Concepts

LaunchGate is built around a simple hierarchy that maps to how teams think about AI quality.

Architecture overview


Organisation
  └── Project
        └── Eval Suite (pass threshold: 0.8)
              ├── Eval Case 1 (scorer: exact_match, threshold: 1.0)
              ├── Eval Case 2 (scorer: contains, threshold: 1.0)
              └── Eval Case 3 (scorer: llm_judge, threshold: 0.7)

Core entities

Entity	Description
Project	Top-level container scoped to your organisation
Eval Suite	A collection of test cases with a pass/fail threshold
Eval Case	A single test — an input, a scorer, and a threshold
Scorer	The evaluation function (exact match, regex, LLM judge, etc.)
Run	One execution of a suite — immutable once complete
BYOK Key	Your encrypted LLM provider key for AI-powered scoring

The eval flow

You call POST /v1/run (or use the SDK/CLI/GitHub Action)
LaunchGate loads the suite and all its active cases
Each case runs its scorer against your output concurrently
Results are aggregated and the pass rate is computed
If pass rate ≥ suite threshold → Cleared for launch ✓
If pass rate < suite threshold → Launch aborted ✗
Results are stored immutably for audit and trend analysis