diff --git a/.buildkite/pipeline.yaml b/.buildkite/pipeline.yaml
index c287cae..1887dc8 100644
--- a/.buildkite/pipeline.yaml
+++ b/.buildkite/pipeline.yaml
@@ -1,4 +1,12 @@
 steps:
+  - label: ":lock: lint workloads"
+    agents:
+      queue: small_cpu_queue_premerge
+    commands:
+      - python3 -m ensurepip --upgrade --default-pip 2>/dev/null || true
+      - python3 -m pip install --user pyyaml
+      - export PATH="$HOME/.local/bin:$PATH" && python3 scripts/lint_workload.py --all
+
   - label: ":pipeline: generate steps"
     agents:
       queue: small_cpu_queue_premerge
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
new file mode 100644
index 0000000..43331c5
--- /dev/null
+++ b/.pre-commit-config.yaml
@@ -0,0 +1,9 @@
+repos:
+  - repo: local
+    hooks:
+      - id: lint-workload
+        name: lint workload recipes
+        entry: python3 scripts/lint_workload.py
+        language: system
+        files: ^workloads/.*\.yaml$
+        pass_filenames: true
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..2ccb1f3
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,73 @@
+# Contributing
+
+How to add or change a workload recipe without breaking the nightly.
+
+## Setup
+
+Install pre-commit so recipes are linted on commit:
+
+```bash
+pip install pre-commit
+pre-commit install
+```
+
+## Add or change a recipe
+
+1. Copy an existing workload that targets the same GPU (e.g.
+   `workloads/minimax_m3_b200.yaml` for B200) into `workloads/<model>_<hw>.yaml`.
+2. Edit the fields for your model and tasks. The recipe schema is documented in
+   [README.md](./README.md). Set `nightly: true` to include it in the nightly
+   schedule; leave it off for opt-in recipes.
+3. Lint locally (see below), open a PR, and validate on hardware before merge.
+
+## Lint (local + CI)
+
+`scripts/lint_workload.py` runs the same structural checks the pipeline relies
+on, with no GPU and no model download:
+
+- the recipe passes `lib/parse_workload.py` (required fields, known GPU profile,
+  well-formed `lm_eval` / `vllm_bench` / `bfcl` blocks, and, when `lm_eval` is
+  installed, task names against the registry);
+- `.buildkite/generate_pipeline.py` emits a buildkite step for it.
+
+```bash
+scripts/lint_workload.py workloads/<your>_<hw>.yaml   # one or more paths
+scripts/lint_workload.py --changed                    # recipes changed vs origin/main
+scripts/lint_workload.py --all                        # every recipe
+```
+
+This is what catches the "commit garbage" cases: a typo'd field, an unknown
+GPU, a malformed `serve_args`, or a task that generates no step.
+
+## What runs on your PR
+
+- **DCO**: every commit needs a `Signed-off-by` line (`git commit -s`).
+- **lint workloads**: the buildkite pipeline runs `lint_workload.py --all` on
+  every PR and fails the build if any recipe is structurally invalid.
+
+The lint does **not** prove accuracy. It only proves the recipe is well-formed
+and will produce a runnable step.
+
+## Validate on hardware (before merge)
+
+Prove the recipe actually serves and scores by triggering a build scoped to
+just your recipe(s), against a known-good vLLM image:
+
+```bash
+bk build create \
+  --pipeline perf-eval \
+  --commit "<your perf-eval branch SHA>" \
+  --branch "<your branch>" \
+  --env "VLLM_IMAGE=<full image URI>" \
+  --env "VLLM_COMMIT=<vLLM SHA>" \
+  --env "WORKLOADS=workloads/<your>_<hw>.yaml"
+```
+
+Link the passing build in your PR so reviewers can see it ran on the target GPU.
+A build takes ~30-90 minutes; GPU queues are shared, so don't trigger duplicate
+builds for the same commit.
