Latency between observing a risk and computing a mitigation action generates an information void that erodes operational certainty. We resolved this void by releasing a forward-looking margin diagnostic in the KAIROS engine. The engine computes the predicted scalar inside the identical scoring substrate as the current state, guaranteeing mathematical parity.
Defining the Kairos Margin
The Kairos Margin (K) defines the strict distance to a state-gate breach. We compute this bounded residual as K = gamma - gamma_floor, where gamma represents the current control margin and gamma_floor is the policy-defined minimum. This metric enforces a zero-epsilon decision policy. A scalar of K < 0 dictates a breached state.
The cyber signal envelope carries kMargin, a gateBreached boolean, and a display regime. Operators read distance, sign, and regime from a single position, removing translation overhead from the execution loop. We classify the margin into three operator-facing regimes using a policy-configurable display tolerance (ε):
| Regime | Condition | Meaning |
|---|---|---|
| Plastic | K > +ε | Above the floor with headroom. |
| Kairos | |K| ≤ ε | At the threshold. The decisive moment. |
| Locked | K < -ε | Past the floor with margin. |
Same-Evaluator Predicted Gamma
Predicted Gamma extends the control margin computation into the action-preview trajectory. When an operator proposes a SOC action, the engine previews the post-action state. We compute Predicted Gamma using the exact compute_lambda_gamma function and calibration artifact that generated the current gamma.
The action mapper supplies an additive metric delta representing the action’s structural impact. The session executes the merge and re-scores the state. The parity between current and predicted values is a structural property of the engine architecture.
The Predicted K Margin
The engine pairs Predicted Gamma with the Predicted K Margin. We compute this paired scalar using the identical residual formula against the predicted state. The engine outputs evaluation.actionGate.predictedKMargin inside the signal envelope. Until real-world calibration data has arrived, the engine serializes these scalars as explicit JSON null.
Enforcing the Diagnostic Boundary
The action gate’s reject decision is governed by warning, loss, and action semantics. We decoupled the execution state from the predicted margin sign. The schema omits a predictedGateBreached boolean and a predicted display regime.
The engine reports structural facts and preserves the operator as the autonomous decision-maker. Enforcing on engine predictions alters the EU AI Act risk classification and shifts NIS2 and DORA audit boundaries. The current interface preserves the operator’s control over the SOC execution loop.
The V2 enabling parameter, reject_on_predicted_floor_crossing, exists at the interface level with a hysteresis buffer. We keep this parameter unimplemented to maintain the diagnostic boundary.
Calibration and Expansion
Two SOC action categories, ObserveOnly and ForensicSnapshot, leave the metric vector unchanged by construction and operate with fully resolved predicted scalars. Five categories, including HostIsolation and CredentialRevocation, await real-world action-impact calibration data and serialize as null. The engine preserves mathematical honesty over populated data fields. We will release AI-safety baseline calibration as the next operational update.
The Structural Integrity Constraints
Four engineering constraints secure the computation pipeline:
- Single populating site: A continuous integration test locates every write to
ActionGateDetail.predicted_gammaand fails the build if more than one computed initializer exists. - Closed-by-default action-mapper registry: Production builds construct the registry from a fixed static list.
- Strict session merge: The merge rejects unknown metrics, non-finite values, and missing baseline snapshots, yielding a
nulloutput. - Explicit-null serialization: The JSON schema retains the key and carries a
nullwhen calibration data is unavailable.
Deploy the KAIROS Substrate as a forward-looking margin diagnostic inside existing analyst workflows. Contact us to apply for design-partner access.