runs, out: Path) -> None:
+    """One panel per config: mAP50 + LR + ckpt preemption markers."""
+    names = [n for n, *_ in CONFIGS if n in runs]
+    ncol = min(3, len(names))
+    nrow = (len(names) + ncol - 1) // ncol
+    fig, axes = plt.subplots(nrow, ncol, figsize=(4.6 * ncol, 3.6 * nrow), squeeze=False)
+
+    for i, name in enumerate(names):
+        ax = axes[i // ncol][i % ncol]
+        df = runs[name]
+        lbl, colour, ls = next((c[1], c[2], c[3]) for c in CONFIGS if c[0] == name)
+
+        ax.plot(df["epoch"], df[MAP50], color=colour, linestyle=ls, marker="o",
+                markersize=4, linewidth=2, label="val mAP@50", zorder=3)
+
+        twin = ax.twinx()
+        twin.plot(df["epoch"], df[LR], color="#666666", linestyle=":", linewidth=1.4, zorder=1)
+        twin.set_ylabel("lr", color="#666666", fontsize=9)
+        twin.tick_params(axis="y", colors="#666666", labelsize=8)
+        twin.spines[["top"]].set_visible(False)
+
+        for j, ep in enumerate(resume_epochs(df)):
+            ax.axvline(ep, color="#7209b7", linewidth=1.2, alpha=0.55, linestyle="-.",
+                       zorder=2, label="ckpt resume" if j == 0 else None)
+        peak = warmup_peak_epoch(df)
+        if peak is not None:
+            ax.axvline(peak, color="#888888", linewidth=1.2, alpha=0.6, zorder=2)
+
+        best = df.loc[df[MAP50].idxmax()]
+        ax.annotate(f"best ep{int(best['epoch'])}: {best[MAP50]:.3f}",
+                    xy=(best["epoch"], best[MAP50]), xytext=(4, -12),
+                    textcoords="offset points", fontsize=8, color="#333333")
+
+        _style(ax, ylabel="val mAP@50", title=lbl)
+        ax.set_ylim(-0.03, 0.9)
+        ax.legend(fontsize=7.5, loc="upper right", framealpha=0.9)
+
+    for k in range(len(names), nrow * ncol):
+        axes[k // ncol][k % ncol].axis("off")
+
+    fig.suptitle(
+        "Per-config detail. Grey line = warmup LR peak, purple = ckpt preemption/resume. "
+        "Collapse tracks the LR peak, not the resumes.",
+        fontsize=11.5, y=1.0,
+    )
+    fig.tight_layout()
+    fig.savefig(out, dpi=200, bbox_inches="tight")
+    plt.close(fig)
+    print(f"  wrote {out.name}")
+
+
+def print_summary(runs) -> None:
+    print("\nPer-config summary (internal YOLO val-split proxy, NOT benchmark eval)\n")
+    hdr = f"{'config':<12} {'eps':>4} {'best ep':>8} {'best mAP50':>11} {'last mAP50':>11} {'last P':>7} {'last R':>7} {'resumes':>8}"
+    print(hdr)
+    print("-" * len(hdr))
+    for name, *_ in CONFIGS:
+        if name not in runs:
+            print(f"{name:<12} {'--':>4}   (dropped before epoch 1 -- args.yaml only)")
+            continue
+        df = runs[name]
+        best = df.loc[df[MAP50].idxmax()]
+        last = df.iloc[-1]
+        print(f"{name:<12} {len(df):>4} {int(best['epoch']):>8} {best[MAP50]:>11.3f} "
+              f"{last[MAP50]:>11.3f} {last[PRECISION]:>7.3f} {last[RECALL]:>7.3f} "
+              f"{len(resume_epochs(df)):>8}")
+    print()
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--runs-dir", type=Path, default=HERE / "runs",
+                    help="directory of <config>/results.csv (default: ./runs)")
+    ap.add_argument("--out-dir", type=Path, default=HERE / "figures",
+                    help="where to write the PNGs (default: ./figures)")
+    args = ap.parse_args()
+
+    print(f"reading {args.runs_dir}")
+    runs = load_runs(args.runs_dir)
+    if not runs:
+        print("no results.csv found -- nothing to plot", file=sys.stderr)
+        return 1
+
