SYNTHETIC TEST LAB · VERSION 0.2.1
Nellie QA Bench
Inspect the evidence. Keep uncertainty visible.
1. Prepare a run
Use the blind packet to collect answers. This bench has no model connection, email transport or persistent storage. Public demo: submit synthetic, redacted data only. Evaluation sends your input and optional grading to Cloudflare for in-memory processing. This app does not persist submitted content, but Cloudflare may process request metadata. Never submit private responses, secrets or confidential grading here.
Download blind packet · Download labeled one-row demoBlind packet — 30 cases: 10 factual · 6 research · 5 logistics · 5 email · 4 ambiguous. The labeled demonstration contains one row (QA2-001).
2. Import answers
JSONL: one object per case. Required: case_id, answer_text, run_timestamp, trace (null or array). Maximum 256 KiB / 30 rows. See the source README for structured claims and evidence fields.
Without a private grading configuration, factual claims remain unassessed. Keep confidential grading local; do not submit it to this public demo. Use response_provenance: external-assistant for independently authored assistant answers; demo-authored is reserved for canned examples.
3. Review findings
No overall score. A fixture match says nothing about real-world truth. Missing evidence remains unassessed.