chemistry-ablations-and-curves-without-an-accelerator

# Ablations and curves when the sandbox has no accelerator **Price the training arm first.** `train-the-named-architecture` is the primary instruction: when the brief names an architecture, a small honest re-implementation trained at whatever scale you can afford answers every architecture-shaped criterion. So before this skill applies, measure the machine — device query, cores, RAM, wall clock left — and price one scaled-down training arm against it, in writing. What follows sits **on top of** that arm wherever it is affordable, and replaces it only where there is genuinely no accelerator and no time. ## What goes wrong The task names an architecture by its parts. The run uses a released implementation, decides that removing a part means retraining, prices retraining, finds it unaffordable, and records the ablation as *not attempted*. The report then describes the parts in prose and labels the description as not a measurement made in this run. Nothing in the run varies the thing the task is about, and every prediction it makes is made at one identical configuration. Two habits make the omission permanent: - **The ablation serves no hypothesis, so the descope ladder cuts it first.** Arms that buy breadth across targets survive; the one arm that would show which part carries the result is the cheapest thing to drop and the first thing dropped. - **A slot is refused because it has no content yet.** Declining to reserve a table row or a figure slot for an arm you have...