+    args.out_dir.mkdir(parents=True, exist_ok=True)
+    print(f"writing {args.out_dir}")
+    fig_learning_curves(runs, args.out_dir / "fig1_learning_curves.png")
+    fig_precision_recall(runs, args.out_dir / "fig2_precision_recall.png")
+    fig_loss_divergence(runs, args.out_dir / "fig3_loss_divergence.png")
+    fig_per_config(runs, args.out_dir / "fig4_per_config.png")
+    print_summary(runs)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11l_pano/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y11l_pano/args.yaml
new file mode 100644
index 0000000..87ae55d
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11l_pano/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /gscratch/scrubbed/jfroehli/yolo_runs/y11l_pano/weights/last.pt
+data: /gscratch/scrubbed/jfroehli/yolo/pano/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 4
+imgsz: 1280
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y11l_pano
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: /gscratch/scrubbed/jfroehli/yolo_runs/y11l_pano/weights/last.pt
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.0
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y11l_pano
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11l_pano/results.csv b/scripts/model_comparison/yolo_baseline/runs/y11l_pano/results.csv
new file mode 100644
index 0000000..1c49fc7
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11l_pano/results.csv
@@ -0,0 +1,11 @@
+epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2,lr/pg3,lr/pg4,lr/pg5,lr/pg6,lr/pg7
+1,7352.39,1.58582,1.33396,1.31376,0.87308,0.67332,0.77945,0.42263,1.39892,1.13264,1.26075,0.00999973,0.00333324,0.00999973,0.00333324,0.00999973,0.00333324,0.00999973,0.00333324
+2,14020.1,1.50486,1.23682,1.26008,0.92014,0.61629,0.7638,0.41958,1.37778,1.26102,1.26183,0.0196697,0.00655658,0.0196697,0.00655658,0.0196697,0.00655658,0.0196697,0.00655658
+3,21144.2,1.47727,1.23784,1.2464,0.82445,0.03805,0.0316,0.0164,2.49774,6.40598,2.03689,0.0290097,0.00966991,0.0290097,0.00966991,0.0290097,0.00966991,0.0290097,0.00966991
+4,7167.47,1.49289,1.26434,1.25522,0,0,0,0,3.29462,8.22706,2.57656,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505
+5,8237.51,1.46674,1.24697,1.2402,0,0,0,0,3.22697,8.22256,2.53027,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934
+6,11551.6,1.43966,1.23225,1.22371,1,6e-05,0.005,0.00368,2.98339,7.89658,2.3733,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175
+7,8063.86,1.40326,1.21359,1.20978,0.944,0.02833,0.02429,0.01449,2.61371,6.89773,2.06625,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901
+8,16122.1,1.26461,1.17284,1.15046,0.97138,0.01066,0.01471,0.00775,2.85682,7.28249,2.19974,0.026535,0.008845,0.026535,0.008845,0.026535,0.008845,0.026535,0.008845
+9,26165.8,1.21197,1.14843,1.1307,0.9728,0.00728,0.005,0.00379,2.98835,7.42767,2.26476,0.02604,0.00868,0.02604,0.00868,0.02604,0.00868,0.02604,0.00868
+10,9855.4,1.26067,1.16203,1.15104,0.75276,0.20365,0.1834,0.08434,2.42545,5.31002,1.88035,0.025545,0.008515,0.025545,0.008515,0.025545,0.008515,0.025545,0.008515
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/args.yaml
new file mode 100644
index 0000000..8c94545
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /gscratch/scrubbed/jfroehli/yolo_runs/y11l_tiles/weights/last.pt
+data: /gscratch/scrubbed/jfroehli/yolo/tiles/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 6
+imgsz: 1024
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y11l_tiles
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: /gscratch/scrubbed/jfroehli/yolo_runs/y11l_tiles/weights/last.pt
