diff --git a/launches/auto-nanogpt-extra-instructions.md b/launches/auto-nanogpt-extra-instructions.md
new file mode 100644
index 000000000..41be53559
--- /dev/null
+++ b/launches/auto-nanogpt-extra-instructions.md
@@ -0,0 +1,61 @@
+# Auto-nanoGPT Launch Extra Instructions
+
+This target is not a physical AI, CFD, or surrogate-modelling task. When the
+generic Senpai system prompt mentions physical metrics, CFD, datasets, or
+architecture search for physical modelling, reinterpret that guidance as:
+optimize the fixed `modded-nanogpt` track 3 optimizer benchmark described in
+`$PROBLEM_DIR/program.md`.
+
+Keep the task close to the public `modded-nanogpt` track 3 benchmark:
+
+- The objective is lower optimizer steps to FineWeb validation loss below 3.28.
+- Keep data, batch size, model architecture, and one forward-backward pass per
+  optimizer step fixed.
+- Focus on optimizer algorithms, optimizer hyperparameters, schedules, and
+  initialization.
+- Do not invent extra data configs, data files, or benchmark rules beyond the
+  target repo's `program.md` and the official track 3 README.
+
+Keep the research portfolio balanced between exploitation and exploration.
+Retuning LR/WD/cooldown is important when giving a new method a fair shot, but
+do not let the run become mostly scalar hyperparameter search. Assign fresh
+optimizer mechanisms, preconditioners, schedule ideas, initialization ideas,
+and pruning ablations of complex stacks.
+
+Be deliberate about step budgets. The launch timeout and max-epoch values are
+hard ceilings, not instructions to run forever. Use tiny runs for smoke tests,
+shorter step budgets for uncertain optimizer screening, and longer predeclared
+seed batches for serious confirmation. Think clearly before each PR about
+whether the run is exploration, tuning, pruning, or confirmation.
+
+Early kill gates are encouraged for obvious crashes, non-finite loss, exploding
+gradients, or hopeless screening runs, but do not use per-run validation loss to
+cherry-pick final steps or seeds. Final claims must report all non-cherry-picked
+runs at a predeclared step count and evaluate the benchmark statistical rule.
+
+Use W&B aggressively. Preserve and extend the starter script's telemetry for
+losses, validation loss, steps-to-target, learning rates, weight decay, gradient
+norms and distributions, parameter norms and distributions, and any
+optimizer-specific diagnostics that help explain why an idea worked or failed.
+The starter script also logs trailing-window train-loss slopes every 10% of the
+run and at the final step; use these slopes to reason about curve shape without
+overreacting to individual noisy steps.
+
+Focus only on your assigned advisor branch, research tag, student list, PR
+stream, and W&B runs. Ignore work that is not labeled with your branch/tag. Use
+the benchmark snapshot checked into this target repo; do not refresh, browse,
+fetch, or mine new upstream PRs, branches, records, issues, or post-launch
+updates during this run.
+
+Prime Intellect's autonomous-run materials are explicitly banned sources for
+agents during this launch. Do not open, fetch, browse, search within, clone,
+cite, summarize, or use:
+
+- `https://www.primeintellect.ai/auto-nanogpt`
+- `https://github.com/PrimeIntellect-ai/experiments-autonomous-speedrunning`
+- any raw GitHub URLs, files, branches, issues, pull requests, or archives under
+  that repository
+
+Those links are named only so you know what not to read. They are comparison
+artifacts for humans after the run, not part of the active experimental
+context.
diff --git a/launches/launch-auto-nanogpt.sh b/launches/launch-auto-nanogpt.sh
new file mode 100755
index 000000000..7e6a77565
--- /dev/null
+++ b/launches/launch-auto-nanogpt.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+target_repo_url="https://github.com/morganmcg1/modded-nanogpt-senpai.git"
+target_repo_branch="senpai-launch-20260515"
+extra_instructions="launches/auto-nanogpt-extra-instructions.md"
+wandb_entity="wandb-applied-ai-team"
+wandb_project="modded-nanogpt-senpai"
+timeout_minutes="30240"
+max_epochs="100000"
+
+env_file="${SENPAI_ENV_FILE:-}"
+if [[ -z "${env_file}" ]]; then
+  if [[ -f ".env" ]]; then
+    env_file=".env"
+  elif [[ -f "${HOME}/ML/senpai/.env" ]]; then
+    env_file="${HOME}/ML/senpai/.env"
+  fi
+fi
+if [[ -n "${env_file}" ]]; then
+  set -a
+  # shellcheck disable=SC1090
+  source "${env_file}"
+  set +a
+fi
+
+if [[ -n "${SENPAI_PYTHON:-}" ]]; then
+  launch_cmd=("${SENPAI_PYTHON}")
+elif python -c "import simple_parsing" >/dev/null 2>&1; then
+  launch_cmd=(python)
+elif command -v uv >/dev/null 2>&1; then
+  launch_cmd=(uv run python)
+else
+  echo "Could not find Python with simple_parsing. Install the Senpai deps or set SENPAI_PYTHON." >&2
+  exit 1
+fi
+
+for rep in 1 2 3 4 5; do
+  tag="auto-nanogpt-1gpu-r${rep}"
+  "${launch_cmd[@]}" k8s/launch.py \
+    --tag "${tag}" \
+    --advisor \
+    --target_repo_url "${target_repo_url}" \
+    --target_repo_branch "${target_repo_branch}" \
+    --advisor_branch "${tag}" \
+    --gh_history_scope fresh \
+    --n_students 8 \
+    --student_prefix "g1r${rep}" \
+    --gpus_per_student 1 \
+    --timeout_minutes "${timeout_minutes}" \
+    --max_epochs "${max_epochs}" \
+    --wandb_entity "${wandb_entity}" \
+    --wandb_project "${wandb_project}" \
+    --extra_instructions "${extra_instructions}" \
+    "$@"
+done