diff --git a/docs/cloud_gpu_training.md b/docs/cloud_gpu_training.md
index 6eed660..efef907 100644
--- a/docs/cloud_gpu_training.md
+++ b/docs/cloud_gpu_training.md
@@ -1,5 +1,39 @@
 # Cloud GPU Training
 
+## Default: Kaggle (free), not Modal
+
+GPU training runs on Kaggle's free T4/P100 sessions. The production Modal
+workspace hosts OpenAdapt's managed runner, compiler and recorder. On
+2026-10-16 GPU training used up its monthly spend limit, and Modal stopped
+production with it. Modal GPU commands in `openadapt_ml.cloud.modal_cloud` now
+refuse to run unless you set `OPENADAPT_ALLOW_MODAL_GPU=1`, set
+`MODAL_ENVIRONMENT` to a non-production environment, and activate a Modal
+profile that is not in `OPENADAPT_MODAL_PROD_WORKSPACES` (default `abrichr`).
+
+```bash
+# one-time: Kaggle API token (Kaggle → Settings → API → Create New Token) and phone verification
+uv run python -m openadapt_ml.cloud.kaggle_cloud stage --bundle <bundle_dir> --config <train.yaml> --username <kaggle_user>
+uv run python -m openadapt_ml.cloud.kaggle_cloud push --new-dataset    # later runs: push (versions the dataset)
+uv run python -m openadapt_ml.cloud.kaggle_cloud status
+uv run python -m openadapt_ml.cloud.kaggle_cloud pull                  # results → training_output/kaggle/output
+# resume: stage again with --checkpoint training_output/kaggle/output/results
+```
+
+The kernel exits with code 3 if the session has no GPU or no internet, so a
+CPU fallback can't pass for a GPU run. T4s have no bf16; the trainer turns
+bf16 off automatically. Sessions are capped at 12 hours, and the GPU quota is
+weekly.
+
+### Modal workspace budget (production)
+
+In the Modal dashboard, go to **Settings → Usage & Billing** for workspace
+`abrichr` and set a **Workspace budget** and a **spend limit**. The hosted
+services use well under $1/month, so a low cap is enough. Starter-plan budgets
+apply to the whole workspace: a separate environment in the same workspace
+still spends the production budget. Run any Modal GPU work in a separate
+workspace.
+
+
 This guide covers using cloud GPUs for training OpenAdapt-ML models. Two providers are supported:
 
 | Provider | Setup Time | Cost | Best For |
diff --git a/openadapt_ml/cloud/kaggle_cloud.py b/openadapt_ml/cloud/kaggle_cloud.py
new file mode 100644
index 0000000..e6e470f
--- /dev/null
+++ b/openadapt_ml/cloud/kaggle_cloud.py
@@ -0,0 +1,196 @@
+"""Kaggle GPU training for OpenAdapt (free T4/P100 sessions).
+
+Kaggle replaces Modal for GPU training. The flow uses only the ``kaggle`` CLI:
+
+1. ``stage``: builds a private dataset folder (bundle, config, optional
+   checkpoint) and a script-kernel folder whose ``kernel-metadata.json`` turns
+   on GPU and internet.
+2. ``push``: ``kaggle datasets create`` (or ``version``), then
+   ``kaggle kernels push``. This starts a free Kaggle session; it never
+   touches Modal.
+3. ``status`` / ``pull``: ``kaggle kernels status`` and
+   ``kaggle kernels output`` bring the results back.
+
+To resume, pass the pulled ``results`` directory as ``--checkpoint`` to the
+next ``stage``; it is uploaded with the dataset as ``checkpoint/``.
+
+Credentials: ``~/.kaggle/access_token`` or ``~/.kaggle/kaggle.json``, or the
+``KAGGLE_USERNAME``/``KAGGLE_KEY`` (or ``KAGGLE_API_TOKEN``) environment
+variables. GPU and internet in kernels need a phone-verified Kaggle account.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+KERNEL_FILE = "kaggle_train_kernel.py"
+STAGE_MARKER = "openadapt_kaggle_stage.json"
+SLUG_RE = re.compile(r"^[a-z0-9][a-z0-9-]{4,48}[a-z0-9]$")
+ACCELERATORS = ("NvidiaTeslaT4", "NvidiaTeslaP100")
+
+
+def _check_slug(value: str, what: str) -> str:
+    if not SLUG_RE.fullmatch(value):
+        raise ValueError(
+            f"{what} must be 6-50 lowercase letters, digits or dashes: {value!r}"
+        )
+    return value
+
+
+def kernel_metadata(
+    username: str, kernel_slug: str, dataset_slug: str, accelerator: str
+) -> dict:
+    """Kaggle kernel-metadata.json for a private GPU script kernel."""
+    if accelerator not in ACCELERATORS:
+        raise ValueError(f"accelerator must be one of {ACCELERATORS}")
+    return {
+        "id": f"{username}/{_check_slug(kernel_slug, 'kernel slug')}",
+        "title": kernel_slug,
+        "code_file": KERNEL_FILE,
+        "language": "python",
+        "kernel_type": "script",
+        "is_private": True,
+        "enable_gpu": True,
+        "enable_internet": True,
+        "machine_shape": accelerator,
+        "dataset_sources": [f"{username}/{_check_slug(dataset_slug, 'dataset slug')}"],
+        "competition_sources": [],
+        "kernel_sources": [],
+    }
+
+
+def dataset_metadata(username: str, dataset_slug: str) -> dict:
+    return {
+        "title": dataset_slug,
+        "id": f"{username}/{_check_slug(dataset_slug, 'dataset slug')}",
+        "licenses": [{"name": "other"}],
+    }
+
+
+def stage(
+    bundle: Path,
+    config: Path,
+    out: Path,
+    username: str,
+    dataset_slug: str,
+    kernel_slug: str,
+    accelerator: str = "NvidiaTeslaT4",
+    checkpoint: Path | None = None,
+) -> tuple[Path, Path]:
+    """Write <out>/dataset and <out>/kernel. Returns both paths."""
+    if not (bundle / "training_data.jsonl").is_file():
+        raise FileNotFoundError(f"{bundle}/training_data.jsonl not found")
+    if not config.is_file():
+        raise FileNotFoundError(config)
+    dataset_dir, kernel_dir = out / "dataset", out / "kernel"
+    for d in (dataset_dir, kernel_dir):
+        if d.exists():
+            shutil.rmtree(d)
+        d.mkdir(parents=True)
+    shutil.copytree(bundle, dataset_dir / "bundle")
+    shutil.copy2(config, dataset_dir / "train_config.yaml")
+    if checkpoint is not None:
+        shutil.copytree(checkpoint, dataset_dir / "checkpoint")
+    (dataset_dir / STAGE_MARKER).write_text(
+        json.dumps(
+            {
+                "dataset": dataset_slug,
+                "kernel": kernel_slug,
+                "has_checkpoint": checkpoint is not None,
+            }
+        )
+    )
+    (dataset_dir / "dataset-metadata.json").write_text(
+        json.dumps(dataset_metadata(username, dataset_slug), indent=2)
+    )
+    shutil.copy2(Path(__file__).with_name(KERNEL_FILE), kernel_dir / KERNEL_FILE)
+    (kernel_dir / "kernel-metadata.json").write_text(
+        json.dumps(
+            kernel_metadata(username, kernel_slug, dataset_slug, accelerator), indent=2
+        )
+    )
+    return dataset_dir, kernel_dir
+
+
+def _kaggle(*args: str) -> int:
+    print("+ kaggle", " ".join(args), flush=True)
+    return subprocess.run(["kaggle", *args]).returncode
+
+
+def push(staged: Path, new_dataset: bool, message: str) -> int:
+    dataset_dir, kernel_dir = staged / "dataset", staged / "kernel"
+    if new_dataset:
+        rc = _kaggle("datasets", "create", "-p", str(dataset_dir), "-r", "zip")
+    else:
+        rc = _kaggle(
+            "datasets", "version", "-p", str(dataset_dir), "-m", message, "-r", "zip"
+        )
+    if rc:
+        return rc
+    print(
+        "Wait until the dataset version is ready before the kernel reads it.",
+        flush=True,
+    )
+    return _kaggle("kernels", "push", "-p", str(kernel_dir))
+
+
+def _kernel_ref(staged: Path) -> str:
+    return json.loads((staged / "kernel" / "kernel-metadata.json").read_text())["id"]
+
+
+def cli_main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Kaggle GPU training for OpenAdapt")
+    sub = parser.add_subparsers(dest="command", required=True)
+    s = sub.add_parser(
+        "stage", help="Write the dataset and kernel folders (no network)"
+    )
+    s.add_argument("--bundle", type=Path, required=True)
+    s.add_argument("--config", type=Path, required=True)
+    s.add_argument("--out", type=Path, default=Path("training_output/kaggle"))
+    s.add_argument("--username", required=True, help="Kaggle username")
+    s.add_argument("--dataset-slug", default="openadapt-ml-train-data")
+    s.add_argument("--kernel-slug", default="openadapt-ml-train")
+    s.add_argument("--accelerator", default="NvidiaTeslaT4", choices=ACCELERATORS)
+    s.add_argument("--checkpoint", type=Path, help="Pulled results dir to resume from")
+    p = sub.add_parser("push", help="Upload the dataset and start the kernel")
+    p.add_argument("--staged", type=Path, default=Path("training_output/kaggle"))
+    p.add_argument(
+        "--new-dataset", action="store_true", help="First upload (datasets create)"
+    )
+    p.add_argument("--message", default="openadapt-ml training data")
+    st = sub.add_parser("status", help="Kernel status")
+    st.add_argument("--staged", type=Path, default=Path("training_output/kaggle"))
+    pl = sub.add_parser("pull", help="Download kernel output (results)")
+    pl.add_argument("--staged", type=Path, default=Path("training_output/kaggle"))
+    pl.add_argument("--out", type=Path, default=Path("training_output/kaggle/output"))
+    args = parser.parse_args(argv)
+
+    if args.command == "stage":
+        d, k = stage(
+            args.bundle,
+            args.config,
+            args.out,
+            args.username,
+            args.dataset_slug,
+            args.kernel_slug,
+            args.accelerator,
+            args.checkpoint,
+        )
+        print(f"dataset: {d}\nkernel:  {k}")
+        return 0
+    if args.command == "push":
+        return push(args.staged, args.new_dataset, args.message)
+    if args.command == "status":
+        return _kaggle("kernels", "status", _kernel_ref(args.staged))
+    args.out.mkdir(parents=True, exist_ok=True)
+    return _kaggle("kernels", "output", _kernel_ref(args.staged), "-p", str(args.out))
+
+
+if __name__ == "__main__":
+    sys.exit(cli_main())
diff --git a/openadapt_ml/cloud/kaggle_train_kernel.py b/openadapt_ml/cloud/kaggle_train_kernel.py
new file mode 100644
index 0000000..bd67cfe
--- /dev/null
+++ b/openadapt_ml/cloud/kaggle_train_kernel.py
@@ -0,0 +1,91 @@
+"""Kaggle kernel entrypoint for openadapt-ml SFT training.