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.0
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y11l_tiles
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/results.csv b/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/results.csv
new file mode 100644
index 0000000..4a4a2b9
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11l_tiles/results.csv
@@ -0,0 +1,4 @@
+epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2,lr/pg3,lr/pg4,lr/pg5,lr/pg6,lr/pg7
+1,18691.1,1.42439,1.78812,1.5495,0.87911,0.46461,0.65494,0.34407,1.29443,2.15475,1.60019,0.00999989,0.0033333,0.00999989,0.0033333,0.00999989,0.0033333,0.00999989,0.0033333
+2,18270.2,1.23796,1.67682,1.38557,0.15692,0.59146,0.30866,0.10046,2.10103,3.69541,2.39119,0.0196699,0.00655663,0.0196699,0.00655663,0.0196699,0.00655663,0.0196699,0.00655663
+3,26482.8,1.20646,1.67449,1.36324,0.26701,0.08767,0.04223,0.00424,3.13441,5.81774,3.60727,0.0290099,0.00966997,0.0290099,0.00966997,0.0290099,0.00966997,0.0290099,0.00966997
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11x_pano/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y11x_pano/args.yaml
new file mode 100644
index 0000000..861695f
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11x_pano/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /gscratch/scrubbed/jfroehli/yolo_runs/y11x_pano/weights/last.pt
+data: /gscratch/scrubbed/jfroehli/yolo/pano/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 2
+imgsz: 1280
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y11x_pano
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: /gscratch/scrubbed/jfroehli/yolo_runs/y11x_pano/weights/last.pt
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.0
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y11x_pano
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11x_pano/results.csv b/scripts/model_comparison/yolo_baseline/runs/y11x_pano/results.csv
new file mode 100644
index 0000000..d3be6ee
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11x_pano/results.csv
@@ -0,0 +1,8 @@
+epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2,lr/pg3,lr/pg4,lr/pg5,lr/pg6,lr/pg7
+1,12202.8,1.58839,1.40362,1.32786,0.89616,0.65338,0.77746,0.42107,1.40861,1.21531,1.26997,0.00999987,0.00333329,0.00999987,0.00333329,0.00999987,0.00333329,0.00999987,0.00333329
+2,23829.7,1.51142,1.29776,1.27756,0.91411,0.31562,0.5522,0.2944,1.5679,2.31553,1.49041,0.0196699,0.00655662,0.0196699,0.00655662,0.0196699,0.00655662,0.0196699,0.00655662
+3,12495.7,1.54439,1.35191,1.29484,0,0,0,0,3.35169,7.97593,2.71784,0.0290099,0.00966996,0.0290099,0.00966996,0.0290099,0.00966996,0.0290099,0.00966996
+4,10320.1,1.54659,1.3661,1.29763,0,0,0,0,4.12017,7.12724,2.49091,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505
+5,12549.5,1.52107,1.3395,1.28476,0,0,0,0,3.99518,7.35629,2.42091,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934
+6,12306.2,1.50048,1.31663,1.2736,0,0,0,0,3.88304,7.61658,2.40362,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175
+7,13165.8,1.47356,1.29413,1.26027,0,0,0,0,3.51034,7.88276,2.40115,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901
diff --git a/scripts/model_comparison/yolo_baseline/runs/y11x_tiles/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y11x_tiles/args.yaml
new file mode 100644
index 0000000..d1320af
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y11x_tiles/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /mmfs1/home/jfroehli/RampNet/yolo11x.pt
