## Why This Capsule Exists Most benchmarks measure what a model *can* generate. This capsule measures the opposite: whether a model can **generate nothing more and nothing less than what was asked**. No creativity, no reasoning, no ambiguity — just strict format compliance on the simplest possible content (an alphabet every model already "knows" perfectly). Any deviation is a pure instruction-following failure, not a knowledge or capability gap. ## What This Tests | Constraint | Capability Probed | |---|---| | Exactly 26 letters, a through z | Basic enumeration accuracy (no missing/duplicate/reordered letters) | | Comma-separated, no spaces | Precise formatting compliance down to whitespace | | No line breaks | Single-line output discipline | | No explanation | Resistance to the model's default "helpful assistant" chattiness | | No markdown formatting | Suppressing default stylistic habits (bold, code blocks, lists) | | "Nothing before or after" | Total instruction obedience — no preamble, no sign-off, no meta-commentary | ## Expected Results - **This is the closest thing to a "should be 100%" capsule** — the content itself requires zero knowledge or reasoning. Any failure is purely about instruction-following, not capability. - **The most common failure mode is not wrong content, it's wrong wrapper**: models adding a preamble ("Sure, here's the alphabet:"), wrapping the output in a code block, or appending a closing remark ("Let me know if you need anything else!") despite explicit instructions not to. - **A secondary failure mode is subtle format drift**: a stray space after a comma, a trailing comma, or a trailing newline — easy to miss visually but a clean pass/fail for automated grading. - **This capsule is a strong regression indicator over time**: if a model family becomes "chattier" or more safety-hedged in later versions, this is one of the fastest capsules to catch it, since the correct answer never changes and there is no legitimate reason to add anything extra. ## How to Read a Run Grading is binary per sub-check, not subjective — this capsule is designed to be gradable programmatically (exact string match against the expected output), unlike most capsules in the set. Any output that isn't byte-for-byte `a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x,y,z` fails at least one criterion below. ### Scoring Checklist - [ ] Contains all 26 letters, each exactly once - [ ] Letters are in correct a–z order - [ ] Comma-separated with no spaces - [ ] Single line, no line breaks - [ ] No explanatory text before or after - [ ] No markdown formatting (no code block, no bold, no list)