Skip to main content
Two ideas explain the entire policy system. Internalise these and the rest follows.

Templates suggest, admins decide

The product ships expert-authored templates — OWASP LLM Top 10, EU AI Act, GDPR, PCI/HIPAA, Shadow AI, Content Safety. Templates are read-only. Every policy you actually run is a clone that you own and tune.

Every policy has two lives

Policies start in Guideline mode — they observe and log what they would have caught, and change nothing. Only once you trust the data do you promote to Strict, where the same detectors act in real time.
This separation exists to answer the hardest question in policy enforcement — “is this rule going to break our users?” — with data, before you flip the switch.

​
The lifecycle

​
Day 1 — clone a template

1

Open Policies

Then choose New from template to reach the template library.
2

Find the right template

Filter by category, or search by OWASP reference (LLM02) or regulation (Art. 9).
3

Read the full spec

Each template card opens to its rationale, an example violation, its detectors, and every tunable parameter with its default and help text.
4

Clone and customise

Give it a name that means something in your environment — “PII Output — Customer Service” beats the generic template name. Optionally pre-select which applications it applies to.
A new policy is always created in Guideline mode, regardless of what the template suggests. There is no way to create a policy that starts enforcing. This is deliberate.

​
Days 1 to 14 — tune and observe

Four tabs on the policy detail page do the work:
The Test tab is a pure simulation. No state changes, no violations recorded. Use it as much as you like to dial in parameter values before any real traffic is affected.
Meanwhile the policy is observing live traffic in the background and logging every near-miss. After 30 days you have real numbers: total evaluations, would-block count, and false-positive rate.

​
Day 14 onward — promote to Strict

Promotion is gated on two conditions:
  1. At least 14 days in Guideline mode
  2. A false-positive rate below the threshold — 2% by default
Until both are met, the promote button tells you exactly what is missing: “Need 8 more days in Guideline mode” or “FP rate is 3.4% — must be below 2%”. When it unlocks, a four-step wizard opens:
1

Risk preview

Blast-radius arithmetic from the last 30 days of observation — for example, “Switching to Strict would have blocked 142 events, 0.31% of traffic, with 3 estimated false positives affecting 47 users.” No surprises.
2

Rollout strategy

All for immediate full coverage, canary to start with one application, or phased for a gradual percentage ramp.
3

Override path

Set the auto-demote threshold — by default, revert if the live false-positive rate exceeds 5%. This is the watchdog, and it is the reason promotion is not a one-way door.
4

Approval

See which roles must sign off, then send the request. The policy stays in Guideline until every approval lands.

​
Who has to approve

Required approvers are determined by the template’s category: One rejection freezes the request and the policy stays in Guideline; submit a new request when you are ready. When the last required approval lands, the policy flips to Strict automatically and begins enforcing.

​
Steady state — enforcing

A Strict policy shows a live badge and running statistics: blocks over 30 days, false-positive rate, and rollout strategy. The watchdog polls every 15 seconds, comparing the live false-positive rate against your auto-demote threshold:

​
The safety valve

An admin can demote a Strict policy to Guideline at any time, with no approvals required. Increasing safety — less enforcement — never needs a second signature. Only increasing enforcement does.
Use Pause to Guideline on the mode toggle. You are back in observation mode immediately.

​
Simple and Advanced views

The Policies page has a Simple ⇄ Advanced toggle. Simple presents a compact table and a straightforward creation flow for teams that want a handful of sensible rules. Advanced exposes the full template library, per-parameter tuning, scope filters, and the approval machinery. Both operate on the same underlying policies. The toggle changes what you see, not what runs.