The four verification layers explained
Every response G8KEPR checks passes through four independent verification layers. Each layer answers a different question about the output, and together they produce a single, auditable verdict about whether the response should be trusted.
Why four layers
A single check can be fooled. A response might be perfectly formatted but factually invented, or accurate but tampered with in transit. By running four focused layers instead of one broad check, G8KEPR catches distinct failure modes independently, so a weakness in one layer is less likely to become a blind spot for the whole system.
The layers
The layers run in sequence, from the most fundamental question to the most policy-specific.
- 1.Integrity — Confirms the response is authentic and unaltered. This layer checks that what your application received matches what the model produced, flagging tampering, truncation, or injection between generation and delivery.
- 2.Grounding — Confirms the response is supported by its sources. When context, documents, or retrieved data are provided, this layer checks that the claims in the output actually trace back to that material, flagging unsupported or fabricated statements.
- 3.Structural — Confirms the response has the shape your application expects. This layer validates format, schema, and required fields, helping keep malformed or incomplete output from reaching a downstream system.
- 4.Constraint — Confirms the response obeys your rules. This layer enforces the policies you define: allowed and blocked content, value limits, and business or safety constraints specific to your use case.
How they combine
Each layer produces its own signal. Rather than treating any single result as final, G8KEPR aggregates all four into one decision:
- ▸Pass — every layer is satisfied and the response is delivered.
- ▸Flag — one or more layers raise a concern; the response is surfaced for review or routed by your policy.
- ▸Block — a layer detects a violation serious enough to stop the response outright.
Every verdict is recorded with the per-layer detail behind it, so you can always see *which* layer acted and *why*.
You can tune the sensitivity of each layer independently. Start with defaults, review your flagged responses, and tighten the constraint layer as your policies mature.
Layers run together in real time and typically add only a small amount of latency to each request.
To adjust thresholds and policies, see Configuring verification.