autoresearch-planlisted

# Bridge: Autoresearch (Plan-Level) **Announce at start:** "I'm using the bridge:autoresearch-plan skill to experimentally compare candidate approaches before we commit to one." ## Purpose karpathy/autoresearch's actual design (an agent iterating on `train.py`, fixed 5-minute budget, single metric `val_bpb`, keep-or-discard) works because three things are nailed down before any iteration starts: the **scope** of what varies, a **fixed budget**, and a **single, directional metric**. This skill exists to enforce that same discipline at the plan stage — after a plan has been shaped (e.g. by gstack's `/autoplan`) but before it's frozen into an implementation spec. Its job is not to write code; it's to spend a small, bounded amount of effort testing 2+ real candidate approaches against real data, so the plan that reaches `writing-plans` / `gstack-to-plan` encodes a validated decision instead of a guess. Do not run this skill on a plan that only has one viable approach — there's nothing to compare, and manufacturing fake alternatives to satisfy this skill defeats its purpose. Skip straight to the next pipeline step in that case and say so. ## Step 1 — Locate the plan and candidate approaches Find the source plan using the same search order as `gstack-to-plan` Step 1 (explicit path arg → `docs/plan.md` → `docs/spec.md` → most recently modified `.md` under `docs/` or `.gstack/`). Read it in full. Look for places where the plan already names more than one option (an "alternativ