+
+## AI assistance
+
+See [CLAUDE.md](./CLAUDE.md) for the disclosure convention (PR body line plus a
+commit trailer for non-trivial agent-authored changes).
diff --git a/lib/parse_workload.py b/lib/parse_workload.py
index 076aec8..4472b29 100644
--- a/lib/parse_workload.py
+++ b/lib/parse_workload.py
@@ -162,7 +162,7 @@ def precision_from_model(model: str) -> str:
 def validate_tasks(tasks: list, path: str) -> None:
     if not tasks:
         sys.exit(f"{path}: missing or empty `lm_eval.tasks`")
-    skip_registry = env_truthy("BENCH_ONLY")
+    skip_registry = env_truthy("BENCH_ONLY") or env_truthy("SKIP_TASK_REGISTRY")
     known = set() if skip_registry else known_task_names()
     for t in tasks:
         extra = set(t) - TASK_FIELDS
diff --git a/scripts/lint_workload.py b/scripts/lint_workload.py
new file mode 100755
index 0000000..fafb6c3
--- /dev/null
+++ b/scripts/lint_workload.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Lint workload recipes before merge -- no GPU, no model download required.
+
+Runs the same structural checks the nightly pipeline relies on, so a broken
+recipe fails in review instead of at 3am on a B200:
+
+  1. ``lib/parse_workload.py`` accepts the recipe -- required fields, known GPU
+     profile, well-formed lm_eval / vllm_bench / bfcl blocks, and (when
+     ``lm_eval`` is importable) task names against the registry.
+  2. ``.buildkite/generate_pipeline.py`` emits a buildkite step for it.
+
+This is the automatic gate; it does NOT prove accuracy. The real proof is a
+scoped hardware run -- see CONTRIBUTING.md ("Validate on hardware").
+
+Usage:
+  scripts/lint_workload.py workloads/foo_b200.yaml [more.yaml ...]
+  scripts/lint_workload.py --changed   # recipes added/changed vs origin/main
+  scripts/lint_workload.py --all       # every workloads/*.yaml
+"""
+
+import argparse
+import glob
+import os
+import subprocess
+import sys
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
+def _run(cmd, extra_env=None):
+    env = dict(os.environ)
+    if extra_env:
+        env.update(extra_env)
+    return subprocess.run(
+        cmd, cwd=ROOT, env=env, capture_output=True, text=True
+    )
+
+
+def _lm_eval_available():
+    return _run([sys.executable, "-c", "import lm_eval"]).returncode == 0
+
+
+def lint_one(path, skip_registry):
+    parse_env = {"SKIP_TASK_REGISTRY": "1"} if skip_registry else {}
+    parsed = _run([sys.executable, "lib/parse_workload.py", path], parse_env)
+    if parsed.returncode != 0:
+        return False, (parsed.stderr or parsed.stdout).strip()
+
+    gen = _run(
+        [sys.executable, ".buildkite/generate_pipeline.py"],
+        {"WORKLOADS": path, "VLLM_IMAGE": "vllm/vllm-openai:lint"},
+    )
+    if gen.returncode != 0:
+        return False, "generate_pipeline: " + gen.stderr.strip()
+    if "label:" not in gen.stdout:
+        return False, "generate_pipeline emitted no step for this recipe"
+    return True, "ok"
+
+
+def changed_workloads():
+    base = os.environ.get("LINT_BASE_REF", "origin/main")
+    diff = _run(
+        ["git", "diff", "--name-only", "--diff-filter=d", base, "--", "workloads/"]
+    )
+    return [ln for ln in diff.stdout.splitlines() if ln.endswith(".yaml")]
+
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    ap.add_argument("paths", nargs="*", help="workload YAML paths to lint")
+    ap.add_argument(
+        "--changed", action="store_true", help="lint recipes changed vs origin/main"
+    )
+    ap.add_argument("--all", action="store_true", help="lint every workloads/*.yaml")
+    args = ap.parse_args()
+
+    os.chdir(ROOT)
+    if args.all:
+        paths = sorted(glob.glob("workloads/*.yaml"))
+    elif args.changed:
+        paths = changed_workloads()
+    else:
+        paths = args.paths
+    if not paths:
+        print("lint_workload: no workloads to lint")
+        return 0
+
+    skip_registry = not _lm_eval_available()
+    if skip_registry:
+        print("lint_workload: lm_eval not importable; skipping task-name validation")
+
+    failed = 0
+    for path in paths:
+        ok, msg = lint_one(path, skip_registry)
+        print(f"{'PASS' if ok else 'FAIL'}  {path}")
+        if not ok:
+            print(f"      {msg}")
+            failed += 1
+    print(f"lint_workload: {len(paths) - failed}/{len(paths)} passed")
+    return 1 if failed else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())