NutriSteppe-AI — LLM Orchestration Verification

1 Orchestration prompt under test

System prompt sent to the model — replace this with your production prompt to verify the real system.

Model (orchestration)

Judge model (scores subjective checks)

Anthropic API key (kept in memory only — never saved or sent anywhere except api.anthropic.com)

2 Results

—

Checks passed

—

Critical violations

release blockers

—

High-severity fails

needs review

—

Release gate

not run

Gate logic: any failed critical case blocks release; any failed high case flags review; otherwise pass. Subjective checks are scored by a judge model and can be overridden by a human reviewer inside each case.

3 Test cases

ID	Category	Severity	Scenario	Status

4 Repository integration — GitHub or GitLab

GitHub

GitLab

API base URL

Repository (owner/repo)

Branch

Dataset file path

Report file path

Personal access token (fine-grained, Contents: Read & write on this repo only — kept in memory only)

Instance URL (gitlab.com or self-hosted)

Project path or numeric ID

Branch

Dataset file path

Report file path

Personal / project access token (scope: api or read_repository + write_repository — kept in memory only)

Self-hosted GitLab: the browser calls the REST API directly, so the instance must allow CORS from wherever this page is hosted (or you host this page on the same instance via GitLab Pages). See the setup instructions for the exact config.

Import local JSON file

NutriSteppe-AI verification harness v1.1 · standalone build with dual GitHub/GitLab repository support · case data ships as nutristeppe_verification_dataset_v1.json