+
+Pushed by ``openadapt_ml.cloud.kaggle_cloud``. It reads the staged dataset
+(bundle, config, optional checkpoint) from /kaggle/input, trains on the
+session GPU, and writes results to /kaggle/working/results so
+``kaggle kernels output`` can pull them back.
+
+It exits with code 3 before installing anything if the session has no GPU or
+no internet: a run that falls back to CPU must not pass for a GPU run.
+"""
+
+import json
+import os
+import shutil
+import socket
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+PIP_SPEC = os.environ.get("OPENADAPT_ML_PIP_SPEC", "openadapt-ml[training]")
+WORK = Path("/kaggle/working")
+RESULTS = WORK / "results"
+
+
+def _find_staged() -> Path:
+    hits = sorted(Path("/kaggle/input").rglob("openadapt_kaggle_stage.json"))
+    if not hits:
+        raise SystemExit("staged dataset not found under /kaggle/input")
+    return hits[0].parent
+
+
+def _preflight() -> None:
+    probe = subprocess.run(["nvidia-smi", "-L"], capture_output=True, text=True)
+    print("GPU", probe.stdout.strip() or probe.stderr.strip(), flush=True)
+    if probe.returncode != 0 or "GPU" not in probe.stdout:
+        print("No GPU in this session; refusing to train on CPU.", flush=True)
+        raise SystemExit(3) from None
+    try:
+        socket.gethostbyname("pypi.org")
+    except OSError as exc:
+        print(f"No internet in this session ({exc}).", flush=True)
+        raise SystemExit(3) from None
+
+
+def main() -> int:
+    _preflight()
+    staged = _find_staged()
+    stage = json.loads((staged / "openadapt_kaggle_stage.json").read_text())
+    print("stage", stage, flush=True)
+    subprocess.run([sys.executable, "-m", "pip", "install", "-q", PIP_SPEC], check=True)
+
+    bundle = WORK / "bundle"
+    if not bundle.exists():
+        shutil.copytree(staged / "bundle", bundle)
+    RESULTS.mkdir(parents=True, exist_ok=True)
+    if (staged / "checkpoint").is_dir():
+        shutil.copytree(
+            staged / "checkpoint", RESULTS / "checkpoint_in", dirs_exist_ok=True
+        )
+
+    cmd = [
+        sys.executable,
+        "-m",
+        "openadapt_ml.scripts.train",
+        "--config",
+        str(staged / "train_config.yaml"),
+        "--jsonl",
+        str(bundle / "training_data.jsonl"),
+        "--output-dir",
+        str(RESULTS),
+    ]
+    print("Running:", " ".join(cmd), flush=True)
+    started = time.time()
+    with open(RESULTS / "train_stdout.log", "w") as log:
+        rc = subprocess.run(
+            cmd, cwd=bundle, stdout=log, stderr=subprocess.STDOUT
+        ).returncode
+    summary = {
+        "status": "completed" if rc == 0 else "failed",
+        "returncode": rc,
+        "elapsed_time": time.time() - started,
+        "stage": stage,
+    }
+    (RESULTS / "kaggle_run.json").write_text(json.dumps(summary, indent=2))
+    print(json.dumps(summary), flush=True)
+    return rc
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/openadapt_ml/cloud/modal_cloud.py b/openadapt_ml/cloud/modal_cloud.py
index 7fedee4..7119e7a 100644
--- a/openadapt_ml/cloud/modal_cloud.py
+++ b/openadapt_ml/cloud/modal_cloud.py
@@ -43,6 +43,11 @@
 import tempfile
 from pathlib import Path
 
+from openadapt_ml.cloud.modal_gpu_guard import (
+    modal_gpu_refusal,
+    require_nonprod_modal_gpu,
+)
+
 # ---------------------------------------------------------------------------
 # Modal app definition (imported lazily so CLI parsing works without modal)
 # ---------------------------------------------------------------------------
@@ -75,6 +80,7 @@ def _build_app():
     Returns:
         (app, training_image, training_volume)
     """
+    require_nonprod_modal_gpu()
     modal = _get_modal()
 
     app = modal.App(MODAL_APP_NAME)
@@ -281,6 +287,7 @@ def _build_inference_app(
     Returns:
         (app, infer_fn) - the app and the inference function handle.
     """
+    require_nonprod_modal_gpu()
     modal = _get_modal()
 
     app = modal.App(INFERENCE_APP_NAME)
@@ -784,6 +791,10 @@ def _cmd_train(args: argparse.Namespace) -> int:
     3. Run training function remotely
     4. Download results
     """
+    refusal = modal_gpu_refusal()
+    if refusal:
+        print(f"Refusing Modal GPU work: {refusal}")
+        return 1
     modal = _get_modal()
 
     # --- Step 0: Auto-convert demos if --demo-dir provided ---
@@ -924,6 +935,10 @@ def _cmd_serve(args: argparse.Namespace) -> int:
     inference function that clients can call via Modal's .remote() API.
     Alternatively, clients can use the HTTP wrapper in Qwen3VLAgent.
     """
+    refusal = modal_gpu_refusal()
+    if refusal:
+        print(f"Refusing Modal GPU work: {refusal}")
+        return 1
     modal = _get_modal()
 
     adapter_remote = None
diff --git a/openadapt_ml/cloud/modal_gpu_guard.py b/openadapt_ml/cloud/modal_gpu_guard.py
new file mode 100644
index 0000000..a2c617f
--- /dev/null
+++ b/openadapt_ml/cloud/modal_gpu_guard.py
@@ -0,0 +1,93 @@
+"""Refuse Modal GPU work on the production workspace.
+
+The production Modal workspace hosts OpenAdapt's managed runner, compiler and
+recorder. On 2026-10-16 GPU training used up that workspace's monthly spend
+limit, and Modal stopped the production services with it. GPU training now
+goes to Kaggle (``openadapt_ml.cloud.kaggle_cloud``).
+
+Modal GPU work runs only when the caller opts in on purpose:
+
+* ``OPENADAPT_ALLOW_MODAL_GPU=1``, and
+* ``MODAL_ENVIRONMENT`` names a non-production environment (not ``main``,
+  ``prod`` or ``production``), and
+* the active Modal profile/workspace is not listed in
+  ``OPENADAPT_MODAL_PROD_WORKSPACES`` (default: ``abrichr``, the workspace that
+  serves app.openadapt.ai).
