# zicula-ai-eval-kit — free sample

Three real files from the full kit (19 files), each **trimmed to taste** —
enough to judge the quality, not enough to gate a prompt:

- `golden-set-preview.csv` — 2 of the 5 demo cases from the kit's working
  sample project: the refund-policy case (context + combined assertions +
  `refusal-handling` judge rubric) and the Japanese case
  (`multilingual-consistency`). Copy the pattern, replace the rows, and you
  have real golden-set cases. The full kit adds 3 more demo cases (format,
  greeting, German warranty), stored offline responses and a baseline run —
  so `python3 eval-runner.py --offline` produces a pass/fail report with a
  diff on the first run, no API key.
- `judge-rubrics-preview.md` — 2 of the 5 built-in judge rubrics
  (`refusal-handling`, `multilingual-consistency`) with their complete 1–5
  anchors, plus how the judge call works. The full file adds the other three
  rubrics (`faithfulness`, `tone`, `format-compliance`), custom inline rubric
  writing, and the honest limits of LLM-as-judge.
- `playbook-14d-preview.md` — Phase 1 (Days 1–2) of the 14-day playbook,
  complete with the "done when" gates, plus one-line summaries of Phases 2–4
  (golden set, CI gate, shipping ritual). The full playbook is 14 days,
  day by day.

What the full kit adds (the other 16 files):

- `eval-runner.py` — the product: runs your cases against any
  OpenAI-compatible endpoint (OpenAI, Ollama, LM Studio, vLLM, ...), scores
  deterministic assertions + LLM-as-judge, writes JSON + Markdown reports
  with run-to-run diffs. Python 3 stdlib only, no pip, no account.
- `golden-set-template.csv` + the full sample project (5 cases, stored
  responses, baseline where the Japanese case fails — the diff section shows
  `fixed: ja-share-01` on your first offline run).
- `multilingual-eval-guide.md` — the EN/JA/PT-BR/DE mirror-case method and
  the per-language traps that catch silent regressions.
- `github-actions-gate.yml` — the CI workflow that fails a PR when the eval
  score drops, with an offline no-key smoke job.
- `README-buyer.md`, `LICENSE.txt` — quickstart (EN/JA/DE) and the single
  developer/team commercial license.
- `screenshots/` — rendered previews of the quickstart, golden-set template,
  judge rubrics, multilingual guide and playbook (the gallery on the product
  page shows these).

One-time purchase $39 · free updates for v1.x · 30-day full refund.