Quvant Auriga
In development
Autonomy runs. Auriga steers.
Auriga evaluates each requested action using identity, delegated authority, policies, control state and available evidence. It returns a clear, reviewable outcome before the agent produces an effect.
THE MOMENT THAT MATTERS
Between intent and effect.
Agents no longer just answer: they may send emails, change data, initiate payments or close tickets. The point where the action can still be governed is the instant before it takes effect. Auriga operates at that transition.
FOUR OUTCOMES
Every action gets a clear answer.
Examples of intended behaviour
Proceeds
A refund to a verified customer, within the defined thresholds. The action may run and the decision remains recorded.
Goes to a person
A refund above the approval threshold. The request reaches the appropriate approver with the necessary context.
Denied
Sending personal data to a recipient outside the allowlist. A rule forbids it, so the action does not run.
Stops
There is not enough evidence to decide. Auriga does not guess: it stops and states what is missing.
HOW IT WORKS
Your conditions, applied where the agent acts.
Intercept the request
Auriga sits at the point where the agent requests the use of a tool, through the framework hook, gateway or policy engine already in place.
Read the context
It identifies who is acting, on whose behalf, on which resource and with what declared impact.
Apply the conditions
It evaluates thresholds, prohibitions, required approvals, control state and required evidence.
Issue a replayable outcome
The policy engine produces a deterministic technical verdict: given the same normalised inputs, rule version and control state, it returns the same outcome.
Enforce the outcome
The system controlling access to the tool receives Auriga's outcome and applies it: it runs the action, holds it for approval or blocks it.
CONTROL VALIDATION
A rule matters only if the control holds.
Auriga and Control Validation perform distinct functions. Auriga decides whether a specific action may proceed; Control Validation verifies, in an isolated environment and within the supported scope, whether a control reacts as declared when it is removed or altered. The result may update the control state used in later decisions, but it demonstrates only the behaviour observed in the executed test.
WHEN RULES ARE NOT ENOUGH
When rules are not enough, Auriga does not decide alone.
When policies and evidence do not support a clear outcome, Auriga holds the action and passes the case to a full Quadriga Review. Analyst, Critic, Synthesiser and Blind Validator work as in every Review; the Blind Validator checks the outcome without receiving the preceding debate. The approver receives the strongest alternative and the observations that could resolve the case.
THE RECORD
Who asked, who decided, what happened.
For every action, the record captures the request, declared identity and authority, the applied policy version, evidence state, any approval, the issued outcome and what happened afterwards. It can be exported for security reviews, audits and incident analysis.
LEARNING FROM OUTCOMES
Improve prioritisation without delegating authority to the model.
Over time, decision records and verified outcomes may form a versioned corpus for developing advisory machine-learning capabilities. These models may detect anomalies, prioritise verification work, suggest which cases require attention and improve evidence selection. Policies, stop conditions and technical verdicts remain deterministic: the model assists; it does not authorise the action.
Training requires independent outcomes, traceable labels and explicit rights to use the data.
IT INTEGRATES, IT DOES NOT REPLACE
It works with what you already have.
Identity and access stay with your IAM. Rules may continue to live in your policy engine. The gateway or hook applies the outcome at the operational point. Credentials remain in your vault: Auriga does not hold them. Decision records remain separate from keys and operational secrets.
WHAT IT IS NOT
Explicit boundaries.
- It is not an identity provider.
- It is not a universal gateway.
- It does not hold agent credentials.
- It does not certify regulatory compliance.
- It does not guarantee that an action is correct or risk-free.
- Machine learning does not replace policies, approvals or stop conditions.
Which actions would you delegate to your agents?
Tell us about the process, the actions and the limits that matter to you. Together, we can assess where Auriga should sit, which conditions should apply and which controls need to be verified.