+
+Modal Starter budgets are workspace-wide, so a separate environment in the
+production workspace still spends the production budget. Use a separate
+workspace (``modal profile activate <other>``) for GPU work.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+
+OPT_IN_ENV = "OPENADAPT_ALLOW_MODAL_GPU"
+PROD_WORKSPACES_ENV = "OPENADAPT_MODAL_PROD_WORKSPACES"
+DEFAULT_PROD_WORKSPACES = ("abrichr",)
+PROD_ENVIRONMENTS = frozenset({"", "main", "prod", "production"})
+
+
+class ModalGpuRefused(RuntimeError):
+    """Raised when Modal GPU work would run on the production workspace."""
+
+
+def _active_profile() -> str:
+    explicit = os.environ.get("MODAL_PROFILE", "").strip()
+    if explicit:
+        return explicit
+    try:
+        out = subprocess.run(
+            ["modal", "profile", "current"],
+            capture_output=True,
+            text=True,
+            timeout=20,
+            check=False,
+        )
+    except (OSError, subprocess.SubprocessError):
+        return ""
+    return out.stdout.strip() if out.returncode == 0 else ""
+
+
+def modal_gpu_refusal(
+    environ: dict[str, str] | None = None, profile: str | None = None
+) -> str | None:
+    """Return why Modal GPU work is refused, or None when it is allowed."""
+    env = os.environ if environ is None else environ
+    if env.get(OPT_IN_ENV) != "1":
+        return (
+            f"Modal GPU work is off by default. Train on Kaggle "
+            f"(python -m openadapt_ml.cloud.kaggle_cloud), or set {OPT_IN_ENV}=1 "
+            "with a non-production Modal workspace and environment."
+        )
+    environment = env.get("MODAL_ENVIRONMENT", "").strip().lower()
+    if environment in PROD_ENVIRONMENTS:
+        return (
+            "MODAL_ENVIRONMENT must name a non-production Modal environment "
+            f"(got {environment or 'unset'!r}; main/prod/production are refused)."
+        )
+    prod = {
+        w.strip().lower()
+        for w in env.get(PROD_WORKSPACES_ENV, ",".join(DEFAULT_PROD_WORKSPACES)).split(
+            ","
+        )
+        if w.strip()
+    }
+    active = (profile if profile is not None else _active_profile()).strip().lower()
+    if not active:
+        return "Could not determine the active Modal profile; refusing GPU work."
+    if active in prod:
+        return (
+            f"Active Modal profile {active!r} is the production workspace. "
+            "Activate a separate workspace (modal profile activate <name>) for GPU work."
+        )
+    return None
+
+
+def require_nonprod_modal_gpu() -> None:
+    """Raise ModalGpuRefused unless Modal GPU work is explicitly allowed."""
+    reason = modal_gpu_refusal()
+    if reason:
+        raise ModalGpuRefused(reason)
diff --git a/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_1/model.pt b/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_1/model.pt
new file mode 100644
index 0000000..af3d2d4
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_1/model.pt differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_2/model.pt b/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_2/model.pt
new file mode 100644
index 0000000..0ac29b8
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_2/model.pt differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/convergence/convergence_data.json b/test_artifacts/grpo_e2e/20261016_215856/convergence/convergence_data.json
new file mode 100644
index 0000000..2ff4776
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/convergence/convergence_data.json
@@ -0,0 +1,383 @@
+{
+  "test_name": "test_e2e_policy_gradient_loss_convergence",
+  "num_steps": 50,
+  "rewards": [
+    1.0,
+    0.0,
+    1.0,
+    0.0
+  ],
+  "advantages": [
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003
+  ],
+  "loss_history": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+  ],
+  "policy_history": [
+    [
+      0.26248958706855774,
+      0.23751039803028107,
+      0.26248958706855774,
+      0.23751039803028107
+    ],
+    [
+      0.27491697669029236,
+      0.22508299350738525,
+      0.27491697669029236,
+      0.22508299350738525
+    ],
+    [
+      0.2872212529182434,
+      0.2127787321805954,
+      0.2872212529182434,
+      0.2127787321805954
+    ],
+    [
+      0.2993438243865967,
+      0.20065616071224213,
+      0.2993438243865967,
+      0.20065616071224213
+    ],
+    [
+      0.3112296760082245,
+      0.1887703388929367,
+      0.3112296760082245,
+      0.1887703388929367
+    ],
+    [
+      0.32282814383506775,
+      0.17717184126377106,
+      0.32282814383506775,
+      0.17717184126377106
+    ],
+    [
+      0.3340938985347748,
+      0.1659061163663864,
+      0.3340938985347748,
+      0.1659061163663864
+    ],
+    [
+      0.34498724341392517,
+      0.15501274168491364,
+      0.34498724341392517,
+      0.15501274168491364
+    ],
+    [
+      0.3554747700691223,
+      0.14452524483203888,
+      0.3554747700691223,
+      0.14452524483203888
+    ],
+    [
+      0.36552929878234863,
+      0.13447070121765137,
+      0.36552929878234863,
+      0.13447070121765137
+    ],
+    [
+      0.3751300871372223,
+      0.1248699352145195,
+      0.3751300871372223,
+      0.1248699352145195
+    ],
+    [
+      0.38426241278648376,
+      0.11573759466409683,
+      0.38426241278648376,
+      0.11573759466409683
+    ],
+    [
+      0.3929175138473511,
+      0.10708249360322952,
+      0.3929175138473511,
+      0.10708249360322952
+    ],
+    [
+      0.40109196305274963,
+      0.09890804439783096,
+      0.40109196305274963,
+      0.09890804439783096
+    ],
+    [
+      0.40878725051879883,
+      0.09121274203062057,
+      0.40878725051879883,
+      0.09121274203062057
+    ],
+    [
+      0.41600921750068665,
+      0.08399078994989395,
+      0.41600921750068665,
+      0.08399078994989395
+    ],
+    [
+      0.42276740074157715,
+      0.07723261415958405,
+      0.42276740074157715,
+      0.07723261415958405
+    ],
+    [
+      0.4290744960308075,
+      0.0709255114197731,
+      0.4290744960308075,
+      0.0709255114197731
+    ],
+    [
+      0.4349457919597626,
+      0.06505422294139862,
+      0.4349457919597626,
+      0.06505422294139862
+    ],
+    [
+      0.44039857387542725,
+      0.05960145220160484,
+      0.44039857387542725,
+      0.05960145220160484
+    ],
+    [
+      0.4454515874385834,
+      0.05454840138554573,
+      0.4454515874385834,
+      0.05454840138554573
+    ],
+    [
+      0.4501247704029083,
+      0.04987524449825287,
+      0.4501247704029083,
+      0.04987524449825287
+    ],
+    [
+      0.45443853735923767,
+      0.04556148499250412,
+      0.45443853735923767,
+      0.04556148499250412
+    ],
+    [
+      0.45841363072395325,
+      0.04158635437488556,
+      0.45841363072395325,
+      0.04158635437488556
+    ],
+    [
+      0.46207091212272644,
+      0.037929099053144455,
+      0.46207091212272644,
+      0.037929099053144455
+    ],
+    [
+      0.4654307961463928,
+      0.03456922248005867,
+      0.4654307961463928,
+      0.03456922248005867
+    ],
+    [
+      0.4685133099555969,
+      0.031486690044403076,
+      0.4685133099555969,
+      0.031486690044403076
+    ],
+    [
+      0.4713379144668579,
+      0.028662102296948433,
+      0.4713379144668579,
+      0.028662102296948433
+    ],
+    [
+      0.4739232361316681,
+      0.026076797395944595,
+      0.4739232361316681,
+      0.026076797395944595
+    ],
+    [
+      0.4762870669364929,
+      0.02371295355260372,
+      0.4762870669364929,
+      0.02371295355260372
+    ],
+    [
+      0.478446364402771,
+      0.02155364491045475,
+      0.478446364402771,
+      0.02155364491045475
+    ],
+    [
+      0.4804171323776245,
+      0.019582878798246384,
+      0.4804171323776245,
+      0.019582878798246384
+    ],
+    [
+      0.4822143614292145,
+      0.017785610631108284,
+      0.4822143614292145,
+      0.017785610631108284
+    ],
+    [
+      0.48385223746299744,
+      0.01614774949848652,
+      0.48385223746299744,
+      0.01614774949848652
+    ],
+    [
+      0.485343873500824,
+      0.014656132087111473,
+      0.485343873500824,
+      0.014656132087111473
+    ],
+    [
+      0.48670148849487305,
+      0.013298512436449528,
+      0.48670148849487305,
+      0.013298512436449528
+    ],
+    [
+      0.48793649673461914,
+      0.0120635274797678,
+      0.48793649673461914,
+      0.0120635274797678
+    ],
+    [
+      0.48905932903289795,
+      0.01094065047800541,
+      0.48905932903289795,
+      0.01094065047800541
+    ],
+    [
+      0.4900798201560974,
+      0.009920168668031693,
+      0.4900798201560974,
+      0.009920168668031693
+    ],
+    [
+      0.49100688099861145,
+      0.008993119932711124,
+      0.49100688099861145,
+      0.008993119932711124
+    ],
+    [
+      0.49184873700141907,
+      0.008151262067258358,
+      0.49184873700141907,
+      0.008151262067258358
+    ],
+    [
+      0.4926129877567291,
+      0.007387027610093355,
+      0.4926129877567291,
+      0.007387027610093355
+    ],
+    [
+      0.4933064877986908,
+      0.006693469360470772,
+      0.4933064877986908,
+      0.006693469360470772
+    ],
+    [
