An operations agent proposes updating every customer record when only the current service region needs changing.
PF SystemsOperational authority for agentic AIStart with one workflow PF Kernel checks a proposed AI action against the organisation's authority before it takes effect. Its compact profile supports commercially flexible deployment, including environments where latency, connectivity and local control matter.
The operational problem
Most governance becomes too blunt when every exception is simply blocked.
PF Kernel uses five outcomes—Allow, Deny, Modify, Step Up and Stop the Line—so useful work can continue while excessive, unsupported or systemically unsafe actions receive proportionate treatment.
Operations example · Modify
One proposed action. One governed consequence.
PF Kernel modifies the proposed action to the current region, preserving the useful intent while constraining the scope.
The original proposal and the effective region-only action remain visible together.
Designed to sit close to action
PF Kernel can be deployed independently or as the final decision authority within PF OS. It used approximately 19 MB RSS in a bounded 10,000-request local benchmark, supporting its lightweight deployment proposition without establishing a certified minimum requirement.
Measured internal performance
A saved internal local benchmark measured 2.2ms median policy evaluation and 2.66ms at p95 in its tested configuration. These figures are not an independent benchmark, production service-level guarantee or certification; hardware and configuration affect results.
Authority remains explicit
ClientBridge may carry the request, PF Memory may provide context and PF Core may preserve evidence. None inherits PF Kernel's decision authority.
PF Kernel · measured local evidence
Lightweight by measured local evidence.
These results come from bounded local tests on a frozen test baseline. They describe what was observed in that environment; they do not establish a certified minimum or a production capacity claim.
Peak RSS
Observed during a bounded 10,000-request local benchmark.
Post-test change
The measured change from baseline to stable post-test RSS.
Highest completed ramp RSS
Observed at concurrency 2,048 with no out-of-memory event. A capacity maximum was not established.
Planning recommendations
Starting points for engineering discussion—not guaranteed requirements.
Plausible for a tightly bounded service; not qualified as a minimum.
Recommended standalone allocation with useful operating headroom.
Conservative planning for larger policies, deeper receipt history or colocated helpers.
Evidence boundary. Internal local evidence, not an independent benchmark, certification, service-level commitment or production sizing guarantee. Hardware, configuration, policy size and surrounding services affect results; storage, networking or external dependencies may constrain a deployment before memory does.
A controlled first step
Which AI action should your organisation govern first?
Start with one consequential workflow, its authority boundary and the evidence needed afterwards.
Request a governed AI briefing