+data: /gscratch/scrubbed/jfroehli/yolo/tiles/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 12
+imgsz: 1024
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y11x_tiles
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: false
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.1
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y11x_tiles
diff --git a/scripts/model_comparison/yolo_baseline/runs/y26_pano/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y26_pano/args.yaml
new file mode 100644
index 0000000..c7e6545
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y26_pano/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /gscratch/scrubbed/jfroehli/yolo_runs/y26_pano/weights/last.pt
+data: /gscratch/scrubbed/jfroehli/yolo/pano/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 4
+imgsz: 1280
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y26_pano
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: /gscratch/scrubbed/jfroehli/yolo_runs/y26_pano/weights/last.pt
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.0
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y26_pano
diff --git a/scripts/model_comparison/yolo_baseline/runs/y26_pano/results.csv b/scripts/model_comparison/yolo_baseline/runs/y26_pano/results.csv
new file mode 100644
index 0000000..c8bf40a
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y26_pano/results.csv
@@ -0,0 +1,15 @@
+epoch,time,train/box_loss,train/cls_loss,train/l1_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/l1_loss,lr/pg0,lr/pg1,lr/pg2,lr/pg3,lr/pg4,lr/pg5,lr/pg6,lr/pg7
+1,8426.62,1.61195,1.77719,0.00269,0.81174,0.61329,0.73792,0.41315,1.36326,1.36852,0.00338,0.00999973,0.00333324,0.00999973,0.00333324,0.00999973,0.00333324,0.00999973,0.00333324
+2,16395.9,1.46926,1.51431,0.00235,0.82394,0.57935,0.72156,0.40969,1.32392,1.45054,0.00327,0.0196697,0.00655658,0.0196697,0.00655658,0.0196697,0.00655658,0.0196697,0.00655658
+3,24412.5,1.46689,1.53096,0.00235,0.85235,0.48441,0.67078,0.37666,1.35046,1.71333,0.00335,0.0290097,0.00966991,0.0290097,0.00966991,0.0290097,0.00966991,0.0290097,0.00966991
+4,7273.01,1.49076,1.56422,0.00239,0.82983,0.40175,0.59898,0.33015,1.39821,1.93435,0.00349,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505,0.028515,0.009505
+5,14507,1.41274,1.48337,0.00225,0.61098,0.2659,0.3472,0.17936,1.64631,3.1187,0.00428,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934,0.02802,0.00934
+6,21562.6,1.3595,1.42764,0.00216,0.21421,0.25785,0.12504,0.06099,1.92558,4.96187,0.00547,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175,0.027525,0.009175
+7,8725.72,1.40228,1.46717,0.00224,0.6522,0.31318,0.40801,0.22175,1.57462,2.687,0.00404,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901,0.02703,0.00901
+8,17262.2,1.31481,1.41433,0.00207,0.6173,0.41195,0.49115,0.27385,1.47443,2.14678,0.00371,0.026535,0.008845,0.026535,0.008845,0.026535,0.008845,0.026535,0.008845
+9,12033.1,1.3292,1.42895,0.0021,0.63274,0.48292,0.56741,0.33055,1.32759,1.73712,0.00327,0.02604,0.00868,0.02604,0.00868,0.02604,0.00868,0.02604,0.00868
+10,10365,1.31301,1.41394,0.00206,0.6415,0.52115,0.60416,0.36099,1.25303,1.59772,0.00306,0.025545,0.008515,0.025545,0.008515,0.025545,0.008515,0.025545,0.008515
+11,19393.6,1.2671,1.39006,0.00196,0.64587,0.53745,0.62133,0.37519,1.2237,1.54846,0.00297,0.02505,0.00835,0.02505,0.00835,0.02505,0.00835,0.02505,0.00835
+12,27905.8,1.2382,1.37647,0.00191,0.6407,0.55002,0.62359,0.3795,1.19926,1.53745,0.00288,0.024555,0.008185,0.024555,0.008185,0.024555,0.008185,0.024555,0.008185