+      0.493935763835907,
+      0.006064228247851133,
+      0.493935763835907,
+      0.006064228247851133
+    ],
+    [
+      0.4945065379142761,
+      0.005493482109159231,
+      0.4945065379142761,
+      0.005493482109159231
+    ],
+    [
+      0.49502408504486084,
+      0.004975910764187574,
+      0.49502408504486084,
+      0.004975910764187574
+    ],
+    [
+      0.4954933226108551,
+      0.004506658762693405,
+      0.4954933226108551,
+      0.004506658762693405
+    ],
+    [
+      0.49591872096061707,
+      0.004081294406205416,
+      0.49591872096061707,
+      0.004081294406205416
+    ],
+    [
+      0.4963041841983795,
+      0.0036957787815481424,
+      0.4963041841983795,
+      0.0036957787815481424
+    ],
+    [
+      0.49665358662605286,
+      0.0033464336302131414,
+      0.49665358662605286,
+      0.0033464336302131414
+    ]
+  ],
+  "final_policy": [
+    0.49665358662605286,
+    0.0033464336302131414,
+    0.49665358662605286,
+    0.0033464336302131414
+  ],
+  "initial_policy": [
+    0.26248958706855774,
+    0.23751039803028107,
+    0.26248958706855774,
+    0.23751039803028107
+  ],
+  "duration_s": 0.017287254333496094
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/convergence/summary.txt b/test_artifacts/grpo_e2e/20261016_215856/convergence/summary.txt
new file mode 100644
index 0000000..5b1a9bc
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/convergence/summary.txt
@@ -0,0 +1,22 @@
+GRPO Loss Convergence Test
+========================================
+Steps: 50
+Rewards: [1.0, 0.0, 1.0, 0.0]
+Initial policy (uniform): ['0.2625', '0.2375', '0.2625', '0.2375']
+Final policy:             ['0.4967', '0.0033', '0.4967', '0.0033']
+Good action total prob:   0.9933
+Bad action total prob:    0.0067
+Early avg loss:           0.0000
+Late avg loss:            0.0000
+
+Per-step loss history (every 5th):
+  Step   0: loss=0.0000  policy=[0.262, 0.238, 0.262, 0.238]
+  Step   5: loss=0.0000  policy=[0.323, 0.177, 0.323, 0.177]
+  Step  10: loss=0.0000  policy=[0.375, 0.125, 0.375, 0.125]
+  Step  15: loss=0.0000  policy=[0.416, 0.084, 0.416, 0.084]
+  Step  20: loss=0.0000  policy=[0.445, 0.055, 0.445, 0.055]
+  Step  25: loss=0.0000  policy=[0.465, 0.035, 0.465, 0.035]
+  Step  30: loss=0.0000  policy=[0.478, 0.022, 0.478, 0.022]
+  Step  35: loss=0.0000  policy=[0.487, 0.013, 0.487, 0.013]
+  Step  40: loss=0.0000  policy=[0.492, 0.008, 0.492, 0.008]
+  Step  45: loss=0.0000  policy=[0.495, 0.005, 0.495, 0.005]
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/model_diff.json b/test_artifacts/grpo_e2e/20261016_215856/model_diff.json
new file mode 100644
index 0000000..dc81180
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/model_diff.json
@@ -0,0 +1,22 @@
+{
+  "embedding.weight": {
+    "initial_norm": 90.39092254638672,
+    "final_norm": 90.38932037353516,
+    "delta_norm": 0.025262756273150444,
+    "changed": true
+  },
+  "head.weight": {
+    "initial_norm": 9.175859451293945,
+    "final_norm": 9.180281639099121,
+    "delta_norm": 0.18086151778697968,
+    "changed": true
+  },
+  "head.bias": {
+    "initial_norm": 1.653204321861267,
+    "final_norm": 1.6514724493026733,
+    "delta_norm": 0.03192901238799095,
+    "changed": true
+  },
+  "_total_delta_norm": 0.18538758670094135,
+  "_any_changed": true
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/policy_gradient_loss_properties.json b/test_artifacts/grpo_e2e/20261016_215856/policy_gradient_loss_properties.json
new file mode 100644
index 0000000..4f163ab
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/policy_gradient_loss_properties.json
@@ -0,0 +1,48 @@
+{
+  "tests": [
+    {
+      "name": "positive_advantage_gradient",
+      "logp": -1.0,
+      "advantage": 1.0,
+      "gradient": -1.0,
+      "expected": "negative (optimizer increases logp)",
+      "passed": true
+    },
+    {
+      "name": "negative_advantage_gradient",
+      "logp": -1.0,
+      "advantage": -1.0,
+      "gradient": 1.0,
+      "expected": "positive (optimizer decreases logp)",
+      "passed": true
+    },
+    {
+      "name": "zero_advantage_gradient",
+      "logp": -1.0,
+      "advantage": 0.0,
+      "gradient": -0.0,
+      "expected": "zero",
+      "passed": true
+    },
+    {
+      "name": "clipping_large_ratio",
+      "ratio": 18.17414536944306,
+      "clipped_ratio": 1.2,
+      "loss": -1.2000000476837158,
+      "expected_loss": -1.2,
+      "passed": true
+    },
+    {
+      "name": "batch_consistency",
+      "individual_losses": [
+        -1.0,
+        0.5,
+        -0.0
+      ],
+      "batch_loss": -0.1666666716337204,
+      "expected_batch_mean": -0.16666666666666666,
+      "passed": true
+    }
+  ],
+  "all_passed": true
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0.json
new file mode 100644
index 0000000..384a2d3
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.20, y=0.50)",
+      "action_type": "click",
+      "x": 63,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.20, y=0.55)",
+      "action_type": "click",
+      "x": 63,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.20, y=0.60)",
+      "action_type": "click",
+      "x": 63,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.20, y=0.65)",
+      "action_type": "click",
+      "x": 63,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_2.png
new file mode 100644
index 0000000..9fd3281
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_3.png
new file mode 100644
index 0000000..8770b85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_0_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1.json
new file mode 100644
index 0000000..5047a60
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click",
+      "x": 96,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click",
+      "x": 96,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.30, y=0.60)",
+      "action_type": "click",
+      "x": 96,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.30, y=0.65)",
+      "action_type": "click",
+      "x": 96,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_2.png
new file mode 100644
index 0000000..268454d
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_3.png
new file mode 100644
index 0000000..fcb3dd3
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_1_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2.json
new file mode 100644
index 0000000..bd58ec2
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click",
+      "x": 128,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click",
+      "x": 128,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.40, y=0.60)",
+      "action_type": "click",
+      "x": 128,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.40, y=0.65)",
+      "action_type": "click",
+      "x": 128,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_2.png
new file mode 100644
index 0000000..5934673
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_3.png
new file mode 100644
index 0000000..6babda9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_2_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3.json
new file mode 100644
index 0000000..916eb51
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click",
+      "x": 160,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click",
+      "x": 160,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.50, y=0.60)",
+      "action_type": "click",
+      "x": 160,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.50, y=0.65)",
+      "action_type": "click",
+      "x": 160,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_2.png
new file mode 100644
index 0000000..4bea64d
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_3.png
new file mode 100644
index 0000000..8010e79
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_3_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4.json
new file mode 100644
index 0000000..1cc6982
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 4,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click",
+      "x": 192,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click",
+      "x": 192,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.60, y=0.60)",
+      "action_type": "click",
+      "x": 192,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.60, y=0.65)",
+      "action_type": "click",
+      "x": 192,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_0.png
new file mode 100644
index 0000000..05c07b6
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_1.png
new file mode 100644
index 0000000..d0b4238
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_2.png
new file mode 100644
index 0000000..faac2ac
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_3.png
new file mode 100644
index 0000000..2c2d4fc
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_4_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5.json
new file mode 100644
index 0000000..d0f509f
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 5,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.70, y=0.50)",
+      "action_type": "click",
+      "x": 224,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.70, y=0.55)",
+      "action_type": "click",
