Protected collaborator review · hard-200 upgrade v1
200 measured-hard seeds, expanded to long-horizon research tasks
Each of the 200 high-value seed requests (0–12.5% measured pass rate) was processed through the same pipeline that produced the 179 corpus: construct preserved, request rewritten as a natural decision-led long-horizon task, clause partition, prompt-bound rubric atoms, judge rubric, and live public-web feasibility probes. Seed pass rates describe the seeds; every final task requires re-measurement.
378 tasks (both corpora)3515 clauses3924 rubric atoms135 feasible243 explicit shortfalls100 pass the quality gate0 human acceptedneeds re-measurement
| dimension | seed mean | final mean |
|---|---|---|
| naturalness | 2.94 | 3.63 |
| quality | 2.71 | 3.72 |
| depth | 3.16 | 4.72 |
| breadth | 2.96 | 4.52 |
| verification_level | 3.83 | 4.65 |
1 · Blind prompt review
Final tasks only — no seeds, rubrics, or scores. Judge naturalness,
scope, and completion boundaries unbiased.Start here →
2 · Full seed-vs-final review
Side-by-side seed and final, measured seed stats, scorecards, clauses,
bound rubric atoms, source packets, and feasibility dispositions.
Open full corpus →
3 · Annotate tasks
Review every prompt and rubric, edit proposed language, record CMU or
Browser-Use decisions side by side, and export a mergeable annotation file.
Open annotation workspace →
Status boundary: internal QA artifact for
human review — not a scored benchmark release, and no task here has human
acceptance yet. Source seed file remains unmodified.