Verify · Verification
RUUT FORGE — the verification layer
RUUT FORGE is the verification layer of the RUUT system. It checks every consequential AI decision against versioned, readable rules before it executes, returns a clear verdict — CERTIFIED, REJECTED or UNCERTAIN — routes what needs a person to human approval, and records why.
Last updated
What it does
What RUUT FORGE does
Screens what goes in
FORGE checks user input and retrieved content for manipulation such as prompt injection, and for data that should not reach a model, before reasoning begins.
Checks every action against readable rules
Each proposed tool call or decision is checked against versioned rule packs — which agent, which system, which action, which limits — written so compliance and safety teams can read and approve them.
Gives a verdict, not a guess
Every check ends CERTIFIED, REJECTED or UNCERTAIN, with its reasons and the evidence used. UNCERTAIN goes to a person, so nothing consequential proceeds on a guess — and every verdict is recorded.
How it connects
Where it sits in the loop
FORGE is the Verify stage of the RUUT Agent Loop and sits on both sides of the reasoning: it guards what enters RUUT BRAIN and checks what BRAIN proposes before any agent acts. Nothing consequential reaches the Act stage without a FORGE verdict.
- Perceive
- Reason
- VerifyFORGE
- Act
Deployment
How it is deployed
FORGE is deployed wherever BRAIN runs — on-premise on EDGEBOX, in a private cloud, or hybrid — and can also be called through its API to check decisions made outside the RUUT system. Rule packs, approval rules and escalation paths are defined with your team.
Typical uses
- Requiring supervisor approval before an agent stops a production line
- Blocking an agent from calling a system outside its allow-list
- Showing an auditor, months later, which rules approved a decision and why
FAQ
Questions about RUUT FORGE
What does it mean to verify an AI decision?
Verifying an AI decision means checking it before it executes: is the input trustworthy, is the proposed action allowed by policy, is the output supported by evidence, and does a person need to approve it. RUUT FORGE runs these checks on every consequential path and records each one.
What does a FORGE verdict say?
Every checked decision ends in one of three verdicts: CERTIFIED, when it passes the rules that apply; REJECTED, when a rule fails, with the reason; or UNCERTAIN, when the evidence is not enough to decide, so a person reviews it. FORGE never calls an AI output true or correct — it records which rules were applied, to which evidence.
How is FORGE different from AI guardrails?
Most guardrails filter text: they block unsafe words in a prompt or a reply. FORGE governs decisions and actions. It checks what an agent is about to do against versioned rules your team has approved, decides whether a person must confirm it, and keeps a record that shows afterwards exactly why an action was allowed or refused.
Who writes the rules FORGE applies?
Rules are written in a readable, versioned rule language and grouped into domain packs — for example, an industrial-safety pack for plants or a pack for a lending workflow. Your compliance or safety team can read, review and approve exactly what the system enforces, and every verdict records the version of the rules it was checked against.
Does FORGE replace human approval?
No. FORGE decides when human approval is needed and routes the decision to the right person; it does not remove people from consequential decisions. You choose which actions always require approval and which can proceed automatically within policy.
Can FORGE stop prompt injection?
FORGE screens inputs and retrieved content for prompt-injection patterns and restricts what any agent is allowed to do, which limits the impact even if an attack is not detected. No single control is perfect, so FORGE combines screening, least-privilege tools, output checks and approvals.
Can FORGE be used without the rest of RUUT?
Yes. FORGE is designed as standalone verification infrastructure with an API, so it can check decisions made by agents and models outside the RUUT system as well as inside it. Within RUUT, every consequential step proposed by RUUT BRAIN passes through FORGE automatically, with no extra integration work.
Related
Works together with
- ReasonOrchestration
RUUT BRAIN
The orchestration layer. Routes reasoning across models and coordinates every agent.
- ActFinancial institutions
RUUT VAULT
Agentic infrastructure for banks and financial institutions — credit, collections and compliance.
- ActPlant safety
RUUT FOREMAN
The plant-safety agent. One clear shift-end view of violations, machine signals and observations.
- ComputeSovereign compute
RUUT EDGEBOX
The on-premise AI appliance that runs the whole RUUT loop inside your own walls.
Let's talk about your system.
No public pricing — every deployment is scoped to your data, infrastructure and risk tolerance.