+      "x": 224,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.70, y=0.60)",
+      "action_type": "click",
+      "x": 224,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.70, y=0.65)",
+      "action_type": "click",
+      "x": 224,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_0.png
new file mode 100644
index 0000000..b3d658c
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_1.png
new file mode 100644
index 0000000..884fcc1
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_2.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_2.png
new file mode 100644
index 0000000..51431ee
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_3.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_3.png
new file mode 100644
index 0000000..1c6ba7a
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/rollout_5_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/summary.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/summary.json
new file mode 100644
index 0000000..c65d591
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_collection/summary.json
@@ -0,0 +1,23 @@
+{
+  "test_name": "test_e2e_rollout_collection_mock",
+  "num_rollouts": 6,
+  "steps_per_rollout": 4,
+  "rewards": [
+    1.0,
+    0.0,
+    1.0,
+    0.0,
+    1.0,
+    0.0
+  ],
+  "advantages": [
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003
+  ],
+  "reward_mean": 0.5,
+  "duration_s": 0.04496502876281738
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0.json
new file mode 100644
index 0000000..fda167c
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_0_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1.json
new file mode 100644
index 0000000..38c7391
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_1_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2.json
new file mode 100644
index 0000000..bc83a6b
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_2_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3.json
new file mode 100644
index 0000000..407fbef
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_0_rollout_3_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0.json
new file mode 100644
index 0000000..408b611
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_0_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1.json
new file mode 100644
index 0000000..7eb87d0
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_1_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2.json
new file mode 100644
index 0000000..fbd67cc
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_2_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3.json b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3.json
new file mode 100644
index 0000000..821c840
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215856/rollout_traces/step_1_rollout_3_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215856/summary.txt b/test_artifacts/grpo_e2e/20261016_215856/summary.txt
new file mode 100644
index 0000000..0246fd0
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/summary.txt
@@ -0,0 +1,25 @@
+============================================================
+GRPO E2E TEST REPORT: test_e2e_training_loop_mock
+============================================================
+Status:               passed
+Duration:             0.04s
+Training steps:       2
+Rollouts per step:    4
+Total rollouts:       8
+
+--- Model Weight Change ---
+Weights changed:      True
+Total delta L2 norm:  0.185388
+  embedding.weight: delta=0.025263 (norm 90.3909 -> 90.3893)
+  head.weight: delta=0.180862 (norm 9.1759 -> 9.1803)
+  head.bias: delta=0.031929 (norm 1.6532 -> 1.6515)
+
+--- Training Metrics ---
+  Step 0: loss=0.0000 reward_mean=0.50 grad_norm=1.0000 valid_terms=8
+  Step 1: loss=0.0000 reward_mean=0.50 grad_norm=1.0000 valid_terms=8
+
+--- Checkpoint ---
+Checkpoint path:      test_artifacts/grpo_e2e/20261016_215856/checkpoint/step_2
+Checkpoint loadable:  True
+
+============================================================
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/test_report.json b/test_artifacts/grpo_e2e/20261016_215856/test_report.json
new file mode 100644
index 0000000..b4e9138
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/test_report.json
@@ -0,0 +1,15 @@
+{
+  "test_name": "test_e2e_training_loop_mock",
+  "start_time": "2026-10-16T21:58:56",
+  "status": "passed",
+  "duration_s": 0.0386500358581543,
+  "num_training_steps": 2,
+  "num_rollouts_per_step": 4,
+  "total_rollouts": 8,
+  "model_weights_changed": true,
+  "total_delta_norm": 0.18538758670094135,
+  "checkpoint_loadable": true,
+  "final_loss": 0.0,
+  "final_reward_mean": 0.5,
+  "final_grad_norm": 0.9999997175711289
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/training_log.json b/test_artifacts/grpo_e2e/20261016_215856/training_log.json
new file mode 100644
index 0000000..3071712
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/training_log.json
@@ -0,0 +1,34 @@
+[
+  {
+    "step": 0,
+    "loss": 0.0,
+    "reward_mean": 0.5,
+    "reward_std": 0.5,
+    "advantages": [
+      0.9999999800000003,
+      -0.9999999800000003,
+      0.9999999800000003,
+      -0.9999999800000003
+    ],
+    "grad_norm": 1.0000000347938705,
+    "lr": 0.001,
+    "valid_terms": 8,
+    "step_time": 0.019833087921142578
+  },
+  {
+    "step": 1,
+    "loss": 0.0,
+    "reward_mean": 0.5,
+    "reward_std": 0.5,
+    "advantages": [
+      0.9999999800000003,
+      -0.9999999800000003,
+      0.9999999800000003,
+      -0.9999999800000003
+    ],
+    "grad_norm": 0.9999997175711289,
+    "lr": 0.001,
+    "valid_terms": 8,
+    "step_time": 0.015851974487304688
+  }
+]
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215856/weight_diff/detailed_diff.json b/test_artifacts/grpo_e2e/20261016_215856/weight_diff/detailed_diff.json
new file mode 100644
index 0000000..6713397
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215856/weight_diff/detailed_diff.json
@@ -0,0 +1,51 @@
+{
+  "parameters": {
+    "embedding.weight": {
+      "before_norm": 91.11665344238281,
+      "after_norm": 91.11591339111328,
+      "delta_l2": 0.012671872042119503,
+      "delta_max": 0.0010254383087158203,
+      "delta_mean": -1.0594574177957838e-06,
+      "shape": [
+        256,
+        32
+      ],
+      "num_elements": 8192,
+      "grad_norm": 3.506531000137329,
+      "grad_max": 1.0161888599395752
+    },
+    "head.weight": {
+      "before_norm": 9.253376960754395,
+      "after_norm": 9.25264835357666,
+      "delta_l2": 0.09050253033638,
+      "delta_max": 0.0010017603635787964,
+      "delta_mean": 4.1243034502258524e-05,
+      "shape": [
+        256,
+        32
+      ],
+      "num_elements": 8192,
+      "grad_norm": 32.2695198059082,
+      "grad_max": 6.838644027709961
+    },
+    "head.bias": {
+      "before_norm": 1.6341360807418823,
+      "after_norm": 1.6350926160812378,
+      "delta_l2": 0.01599622517824173,
+      "delta_max": 0.0010016560554504395,
+      "delta_mean": 1.566299397381954e-05,
+      "shape": [
+        256
+      ],
+      "num_elements": 256,
+      "grad_norm": 4.00044059753418,
+      "grad_max": 2.012416124343872
+    }
+  },
+  "summary": {
+    "total_params": 16640,
+    "trainable_params": 16640,
+    "changed_param_groups": 3,
+    "total_param_groups": 3
+  }
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_1/model.pt b/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_1/model.pt
new file mode 100644
index 0000000..7efb5f1
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_1/model.pt differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_2/model.pt b/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_2/model.pt
new file mode 100644
index 0000000..c49a2b7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_2/model.pt differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/convergence/convergence_data.json b/test_artifacts/grpo_e2e/20261016_215929/convergence/convergence_data.json
new file mode 100644
index 0000000..0a9af61
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/convergence/convergence_data.json
@@ -0,0 +1,383 @@
+{
+  "test_name": "test_e2e_policy_gradient_loss_convergence",
+  "num_steps": 50,
+  "rewards": [
+    1.0,
+    0.0,
+    1.0,
+    0.0
+  ],
+  "advantages": [
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003
+  ],
+  "loss_history": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+  ],
+  "policy_history": [
+    [
+      0.26248958706855774,
+      0.23751039803028107,
+      0.26248958706855774,
+      0.23751039803028107
+    ],
+    [
+      0.27491697669029236,
+      0.22508299350738525,
+      0.27491697669029236,
+      0.22508299350738525
+    ],
+    [
+      0.2872212529182434,
+      0.2127787321805954,
+      0.2872212529182434,
+      0.2127787321805954
+    ],
+    [
+      0.2993438243865967,
+      0.20065616071224213,
+      0.2993438243865967,
+      0.20065616071224213
+    ],
+    [
+      0.3112296760082245,
+      0.1887703388929367,
+      0.3112296760082245,
+      0.1887703388929367
+    ],
+    [
+      0.32282814383506775,
+      0.17717184126377106,
+      0.32282814383506775,
+      0.17717184126377106
+    ],
+    [
+      0.3340938985347748,
+      0.1659061163663864,
+      0.3340938985347748,
