Set, Prove & Hold
The difference between a rule an AI has been asked to follow and a control designed to resist bypass.
Set, Prove & Hold (4 + 4 + 4)
Set is sequential before build, Prove supplies four pre-launch tests, and Hold runs four concurrent disciplines after launch.
Set
Steps before you build
Prove
Tests before you launch
Hold
Disciplines after launch
Name what must never happen
Write a rule someone could try to break this afternoon.
- Write a "never", not a value: "be helpful and safe" cannot be tested.
- Write it for this system rather than copying a vendor's generic list.
- Use the applicable law, the damage this system could cause, and known AI failure modes as inputs.
Where the rule is enforced decides what it can resist
The same customer-data rule can be expressed in four places. The layer changes what reaches the AI and what limitations remain.
What it is
The AI receives only task-needed data and tools
Customer-data example
Correctly scoped retrieval returns records for the signed-in customer only.
Scoping is before the model: only task-needed, authorised records reach the AI.
What reaches AI
Only authorised, task-needed customer records reach the AI.
Limitation
It depends on correct identity, authorization, and scoping design.
Build each rule into the strongest layer that can carry it. Architecture limits exposure; runtime checks add an independent boundary; policy and prompts remain useful but are not equivalent controls.
The layer changes how you build, test, and re-test
Compare four enforcement layers across the Set, Prove, and Hold lifecycle.
| Layer | Set — build it | Prove — test it | Hold — re-test it |
|---|---|---|---|
| 1 · Policy | Write it in a document | Review whether the policy is complete and communicated; it does not itself execute a technical test. | Review when the policy or use changes. |
| 2 · Prompt | Put it in instructions | Attack the instruction and record residual failures. | Re-test on model or instruction changes. |
| 3 · Runtime | Add an external input or output check | Attack it and measure wrong blocks. | Re-test when the scanner, model, or supported languages change. |
| 4 · Architecture | Scope data and tools to the task | Verify authorization and scoping are in place. | Review when identity, data access, tools, or architecture change. |
Layer choice is a maintenance decision as well as a security decision: weaker, change-sensitive controls generally demand more frequent evidence.
Track the number of active guardrails at each layer, alongside dated test evidence and legitimate-work block rates.
Where to start
Three practical moves, in order.
- 1Inventory the guardrails you think you haveMark the enforcement layer for each and review how many rely only on policy or instructions.
- 2Try to break the three that matter mostUse the deployed system and record the attempts, versions, results, and legitimate-work effects.
- 3Move one guardrail to a stronger layerStart by narrowing the data and tools available to the task.
Turn guardrails into evidence
We can help your team map controls, test the deployed system, and establish a reviewable operating rhythm.
Start a conversationAgent Authority Model
Authority decides what a handover may do; guardrails decide how strongly its boundaries are enforced.