Skip to Content
DocumentationTraining EngineHow Training Works

How Training Works

The Self-Improving Flywheel

Every analysis trains a model on YOUR codebase patterns. Your 50th analysis is dramatically better than your 1st.

What It Does

QuantumVerifi learns from every analysis you run. Passing tests are automatically harvested as training data, and when enough data accumulates, a custom AI model is trained specifically for your codebase.

The result: test generation that understands your coding style, your frameworks, your naming conventions, and your architectural patterns — getting better with every analysis.

How It Works

  1. Analyse — You run an analysis on your codebase
  2. Harvest — Passing, execution-verified tests are automatically extracted as training data
  3. Train — When enough data accumulates, a custom model adapter is trained for your tenant
  4. Route — Future test generation uses your trained model instead of the generic base model
  5. Repeat — Each analysis adds more data, making the model progressively better

This creates a flywheel: more analyses → better training data → better models → better tests → more passing tests → even better training data.

Quality Gates

Not all tests become training data. Only tests that meet strict quality criteria are harvested:

  • Analysis must achieve grade C or above
  • Tests must be execution-verified (actually ran and passed in a sandbox)
  • Content must be substantive (no trivial stubs)
  • Duplicates are automatically filtered via content hashing

Transparent Fallback

Your trained model is used transparently — if it ever produces lower-quality output, the system automatically falls back to the base AI model. You never see a degradation in quality.

Available On

The training engine is available on Scale and Enterprise plans. All other plans benefit from QuantumVerifi’s shared learning across the platform.

Training EngineDeployment