+      0.1659061163663864
+    ],
+    [
+      0.34498724341392517,
+      0.15501274168491364,
+      0.34498724341392517,
+      0.15501274168491364
+    ],
+    [
+      0.3554747700691223,
+      0.14452524483203888,
+      0.3554747700691223,
+      0.14452524483203888
+    ],
+    [
+      0.36552929878234863,
+      0.13447070121765137,
+      0.36552929878234863,
+      0.13447070121765137
+    ],
+    [
+      0.3751300871372223,
+      0.1248699352145195,
+      0.3751300871372223,
+      0.1248699352145195
+    ],
+    [
+      0.38426241278648376,
+      0.11573759466409683,
+      0.38426241278648376,
+      0.11573759466409683
+    ],
+    [
+      0.3929175138473511,
+      0.10708249360322952,
+      0.3929175138473511,
+      0.10708249360322952
+    ],
+    [
+      0.40109196305274963,
+      0.09890804439783096,
+      0.40109196305274963,
+      0.09890804439783096
+    ],
+    [
+      0.40878725051879883,
+      0.09121274203062057,
+      0.40878725051879883,
+      0.09121274203062057
+    ],
+    [
+      0.41600921750068665,
+      0.08399078994989395,
+      0.41600921750068665,
+      0.08399078994989395
+    ],
+    [
+      0.42276740074157715,
+      0.07723261415958405,
+      0.42276740074157715,
+      0.07723261415958405
+    ],
+    [
+      0.4290744960308075,
+      0.0709255114197731,
+      0.4290744960308075,
+      0.0709255114197731
+    ],
+    [
+      0.4349457919597626,
+      0.06505422294139862,
+      0.4349457919597626,
+      0.06505422294139862
+    ],
+    [
+      0.44039857387542725,
+      0.05960145220160484,
+      0.44039857387542725,
+      0.05960145220160484
+    ],
+    [
+      0.4454515874385834,
+      0.05454840138554573,
+      0.4454515874385834,
+      0.05454840138554573
+    ],
+    [
+      0.4501247704029083,
+      0.04987524449825287,
+      0.4501247704029083,
+      0.04987524449825287
+    ],
+    [
+      0.45443853735923767,
+      0.04556148499250412,
+      0.45443853735923767,
+      0.04556148499250412
+    ],
+    [
+      0.45841363072395325,
+      0.04158635437488556,
+      0.45841363072395325,
+      0.04158635437488556
+    ],
+    [
+      0.46207091212272644,
+      0.037929099053144455,
+      0.46207091212272644,
+      0.037929099053144455
+    ],
+    [
+      0.4654307961463928,
+      0.03456922248005867,
+      0.4654307961463928,
+      0.03456922248005867
+    ],
+    [
+      0.4685133099555969,
+      0.031486690044403076,
+      0.4685133099555969,
+      0.031486690044403076
+    ],
+    [
+      0.4713379144668579,
+      0.028662102296948433,
+      0.4713379144668579,
+      0.028662102296948433
+    ],
+    [
+      0.4739232361316681,
+      0.026076797395944595,
+      0.4739232361316681,
+      0.026076797395944595
+    ],
+    [
+      0.4762870669364929,
+      0.02371295355260372,
+      0.4762870669364929,
+      0.02371295355260372
+    ],
+    [
+      0.478446364402771,
+      0.02155364491045475,
+      0.478446364402771,
+      0.02155364491045475
+    ],
+    [
+      0.4804171323776245,
+      0.019582878798246384,
+      0.4804171323776245,
+      0.019582878798246384
+    ],
+    [
+      0.4822143614292145,
+      0.017785610631108284,
+      0.4822143614292145,
+      0.017785610631108284
+    ],
+    [
+      0.48385223746299744,
+      0.01614774949848652,
+      0.48385223746299744,
+      0.01614774949848652
+    ],
+    [
+      0.485343873500824,
+      0.014656132087111473,
+      0.485343873500824,
+      0.014656132087111473
+    ],
+    [
+      0.48670148849487305,
+      0.013298512436449528,
+      0.48670148849487305,
+      0.013298512436449528
+    ],
+    [
+      0.48793649673461914,
+      0.0120635274797678,
+      0.48793649673461914,
+      0.0120635274797678
+    ],
+    [
+      0.48905932903289795,
+      0.01094065047800541,
+      0.48905932903289795,
+      0.01094065047800541
+    ],
+    [
+      0.4900798201560974,
+      0.009920168668031693,
+      0.4900798201560974,
+      0.009920168668031693
+    ],
+    [
+      0.49100688099861145,
+      0.008993119932711124,
+      0.49100688099861145,
+      0.008993119932711124
+    ],
+    [
+      0.49184873700141907,
+      0.008151262067258358,
+      0.49184873700141907,
+      0.008151262067258358
+    ],
+    [
+      0.4926129877567291,
+      0.007387027610093355,
+      0.4926129877567291,
+      0.007387027610093355
+    ],
+    [
+      0.4933064877986908,
+      0.006693469360470772,
+      0.4933064877986908,
+      0.006693469360470772
+    ],
+    [
+      0.493935763835907,
+      0.006064228247851133,
+      0.493935763835907,
+      0.006064228247851133
+    ],
+    [
+      0.4945065379142761,
+      0.005493482109159231,
+      0.4945065379142761,
+      0.005493482109159231
+    ],
+    [
+      0.49502408504486084,
+      0.004975910764187574,
+      0.49502408504486084,
+      0.004975910764187574
+    ],
+    [
+      0.4954933226108551,
+      0.004506658762693405,
+      0.4954933226108551,
+      0.004506658762693405
+    ],
+    [
+      0.49591872096061707,
+      0.004081294406205416,
+      0.49591872096061707,
+      0.004081294406205416
+    ],
+    [
+      0.4963041841983795,
+      0.0036957787815481424,
+      0.4963041841983795,
+      0.0036957787815481424
+    ],
+    [
+      0.49665358662605286,
+      0.0033464336302131414,
+      0.49665358662605286,
+      0.0033464336302131414
+    ]
+  ],
+  "final_policy": [
+    0.49665358662605286,
+    0.0033464336302131414,
+    0.49665358662605286,
+    0.0033464336302131414
+  ],
+  "initial_policy": [
+    0.26248958706855774,
+    0.23751039803028107,
+    0.26248958706855774,
+    0.23751039803028107
+  ],
+  "duration_s": 0.017443180084228516
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/convergence/summary.txt b/test_artifacts/grpo_e2e/20261016_215929/convergence/summary.txt
new file mode 100644
index 0000000..5b1a9bc
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/convergence/summary.txt
@@ -0,0 +1,22 @@
+GRPO Loss Convergence Test
+========================================
+Steps: 50
+Rewards: [1.0, 0.0, 1.0, 0.0]
+Initial policy (uniform): ['0.2625', '0.2375', '0.2625', '0.2375']
+Final policy:             ['0.4967', '0.0033', '0.4967', '0.0033']
+Good action total prob:   0.9933
+Bad action total prob:    0.0067
+Early avg loss:           0.0000
+Late avg loss:            0.0000
+
+Per-step loss history (every 5th):
+  Step   0: loss=0.0000  policy=[0.262, 0.238, 0.262, 0.238]
+  Step   5: loss=0.0000  policy=[0.323, 0.177, 0.323, 0.177]
+  Step  10: loss=0.0000  policy=[0.375, 0.125, 0.375, 0.125]
+  Step  15: loss=0.0000  policy=[0.416, 0.084, 0.416, 0.084]
+  Step  20: loss=0.0000  policy=[0.445, 0.055, 0.445, 0.055]
+  Step  25: loss=0.0000  policy=[0.465, 0.035, 0.465, 0.035]
+  Step  30: loss=0.0000  policy=[0.478, 0.022, 0.478, 0.022]
+  Step  35: loss=0.0000  policy=[0.487, 0.013, 0.487, 0.013]
+  Step  40: loss=0.0000  policy=[0.492, 0.008, 0.492, 0.008]
+  Step  45: loss=0.0000  policy=[0.495, 0.005, 0.495, 0.005]
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/model_diff.json b/test_artifacts/grpo_e2e/20261016_215929/model_diff.json
new file mode 100644
index 0000000..1c5fdb2
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/model_diff.json
@@ -0,0 +1,22 @@
+{
+  "embedding.weight": {
+    "initial_norm": 91.4769058227539,
+    "final_norm": 91.474609375,
+    "delta_norm": 0.025375721976161003,
+    "changed": true
+  },
+  "head.weight": {
+    "initial_norm": 9.211176872253418,
+    "final_norm": 9.209989547729492,
+    "delta_norm": 0.18093866109848022,
+    "changed": true
+  },
+  "head.bias": {
+    "initial_norm": 1.5798417329788208,
+    "final_norm": 1.580965280532837,
+    "delta_norm": 0.031921159476041794,
+    "changed": true
+  },
+  "_total_delta_norm": 0.1854769170765381,
+  "_any_changed": true
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/policy_gradient_loss_properties.json b/test_artifacts/grpo_e2e/20261016_215929/policy_gradient_loss_properties.json
new file mode 100644
index 0000000..4f163ab
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/policy_gradient_loss_properties.json
@@ -0,0 +1,48 @@
+{
+  "tests": [
+    {
+      "name": "positive_advantage_gradient",
+      "logp": -1.0,
+      "advantage": 1.0,
+      "gradient": -1.0,
+      "expected": "negative (optimizer increases logp)",
+      "passed": true
+    },
+    {
+      "name": "negative_advantage_gradient",
+      "logp": -1.0,
+      "advantage": -1.0,
+      "gradient": 1.0,
+      "expected": "positive (optimizer decreases logp)",
+      "passed": true
+    },
+    {
+      "name": "zero_advantage_gradient",
+      "logp": -1.0,
+      "advantage": 0.0,
+      "gradient": -0.0,
+      "expected": "zero",
+      "passed": true
+    },
+    {
+      "name": "clipping_large_ratio",
+      "ratio": 18.17414536944306,
+      "clipped_ratio": 1.2,
+      "loss": -1.2000000476837158,
+      "expected_loss": -1.2,
+      "passed": true
+    },
+    {
+      "name": "batch_consistency",
+      "individual_losses": [
+        -1.0,
+        0.5,
+        -0.0
+      ],
+      "batch_loss": -0.1666666716337204,
+      "expected_batch_mean": -0.16666666666666666,
