Enterprise Core Runtime Guard for Autonomous AI Systems.
DriftTrace evaluates behavioral drift inside autonomous agent execution flows and produces a structured Drift Event in real time.
POST /evaluate
Input:
- objective
- steps
- context
Output:
- drift_score
- severity
- objective_fidelity
- reason
- recommendation
- verdict
DriftTrace evaluates agent execution steps against the original objective.
If deviation exceeds a defined threshold, a structured Drift Event is generated.
The consuming platform decides whether to:
- Allow execution
- Trigger review
- Block action
DriftTrace does not replace orchestration. It produces a runtime signal that can be consumed by enforcement or workflow systems.
- Objective representation tracking
- Multi step directional deviation scoring
- Cumulative divergence monitoring
- Threshold based drift classification
- Structured runtime telemetry emission
Each evaluation produces:
- drift_score
- severity
- objective_fidelity
- reason
- recommendation
- verdict
Signals are export ready for SIEM, XDR, orchestration engines, or governance pipelines.
Install dependencies:
pip install -r requirements.txt
Run the API locally:
python -m uvicorn app:app --reload
Open:
DriftTrace is a Core Runtime Guard module.
It transforms hidden reasoning drift into structured governance telemetry during execution.
We are opening a limited design partner program for organizations building autonomous AI agents.
See details in: DESIGN_PARTNER.md
For enterprise design partnership inquiries: aivault@aivaultsystems.com