diff --git a/AGENTS.md b/AGENTS.md
index 46c9344..a2b6b02 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -13,3 +13,17 @@ Training, eval, and vLLM-launch scripts write reproducibility artifacts so runs
 **Always pass `--provenance-dir` when running `scripts/launch_vllm.py`** — point it at the associated run's output/checkpoint directory (e.g. `--provenance-dir <output_dir>`) so the vLLM provenance is co-located with the training/eval it belongs to.
 
 When publishing a model or eval results (HuggingFace, GitHub comments, etc.), always include the provenance artifacts (`train_command.txt`, `eval_command.txt`, `vllm_command.txt`, `checkpoint_sha256.txt`, `drafter_checkpoint_sha256.txt`, patches) so the run can be reproduced.
+
+## Agent skills
+
+### Issue tracker
+
+Issues and specs are tracked in GitHub Issues in `rahul-tuli/speculators`; use `gh --repo rahul-tuli/speculators`. See `docs/agents/issue-tracker.md`.
+
+### Triage labels
+
+Use the default triage labels: `needs-triage`, `needs-info`, `ready-for-agent`, `ready-for-human`, and `wontfix`. See `docs/agents/triage-labels.md`.
+
+### Domain docs
+
+This is a single-context repo using root `CONTEXT.md` and `docs/adr/`. See `docs/agents/domain.md`.
diff --git a/scripts/evaluate/evaluate.py b/scripts/evaluate/evaluate.py
index f121402..14a6dea 100644
--- a/scripts/evaluate/evaluate.py
+++ b/scripts/evaluate/evaluate.py
@@ -263,14 +263,11 @@ def _run_subset(
         logger.warning("[%s] No speculative decoding metrics found", subset)
 
     if is_sweep:
-        rows = parse_sweep_results(
-            run_output,
-            spec if has_spec else None,
-        )
+        # Counters bracket the entire sweep, not individual load points.
+        rows = parse_sweep_results(run_output)
         if rows:
             if perf_csv is None:
-                acc_cols = acceptance_csv_columns(spec) if has_spec else []
-                cols = BASE_CSV_COLUMNS + acc_cols
+                cols = BASE_CSV_COLUMNS
                 perf_csv = CsvWriter(
                     output_dir / "perf_results.csv",
                     cols,
@@ -357,7 +354,7 @@ def run_benchmark(args: argparse.Namespace) -> None:
         if mt is not None:
             all_max_tokens[label] = mt
 
-    if acceptance_csv is None:
+    if acceptance_csv is None and not args.target_only:
         logger.error("No acceptance metrics collected from any subset")
         sys.exit(1)
 
@@ -389,12 +386,19 @@ def main() -> None:
     )
     parser.add_argument(
         "mode",
-        choices=["throughput", "sweep"],
+        choices=["throughput", "sweep", "point"],
         help=(
             "throughput: max-rate run for acceptance rates; "
             "sweep: full benchmarking pipeline"
         ),
     )
+    parser.add_argument("--target-only", action="store_true")
+    parser.add_argument(
+        "--identity", type=Path, help="Immutable point run identity JSON"
+    )
+    parser.add_argument("--warmup-requests", type=int, default=2)
+    parser.add_argument("--max-tokens", type=int, default=4096)
+    parser.add_argument("--eval-seed", type=int, default=0)
     parser.add_argument(
         "--target",
         required=True,
@@ -470,7 +474,12 @@ def main() -> None:
         else:
             args.output_dir = f"perf_results_{timestamp}"
 
-    run_benchmark(args)
+    if args.mode == "point":
+        from point import run_point  # noqa: PLC0415
+
+        run_point(args)
+    else:
+        run_benchmark(args)
 
 
 if __name__ == "__main__":
diff --git a/scripts/evaluate/perf_utils.py b/scripts/evaluate/perf_utils.py
index cc2001a..f6048d6 100644
--- a/scripts/evaluate/perf_utils.py
+++ b/scripts/evaluate/perf_utils.py
@@ -567,7 +567,7 @@ def run_guidellm(
         data = f"kind=huggingface,source={dataset}"
         data += f",load_kwargs.data_files={subset}.jsonl,load_kwargs.split=train"
     else:
-        data = f"kind=json_file,path={dataset}"
+        data = f"kind=json_file,path={dataset},load_kwargs.split=train"
     cmd.extend(["--data", data])
 