+      "passed": true
+    }
+  ],
+  "all_passed": true
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0.json
new file mode 100644
index 0000000..384a2d3
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.20, y=0.50)",
+      "action_type": "click",
+      "x": 63,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.20, y=0.55)",
+      "action_type": "click",
+      "x": 63,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.20, y=0.60)",
+      "action_type": "click",
+      "x": 63,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.20, y=0.65)",
+      "action_type": "click",
+      "x": 63,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_2.png
new file mode 100644
index 0000000..9fd3281
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_3.png
new file mode 100644
index 0000000..8770b85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_0_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1.json
new file mode 100644
index 0000000..5047a60
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click",
+      "x": 96,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click",
+      "x": 96,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.30, y=0.60)",
+      "action_type": "click",
+      "x": 96,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.30, y=0.65)",
+      "action_type": "click",
+      "x": 96,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_2.png
new file mode 100644
index 0000000..268454d
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_3.png
new file mode 100644
index 0000000..fcb3dd3
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_1_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2.json
new file mode 100644
index 0000000..bd58ec2
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click",
+      "x": 128,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click",
+      "x": 128,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.40, y=0.60)",
+      "action_type": "click",
+      "x": 128,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.40, y=0.65)",
+      "action_type": "click",
+      "x": 128,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_2.png
new file mode 100644
index 0000000..5934673
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_3.png
new file mode 100644
index 0000000..6babda9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_2_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3.json
new file mode 100644
index 0000000..916eb51
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click",
+      "x": 160,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click",
+      "x": 160,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.50, y=0.60)",
+      "action_type": "click",
+      "x": 160,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.50, y=0.65)",
+      "action_type": "click",
+      "x": 160,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_2.png
new file mode 100644
index 0000000..4bea64d
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_3.png
new file mode 100644
index 0000000..8010e79
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_3_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4.json
new file mode 100644
index 0000000..1cc6982
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 4,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "advantage": 0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click",
+      "x": 192,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click",
+      "x": 192,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.60, y=0.60)",
+      "action_type": "click",
+      "x": 192,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.60, y=0.65)",
+      "action_type": "click",
+      "x": 192,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_0.png
new file mode 100644
index 0000000..05c07b6
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_1.png
new file mode 100644
index 0000000..d0b4238
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_2.png
new file mode 100644
index 0000000..faac2ac
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_3.png
new file mode 100644
index 0000000..2c2d4fc
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_4_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5.json
new file mode 100644
index 0000000..d0f509f
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5.json
@@ -0,0 +1,38 @@
+{
+  "rollout_idx": 5,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "advantage": -0.9999999800000003,
+  "num_steps": 4,
+  "instruction": "Open Notepad and type Hello World",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.70, y=0.50)",
+      "action_type": "click",
+      "x": 224,
+      "y": 120
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.70, y=0.55)",
+      "action_type": "click",
+      "x": 224,
+      "y": 132
+    },
+    {
+      "step_idx": 2,
+      "action_text": "CLICK(x=0.70, y=0.60)",
+      "action_type": "click",
+      "x": 224,
+      "y": 144
+    },
+    {
+      "step_idx": 3,
+      "action_text": "CLICK(x=0.70, y=0.65)",
+      "action_type": "click",
+      "x": 224,
+      "y": 156
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_0.png
new file mode 100644
index 0000000..b3d658c
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_1.png
new file mode 100644
index 0000000..884fcc1
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_2.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_2.png
new file mode 100644
index 0000000..51431ee
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_2.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_3.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_3.png
new file mode 100644
index 0000000..1c6ba7a
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/rollout_5_step_3.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/summary.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/summary.json
new file mode 100644
index 0000000..3ce8547
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_collection/summary.json
@@ -0,0 +1,23 @@
+{
+  "test_name": "test_e2e_rollout_collection_mock",
+  "num_rollouts": 6,
+  "steps_per_rollout": 4,
+  "rewards": [
+    1.0,
+    0.0,
+    1.0,
+    0.0,
+    1.0,
+    0.0
+  ],
+  "advantages": [
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003,
+    0.9999999800000003,
+    -0.9999999800000003
+  ],
+  "reward_mean": 0.5,
+  "duration_s": 0.0335538387298584
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0.json
new file mode 100644
index 0000000..fda167c
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_0_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1.json
new file mode 100644
index 0000000..38c7391
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_1_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2.json
new file mode 100644
index 0000000..bc83a6b
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_2_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3.json
new file mode 100644
index 0000000..407fbef
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3.json
@@ -0,0 +1,20 @@
+{
+  "step": 0,
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_0_rollout_3_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0.json
new file mode 100644
index 0000000..408b611
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 0,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.30, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.30, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_0.png
new file mode 100644
index 0000000..673fea7
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_1.png
new file mode 100644
index 0000000..466b013
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_0_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1.json
new file mode 100644
index 0000000..7eb87d0
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 1,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.40, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.40, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_0.png
new file mode 100644
index 0000000..f5db13f
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_1.png
new file mode 100644
index 0000000..bcf5af9
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_1_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2.json
new file mode 100644
index 0000000..fbd67cc
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 2,
+  "task_id": "mock_task_001",
+  "reward": 1.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.50, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.50, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_0.png
new file mode 100644
index 0000000..71ff70b
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_1.png
new file mode 100644
index 0000000..512ad85
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_2_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3.json b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3.json
new file mode 100644
index 0000000..821c840
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3.json
@@ -0,0 +1,20 @@
+{
+  "step": 1,
+  "rollout_idx": 3,
+  "task_id": "mock_task_001",
+  "reward": 0.0,
+  "num_steps": 2,
+  "instruction": "Click the OK button",
+  "actions": [
+    {
+      "step_idx": 0,
