[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"capsule:91e1cd9f-f12c-4a90-b146-3a46d87b980c":3},["Reactive",4],{"id":5,"user_id":6,"title":7,"prompt_content":8,"prompt_type":9,"is_locked":10,"is_public":10,"fork_of":11,"created_at":12,"updated_at":13,"category":14,"brief":15,"attachments":16,"expected_markers":17,"user":19},"91e1cd9f-f12c-4a90-b146-3a46d87b980c","695c8e6c-9654-4858-b2d3-925adefacc35","Token Precision Test","Output only the 26 lowercase letters of the English alphabet, comma-separated, no spaces, no line breaks, no explanation, no markdown formatting, nothing before or after. Example format: a,b,c,d","text",true,null,"2026-07-03T02:18:15.269673+00:00","2026-07-03T02:19:35.993222+00:00","coding","## Why This Capsule Exists\n\nMost 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.\n\n## What This Tests\n\n| Constraint | Capability Probed |\n|---|---|\n| Exactly 26 letters, a through z | Basic enumeration accuracy (no missing\u002Fduplicate\u002Freordered letters) |\n| Comma-separated, no spaces | Precise formatting compliance down to whitespace |\n| No line breaks | Single-line output discipline |\n| No explanation | Resistance to the model's default \"helpful assistant\" chattiness |\n| No markdown formatting | Suppressing default stylistic habits (bold, code blocks, lists) |\n| \"Nothing before or after\" | Total instruction obedience — no preamble, no sign-off, no meta-commentary |\n\n## Expected Results\n\n- **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.\n- **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.\n- **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\u002Ffail for automated grading.\n- **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.\n\n## How to Read a Run\n\nGrading 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.\n\n### Scoring Checklist\n- [ ] Contains all 26 letters, each exactly once\n- [ ] Letters are in correct a–z order\n- [ ] Comma-separated with no spaces\n- [ ] Single line, no line breaks\n- [ ] No explanatory text before or after\n- [ ] No markdown formatting (no code block, no bold, no list)",[],{"items":18,"matchAll":10},[],{"username":20,"avatar_url":21},"Ezarwebmaster","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F42354978?v=4"]