ACI Safety & Policy enforced the positive controls across ACI Personal Agents and ACI Edge Runtime, while keeping the negative-control rate at zero on the personal-agent path and 0.075 on the current edge configuration.
1.0
Personal positive-control enforcement
Typed rule applied
1.0
Edge positive-control enforcement
Typed rule applied
0.0
Personal negative-control rate
No spillover in this run
0.075
Edge negative-control rate
Measured under the current edge configuration
April 2026
Policy products are easy to describe and harder to measure. Buyers do not just need a statement that rules exist. They need to know whether those rules are enforced on the deployed surface and what the spillover rate looks like when the rule should not fire.
This benchmark measures that contract directly on two deployment modes that matter in practice: personal-agent software and edge-runtime systems. The same typed rule is evaluated against positive and negative controls so the result is quantitative rather than anecdotal.
The same typed policy surface can be evaluated across personal-agent and edge-runtime deployments rather than being rephrased as separate prompt-only guardrails for each environment.
The benchmark measures whether the targeted rule actually forces the intended result on the positive controls, not just whether a model usually follows a policy instruction.
Negative-control rates are recorded explicitly. That gives deployers a concrete tuning and operating metric rather than leaving policy behavior at the level of anecdotes.
The result shows full positive-control enforcement on both surfaces. The personal-agent path held the negative-control rate at zero in this run. The edge path held full positive enforcement and recorded a measurable negative-control rate of 0.075 under the current configuration.
| Surface | Positive-control enforcement | Negative-control rate | Interpretation |
|---|---|---|---|
| ACI Personal Agents | 1.0 | 0.0 | The typed rule enforced the positive controls while leaving the negative controls unchanged in this benchmark run. |
| ACI Edge Runtime | 1.0 | 0.075 | The edge path enforced every positive control and recorded a small negative-control rate under the current runtime configuration. |
What the result says
These results place ACI Safety & Policy as a measurable enforcement layer for personal-agent and edge deployments. Operators can review positive-control enforcement and negative-control spillover directly, then tune the rule surface on numbers rather than relying on prompt wording alone.
The safety add-on matters wherever the deployment surface has to enforce explicit rules rather than hoping policy survives as prompt language.
For local assistants and operator-facing software, the benchmark shows that policy enforcement can be attached to the same device-local state that handles personalization.
For edge systems, the result shows that typed rules can be tested directly against the deployed runtime rather than treated as a soft instruction layer around it.
The point of the product is not only to block something once. It is to make policy behavior measurable, reviewable, and consistent across deployment modes.
Explore where ACI Safety & Policy fits in device-local agents and edge deployments that need measurable hard control.