+      "action_text": "CLICK(x=0.60, y=0.50)",
+      "action_type": "click"
+    },
+    {
+      "step_idx": 1,
+      "action_text": "CLICK(x=0.60, y=0.55)",
+      "action_type": "click"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_0.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_0.png
new file mode 100644
index 0000000..0191bdf
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_0.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_1.png b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_1.png
new file mode 100644
index 0000000..da6a005
Binary files /dev/null and b/test_artifacts/grpo_e2e/20261016_215929/rollout_traces/step_1_rollout_3_screenshot_1.png differ
diff --git a/test_artifacts/grpo_e2e/20261016_215929/summary.txt b/test_artifacts/grpo_e2e/20261016_215929/summary.txt
new file mode 100644
index 0000000..464d3a6
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/summary.txt
@@ -0,0 +1,25 @@
+============================================================
+GRPO E2E TEST REPORT: test_e2e_training_loop_mock
+============================================================
+Status:               passed
+Duration:             0.04s
+Training steps:       2
+Rollouts per step:    4
+Total rollouts:       8
+
+--- Model Weight Change ---
+Weights changed:      True
+Total delta L2 norm:  0.185477
+  embedding.weight: delta=0.025376 (norm 91.4769 -> 91.4746)
+  head.weight: delta=0.180939 (norm 9.2112 -> 9.2100)
+  head.bias: delta=0.031921 (norm 1.5798 -> 1.5810)
+
+--- Training Metrics ---
+  Step 0: loss=0.0000 reward_mean=0.50 grad_norm=1.0000 valid_terms=8
+  Step 1: loss=0.0000 reward_mean=0.50 grad_norm=1.0000 valid_terms=8
+
+--- Checkpoint ---
+Checkpoint path:      test_artifacts/grpo_e2e/20261016_215929/checkpoint/step_2
+Checkpoint loadable:  True
+
+============================================================
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/test_report.json b/test_artifacts/grpo_e2e/20261016_215929/test_report.json
new file mode 100644
index 0000000..a187795
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/test_report.json
@@ -0,0 +1,15 @@
+{
+  "test_name": "test_e2e_training_loop_mock",
+  "start_time": "2026-10-16T21:59:29",
+  "status": "passed",
+  "duration_s": 0.036730051040649414,
+  "num_training_steps": 2,
+  "num_rollouts_per_step": 4,
+  "total_rollouts": 8,
+  "model_weights_changed": true,
+  "total_delta_norm": 0.1854769170765381,
+  "checkpoint_loadable": true,
+  "final_loss": 0.0,
+  "final_reward_mean": 0.5,
+  "final_grad_norm": 1.0000000528613704
+}
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/training_log.json b/test_artifacts/grpo_e2e/20261016_215929/training_log.json
new file mode 100644
index 0000000..ba6679a
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/training_log.json
@@ -0,0 +1,34 @@
+[
+  {
+    "step": 0,
+    "loss": 0.0,
+    "reward_mean": 0.5,
+    "reward_std": 0.5,
+    "advantages": [
+      0.9999999800000003,
+      -0.9999999800000003,
+      0.9999999800000003,
+      -0.9999999800000003
+    ],
+    "grad_norm": 1.0000001900372488,
+    "lr": 0.001,
+    "valid_terms": 8,
+    "step_time": 0.01689004898071289
+  },
+  {
+    "step": 1,
+    "loss": 0.0,
+    "reward_mean": 0.5,
+    "reward_std": 0.5,
+    "advantages": [
+      0.9999999800000003,
+      -0.9999999800000003,
+      0.9999999800000003,
+      -0.9999999800000003
+    ],
+    "grad_norm": 1.0000000528613704,
+    "lr": 0.001,
+    "valid_terms": 8,
+    "step_time": 0.015859127044677734
+  }
+]
\ No newline at end of file
diff --git a/test_artifacts/grpo_e2e/20261016_215929/weight_diff/detailed_diff.json b/test_artifacts/grpo_e2e/20261016_215929/weight_diff/detailed_diff.json
new file mode 100644
index 0000000..7467ea5
--- /dev/null
+++ b/test_artifacts/grpo_e2e/20261016_215929/weight_diff/detailed_diff.json
@@ -0,0 +1,51 @@
+{
+  "parameters": {
+    "embedding.weight": {
+      "before_norm": 90.87411499023438,
+      "after_norm": 90.87342071533203,
+      "delta_l2": 0.012669733725488186,
+      "delta_max": 0.0010232925415039062,
+      "delta_mean": 5.214037628320511e-07,
+      "shape": [
+        256,
+        32
+      ],
+      "num_elements": 8192,
+      "grad_norm": 3.2811145782470703,
+      "grad_max": 0.7971603870391846
+    },
+    "head.weight": {
+      "before_norm": 9.232221603393555,
+      "after_norm": 9.232917785644531,
+      "delta_l2": 0.09049731492996216,
+      "delta_max": 0.0010017603635787964,
+      "delta_mean": -0.00018775406351778656,
+      "shape": [
+        256,
+        32
+      ],
+      "num_elements": 8192,
+      "grad_norm": 32.86104965209961,
+      "grad_max": 7.829895496368408
+    },
+    "head.bias": {
+      "before_norm": 1.5759165287017822,
+      "after_norm": 1.576166033744812,
+      "delta_l2": 0.015995899215340614,
+      "delta_max": 0.0010016560554504395,
+      "delta_mean": -7.669492333661765e-06,
+      "shape": [
+        256
+      ],
+      "num_elements": 256,
+      "grad_norm": 3.992352247238159,
+      "grad_max": 2.0116944313049316
+    }
+  },
+  "summary": {
+    "total_params": 16640,
+    "trainable_params": 16640,
+    "changed_param_groups": 3,
+    "total_param_groups": 3
+  }
+}
\ No newline at end of file
diff --git a/tests/test_kaggle_cloud_and_modal_guard.py b/tests/test_kaggle_cloud_and_modal_guard.py
new file mode 100644
index 0000000..fef3c44
--- /dev/null
+++ b/tests/test_kaggle_cloud_and_modal_guard.py
@@ -0,0 +1,88 @@
+"""Kaggle staging and the Modal GPU production guard (no network)."""
+
+from __future__ import annotations
+
+import json
+
+import pytest
+
+from openadapt_ml.cloud import kaggle_cloud
+from openadapt_ml.cloud.modal_gpu_guard import modal_gpu_refusal
+
+
+def test_guard_refuses_by_default():
+    assert "off by default" in modal_gpu_refusal({}, profile="other")
+
+
+@pytest.mark.parametrize("environment", ["", "main", "prod", "Production"])
+def test_guard_refuses_production_environments(environment):
+    env = {"OPENADAPT_ALLOW_MODAL_GPU": "1", "MODAL_ENVIRONMENT": environment}
+    assert "non-production" in modal_gpu_refusal(env, profile="research")
+
+
+def test_guard_refuses_production_workspace():
+    env = {"OPENADAPT_ALLOW_MODAL_GPU": "1", "MODAL_ENVIRONMENT": "gpu-research"}
+    assert "production workspace" in modal_gpu_refusal(env, profile="abrichr")
+    assert modal_gpu_refusal(env, profile="") is not None
+
+
+def test_guard_allows_explicit_nonprod_target():
+    env = {"OPENADAPT_ALLOW_MODAL_GPU": "1", "MODAL_ENVIRONMENT": "gpu-research"}
+    assert modal_gpu_refusal(env, profile="openadapt-research") is None
+
+
+def test_modal_train_cli_refuses_without_opt_in(monkeypatch, tmp_path, capsys):
+    from openadapt_ml.cloud.modal_cloud import cli_main
+
+    monkeypatch.delenv("OPENADAPT_ALLOW_MODAL_GPU", raising=False)
+    rc = cli_main(["train", "--bundle", str(tmp_path)])
+    assert rc == 1
+    assert "Refusing Modal GPU work" in capsys.readouterr().out
+
+
+def test_stage_writes_gpu_kernel_and_dataset(tmp_path):
+    bundle = tmp_path / "bundle"
+    bundle.mkdir()
+    (bundle / "training_data.jsonl").write_text("{}\n")
+    config = tmp_path / "c.yaml"
+    config.write_text("model: x\n")
+    ckpt = tmp_path / "results"
+    ckpt.mkdir()
+    (ckpt / "adapter.bin").write_bytes(b"0")
+    d, k = kaggle_cloud.stage(
+        bundle,
+        config,
+        tmp_path / "out",
+        "someone",
+        "oa-train-data",
+        "oa-train",
+        checkpoint=ckpt,
+    )
+    meta = json.loads((k / "kernel-metadata.json").read_text())
+    assert (
+        meta["enable_gpu"] is True
+        and meta["enable_internet"] is True
+        and meta["is_private"] is True
+    )
+    assert meta["dataset_sources"] == ["someone/oa-train-data"]
+    assert (k / meta["code_file"]).is_file()
+    assert (d / "bundle" / "training_data.jsonl").is_file()
+    assert (d / "checkpoint" / "adapter.bin").is_file()
+    assert (
+        json.loads((d / "dataset-metadata.json").read_text())["id"]
+        == "someone/oa-train-data"
+    )
+
+
+def test_stage_rejects_bad_slug_and_missing_bundle(tmp_path):
+    with pytest.raises(ValueError):
+        kaggle_cloud.kernel_metadata("u", "Bad Slug", "oa-train-data", "NvidiaTeslaT4")
+    with pytest.raises(FileNotFoundError):
+        kaggle_cloud.stage(
+            tmp_path,
+            tmp_path / "c.yaml",
+            tmp_path / "o",
+            "u",
+            "oa-train-data",
+            "oa-train",
+        )
diff --git a/tests/test_modal_cloud.py b/tests/test_modal_cloud.py
index ff2e274..73c91b0 100644
--- a/tests/test_modal_cloud.py
+++ b/tests/test_modal_cloud.py
@@ -9,6 +9,16 @@
 
 import pytest
 
+
+@pytest.fixture(autouse=True)
+def _allow_modal_gpu_in_mocked_tests():
+    """These tests mock Modal; the production GPU guard is tested separately."""
+    with (
+        patch("openadapt_ml.cloud.modal_cloud.modal_gpu_refusal", return_value=None),
+        patch("openadapt_ml.cloud.modal_cloud.require_nonprod_modal_gpu"),
+    ):
+        yield
+
 # ---------------------------------------------------------------------------
 # CLI argument parsing tests
 # ---------------------------------------------------------------------------