Core controls · Implementation
From design-time testing to runtime verification.
Traditional software is tested before it ships. AI can't be, not fully. This is the implementation library: classify the risk, apply the layers, and deselect what you don't need.
The essentials
Seven steps, in order.
Start here and work down. Each step builds on the one before it; branch into specialised controls once the essentials are in place.
01
Risk Tiers
Classify your system by impact, so the controls match the consequence.
02
Risk Assessment
Quantify control effectiveness and residual risk per tier.
03
Controls
Implement the layered pattern: guardrails, reviewing controls, oversight.
04
Agentic
Add controls when your agent can invoke tools or take actions.
05
IAM Governance
Identity, lifecycle, and delegation for human and non-human callers.
06
Judge Assurance
Measure and calibrate the model-as-judge so you can trust its verdicts.
07
Checklist
Track implementation progress against the controls you selected.
Specialised controls
Reach for these when your deployment needs them.
Match the add-on controls to what you're actually running. Deselect the rest.
Multimodal
Image, audio, and video inputs that bypass text-based guardrails.
Reasoning models
Chain-of-thought models where the reasoning itself needs scrutiny.
Streaming
Real-time outputs that commit before a full review can finish.
Memory & context
Long context and persistent memory that carry risk across sessions.
Multi-agent
The primer for many-agent systems, before you go deep on MASO.
PACE resilience
What each layer does when it fails, and the safe state it falls back to.
The principle
Match controls to risk. Guardrails are necessary but not sufficient, the judge is assurance not control, and humans remain accountable. Apply the right controls at the right time, for the right reasons.
All the principles, in full →