     cmd.extend(["--data-column-mapper", data_column_mapper])
diff --git a/scripts/launch_vllm.py b/scripts/launch_vllm.py
index 47eb77b..2bafbbd 100644
--- a/scripts/launch_vllm.py
+++ b/scripts/launch_vllm.py
@@ -244,8 +244,8 @@ def parse_args():
     eval_parser.add_argument(
         "--spec-model",
         type=str,
-        required=True,
-        help="Drafter model name or path",
+        default=None,
+        help="Drafter model name or path (omit for target-only evaluation)",
     )
     eval_parser.add_argument(
         "--spec-tokens",
@@ -503,9 +503,11 @@ def _build_eval_cmd(args, vllm_args):
         "vllm.entrypoints.cli.main",
         "serve",
         args.model,
-        "--spec-model",
-        args.spec_model,
     ]
+    if args.spec_model is not None:
+        cmd.extend(["--spec-model", args.spec_model])
+    elif args.spec_tokens is not None or args.spec_method is not None:
+        raise ValueError("--spec-tokens/--spec-method require --spec-model")
     if args.spec_tokens is not None:
         cmd.extend(["--spec-tokens", str(args.spec_tokens)])
     if args.spec_method is not None:
diff --git a/tests/e2e/smoke/test_finetuning_sanity.py b/tests/e2e/smoke/test_finetuning_sanity.py
index 4a13150..fcc93a1 100644
--- a/tests/e2e/smoke/test_finetuning_sanity.py
+++ b/tests/e2e/smoke/test_finetuning_sanity.py
@@ -42,6 +42,7 @@ def test_finetuning_weight_sanity(tmp_path: Path):
     # at least MIN_CHANGED tensors must have rel_l1 > REL_L1_MIN
     # to ensure weights actually changed.
     LR = "1e-5"
+    MUON_LR = "1e-4"
     REL_L1_MAX = 0.05
     REL_L1_MIN = 1e-4
     MIN_CHANGED = 3
@@ -93,6 +94,8 @@ def test_finetuning_weight_sanity(tmp_path: Path):
             "2",
             "--lr",
             LR,
+            "--muon-lr",
+            MUON_LR,
             "--total-seq-len",
             "2048",
             "--num-workers",
diff --git a/tests/unit/cli/test_cli.py b/tests/unit/cli/test_cli.py
index fbc2559..ebecb73 100644
--- a/tests/unit/cli/test_cli.py
+++ b/tests/unit/cli/test_cli.py
@@ -5,7 +5,9 @@ from typer.testing import CliRunner
 
 from speculators.cli import app
 
-runner = CliRunner()
+# Rich uses COLUMNS when rendering help, so keep assertions independent of the
+# terminal width provided by the environment running the tests.
+runner = CliRunner(env={"COLUMNS": "200"})
 
 
 def unstyled_output(result):
diff --git a/tests/unit/scripts/test_launch_vllm.py b/tests/unit/scripts/test_launch_vllm.py
index de34b5c..b751480 100644
--- a/tests/unit/scripts/test_launch_vllm.py
+++ b/tests/unit/scripts/test_launch_vllm.py
@@ -1,7 +1,9 @@
 import os
+from argparse import Namespace
 
 from scripts.launch_vllm import (
     DEFAULT_RENDERER_NUM_WORKERS,
+    _build_eval_cmd,
     _enable_scale_out_endpoints,
     _preprocessing_workers,
     _set_render_thread_defaults,
@@ -87,3 +89,12 @@ def test_sizing_uses_one_combined_budget():
     assert preprocessing_workers == 30
     assert api_servers == 7
     assert preprocessing_workers * 3 + api_servers * 4 <= int(160 * 0.75)
+
+
+def test_target_only_eval_omits_drafter_flags():
+    args = Namespace(
+        model="target", spec_model=None, spec_tokens=None, spec_method=None
+    )
+    cmd = _build_eval_cmd(args, ["--dtype", "bfloat16"])
+    assert cmd[-3:] == ["target", "--dtype", "bfloat16"]
+    assert "--spec-model" not in cmd