An existing AI application proposes an action using governed organisational context.
PF SystemsOperational authority for agentic AIStart with one workflow PF OS is the three-component operating system at the centre of PF Systems. Its roles are deliberately separated so that knowledge does not become authority and evidence does not become a decision-maker.
The operational problem
One oversized platform is not the right answer for every workflow.
A customer-service workflow, a payment instruction and a robot at the edge have different infrastructure, latency and assurance needs. PF OS is modular so organisations can begin with the smallest useful footprint and expand from evidence.
Complete governed workflow
One proposed action. One governed consequence.
PF Memory supplies governed knowledge, PF Kernel decides the permitted outcome and PF Core preserves the linked evidence.
PF Trace can display the received evidence to an authorised operator, while ClientBridge connects the surrounding estate.
PF Memory knows
PF Memory supplies governed knowledge and context, including origin, permissions and lifecycle. It can govern an existing memory or retrieval system without requiring replacement. It does not decide or execute.
PF Core proves
PF Core preserves linked evidence, lineage and the material needed to trace, check and deterministically replay the governed evaluation. It does not make the underlying AI deterministic and it does not decide.
PF Kernel decides
PF Kernel applies organisational authority to a proposed action and returns Allow, Deny, Modify, Step Up or Stop the Line.
PF OS · bounded triad evidence
A complete triad with a modest observed footprint.
PF OS comprises PF Memory, PF Core and PF Kernel. In one bounded local qualification stage, the triad processed 32 sealed capture cases at concurrency four. The figures below distinguish direct observations from planning guidance.
Qualification cases
Observed across the 32-case local stage at concurrency four.
Median latency
Measured p50 for the same bounded qualification stage.
p95 latency
Measured p95 for the same bounded qualification stage.
Planning proxy
Sum of 106.0 MiB self and 9.9 MiB child maxima; not an exact simultaneous whole-stack footprint.
Planning recommendations
Starting points for engineering discussion—not guaranteed requirements.
Likely sufficient for a tightly bounded exercise; not a general requirement.
Practical starting allocation for a controlled local PF OS triad.
Recommended planning allowance for policy, receipts, validation and transient work.
Conservative colocated-service budget with ClientBridge, monitoring or evidence helpers; excludes a full desktop estate.
Deployment shapeHost-planning guidance
PF Kernel alone4 GiB lightweight host; 8 GiB is a comfortable baseline.
PF OS triad8 GiB host for bounded local operation.
PF OS + ClientBridge or operator tooling16 GiB host recommended.
Full local demonstration estate16 GiB practical starting point; 32 GiB preferred.
Evidence boundary. Internal local evidence and engineering planning guidance only. These figures are not certified minimum requirements, production capacity limits, current demonstration-estate measurements or guarantees across different hosts. Browser, container, model, monitoring and operator-tool costs must be sized separately.
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