+13,10029.7,1.25905,1.37167,0.00196,0.65355,0.5757,0.64884,0.3978,1.17371,1.48241,0.00279,0.02406,0.00802,0.02406,0.00802,0.02406,0.00802,0.02406,0.00802
+14,19929.1,1.23581,1.3696,0.0019,0.65777,0.59101,0.65936,0.40711,1.15456,1.44529,0.00272,0.023565,0.007855,0.023565,0.007855,0.023565,0.007855,0.023565,0.007855
diff --git a/scripts/model_comparison/yolo_baseline/runs/y26_tiles/args.yaml b/scripts/model_comparison/yolo_baseline/runs/y26_tiles/args.yaml
new file mode 100644
index 0000000..8461e64
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y26_tiles/args.yaml
@@ -0,0 +1,116 @@
+task: detect
+mode: train
+model: /gscratch/scrubbed/jfroehli/yolo_runs/y26_tiles/weights/last.pt
+data: /gscratch/scrubbed/jfroehli/yolo/tiles/data.yaml
+epochs: 60
+time: null
+patience: 20
+batch: 6
+imgsz: 1024
+save: true
+save_period: -1
+cache: false
+device: '0'
+workers: 8
+project: /gscratch/scrubbed/jfroehli/yolo_runs
+name: y26_tiles
+exist_ok: true
+pretrained: true
+cls_remap: true
+optimizer: auto
+verbose: true
+seed: 0
+deterministic: true
+single_cls: false
+rect: false
+cos_lr: false
+close_mosaic: 10
+resume: /gscratch/scrubbed/jfroehli/yolo_runs/y26_tiles/weights/last.pt
+amp: true
+fraction: 1.0
+profile: false
+freeze: null
+multi_scale: 0.0
+compile: false
+channels_last: false
+overlap_mask: true
+mask_ratio: 4
+dropout: 0.0
+val: true
+split: val
+save_json: false
+conf: null
+iou: 0.7
+max_det: 300
+quantize: null
+dnn: false
+plots: true
+end2end: null
+source: null
+vid_stride: 1
+stream_buffer: false
+visualize: false
+augment: false
+agnostic_nms: false
+classes: null
+retina_masks: false
+embed: null
+show: false
+save_frames: false
+save_txt: false
+save_conf: false
+save_crop: false
+show_labels: true
+show_conf: true
+show_boxes: true
+line_width: null
+format: torchscript
+keras: false
+optimize: false
+dynamic: false
+simplify: true
+opset: null
+workspace: null
+nms: false
+lr0: 0.01
+lrf: 0.01
+momentum: 0.937
+weight_decay: 0.0005
+warmup_epochs: 3.0
+warmup_momentum: 0.8
+warmup_bias_lr: 0.0
+distill_model: null
+dis: 6.0
+box: 7.5
+cls: 0.5
+cls_pw: 0.0
+dfl: 1.5
+pose: 12.0
+kobj: 1.0
+rle: 1.0
+angle: 1.0
+dlog: 1.0
+dgrad: 0.5
+dlam: 1.0
+nbs: 64
+hsv_h: 0.015
+hsv_s: 0.7
+hsv_v: 0.4
+degrees: 0.0
+translate: 0.1
+scale: 0.5
+shear: 0.0
+perspective: 0.0
+flipud: 0.0
+fliplr: 0.5
+bgr: 0.0
+mosaic: 1.0
+mixup: 0.0
+cutmix: 0.0
+copy_paste: 0.0
+copy_paste_mode: flip
+auto_augment: randaugment
+erasing: 0.4
+cfg: null
+tracker: tracktrack.yaml
+save_dir: /mmfs1/gscratch/scrubbed/jfroehli/yolo_runs/y26_tiles
diff --git a/scripts/model_comparison/yolo_baseline/runs/y26_tiles/results.csv b/scripts/model_comparison/yolo_baseline/runs/y26_tiles/results.csv
new file mode 100644
index 0000000..d5c093a
--- /dev/null
+++ b/scripts/model_comparison/yolo_baseline/runs/y26_tiles/results.csv
@@ -0,0 +1,4 @@
+epoch,time,train/box_loss,train/cls_loss,train/l1_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/l1_loss,lr/pg0,lr/pg1,lr/pg2,lr/pg3,lr/pg4,lr/pg5,lr/pg6,lr/pg7
+1,20051.5,1.42301,2.03216,0.00647,0.66595,0.55011,0.64742,0.35579,1.20557,1.79447,0.00631,0.00999989,0.0033333,0.00999989,0.0033333,0.00999989,0.0033333,0.00999989,0.0033333
+2,19265.3,1.2233,1.72429,0.00533,0.60321,0.53767,0.60144,0.35348,1.11243,1.81904,0.0057,0.0196699,0.00655663,0.0196699,0.00655663,0.0196699,0.00655663,0.0196699,0.00655663
+3,27505.8,1.20769,1.74667,0.00523,0.51781,0.18848,0.27961,0.11929,2.03582,2.99481,0.01242,0.0290099,0.00966997,0.0290099,0.00966997,0.0290099,0.00966997,0.0290099,0.00966997