diff --git a/tests/test_rf100vl_train.py b/tests/test_rf100vl_train.py
index 3d87aef..267f31b 100644
--- a/tests/test_rf100vl_train.py
+++ b/tests/test_rf100vl_train.py
@@ -85,6 +85,11 @@ def test_generate_data_yaml_uses_all_splits_and_sorted_category_ids(tmp_path):
     # (ecdetseg/configs/ecdet/ecdet.yml), and LibreYOLO's ECConfig defaults to
     # amp with the inherited float16 dtype. fp32 would match neither.
     "ec": "fp16",
+    # rf-detr trains under autocast with the dtype hardcoded to torch.bfloat16
+    # (rfdetr/engine.py::get_autocast_args at tag 1.2.0, the release current
+    # when Roboflow published their RF100-VL numbers) and ModelConfig.amp
+    # defaults to True. fp32 would be a deviation, not the conservative choice.
+    "rfdetr": "bfloat16",
 }
 
 
@@ -146,11 +151,35 @@ def test_dense_rfdetr_selects_fallback_and_micro_dataset_keeps_one_batch():
     }
     plan = rf100vl_train.select_batch_plan(recipe, spec, facts)
     assert plan["run_variant"] == "fallback"
-    assert plan["physical_batch"] == 2
+    assert plan["physical_batch"] == 4
     assert plan["effective_batch"] == 16
     assert plan["expected_batches_per_epoch_minimum"] == 1
 
 
+def test_rfdetr_recipe_matches_roboflows_own_rf100vl_settings():
+    """RF-DETR's authors state their RF100-VL numbers came from the rf-detr
+    defaults with one override: batch 16 and grad accum 1, not the library's
+    default batch 4 / accum 4. Autocast in rf-detr 1.2.0 is hardcoded to
+    bfloat16, so fp32 is a deviation too. Both are easy to reintroduce by
+    copying another family's recipe, and neither is visible in a result table,
+    so pin them here.
+    """
+    recipe = rf100vl_train.load_recipe(
+        rf100vl_train.recipe_path_for_family("rfdetr"),
+        family="rfdetr",
+    )
+    protocol = recipe["protocol"]
+    assert protocol["physical_batch"] == 16, "Roboflow ran batch 16, grad accum 1"
+    assert protocol["precision"] == "bfloat16", "rf-detr autocasts to bfloat16"
+    # A post-resize image cache pins one resolution per image, which silently
+    # defeats the multi_scale sampling this recipe asks for.
+    assert not protocol.get("cache", False)
+    assert recipe["train"]["multi_scale"] is True
+    # Sizes n/s/m inherit the batch; only l steps down, and only for VRAM.
+    for size in ("n", "s", "m"):
+        assert "physical_batch" not in recipe["sizes"][size]
+
+
 def test_rfdetr_l_oom_fallback_reduces_batch():
     spec = get_spec("rfdetr-l")
     recipe = rf100vl_train.load_recipe(
diff --git a/va_bench/recipes/rf100vl/rfdetr.json b/va_bench/recipes/rf100vl/rfdetr.json
index 8054568..1a707b2 100644
--- a/va_bench/recipes/rf100vl/rfdetr.json
+++ b/va_bench/recipes/rf100vl/rfdetr.json
@@ -6,23 +6,36 @@
     "commit": "3e3c9dffc9e4bd41a363f1dd6ca1c736e145a5d8",
     "paths": ["libreyolo/models/rfdetr/config.py"],
     "license": "MIT",
-    "normalization": "LibreYOLO RF-DETR fine-tune defaults under the RF100-VL fixed skeleton"
+    "normalization": "Roboflow's own RF100-VL settings for RF-DETR under the RF100-VL fixed skeleton",
+    "upstream_reference": {
+      "repository": "https://github.com/roboflow/rf-detr",
+      "ref": "1.2.0",
+      "paths": ["rfdetr/config.py", "rfdetr/engine.py"],
+      "license": "Apache-2.0",
+      "protocol_statement": "https://github.com/roboflow/rf-detr/issues/266#issuecomment-3110260396",
+      "quote": "The RF-DETR numbers for the current release are using the default arguments but with batch size 16 and grad accum 1",
+      "known_deviations": [
+        "Roboflow initialised from private Objects365 checkpoints that were never released; we start from the public COCO rf-detr-*.pth. Maintainer estimates the gap at ~0.1 mAP.",
+        "Roboflow's published table predates their own fix for an incorrectly initialised detection head, so their numbers are not reproducible by any current code.",
+        "Size l runs physical_batch 2 because 704 px at batch 16 does not fit 16 GB; every other size runs Roboflow's batch 16 / grad accum 1.",
+        "Resize uses cv2 INTER_LINEAR without antialias."
+      ]
+    }
   },
   "protocol": {
     "epochs": 100,
     "effective_batch": 16,
-    "physical_batch": 4,
+    "physical_batch": 16,
     "selection_metric": "valid_mAP50_95",
     "eval_interval": 1,
     "patience": 0,
     "ema": true,
     "seed": 0,
-    "precision": "fp32",
-    "cache": "disk"
+    "precision": "bfloat16"
   },
   "dense_oom_fallback": {
     "max_annotations_per_image_threshold": 200,
-    "physical_batch": 2
+    "physical_batch": 4
   },
   "sizes": {
     "n": {"imgsz": 384},
@@ -55,9 +68,8 @@
     "multi_scale": true,
     "expanded_scales": true,
     "crop_resize_prob": 0.5,
-    "backbone_lr_mult": 0.1,
     "clip_max_norm": 0.1,
     "ema_decay": 0.993,
-    "workers": 0
+    "workers": 2
   }
 }