VectorCertain
USPTO PROVISIONAL · FILED JAN 2026
MONITORING ENSEMBLE 828 MODELS
BLOCKING AT 3.6 ms
Execution governance · Companion to the HCF2-SG hub

MRM-CFS — Millisecond Execution Governance

If HCF2-SG decides whether an AI decision can be trusted, MRM-CFS decides whether that trusted decision may be acted upon right now — in 3.6 milliseconds.

11,197 VERIFIED TEST EXECUTIONS
PARALLEL SUPERVISION · MINIMAL-RISK FALLBACK CLAIMED

MRM-CFS is VectorCertain's supervisory execution-governance layer: a behavioral ensemble of 828 minimal-resource monitoring models running in parallel with — never serialized into — the AI system it governs. It enforces runtime constraints, gates permission-to-act at the moment of execution, and falls back to minimal-risk behavior when constraints cannot be satisfied. Blocking operates at 3.6 milliseconds, and the governance pipeline claims are supported by 11,197 verified test executions [VC-C].

The layer consumes trust levels from the HCF2-SG hub and answers the question the hub deliberately leaves open: a decision whose evidence is independent enough to trust may still be unsafe to execute in this moment, in this environment, under these constraints. Execution governance is that second, time-bound judgment. Filed as a USPTO provisional in January 2026, with a companion standalone filing claiming the monitoring ensemble as an independent article of manufacture [VC-C].

Request the technical briefing

§ 1 · WHY LATENCY DECIDESWhy Milliseconds Are the Deployment Boundary

Supervision that costs milliseconds is deployable in production; supervision that costs seconds is not. A trading loop, a network enforcement path, and a vehicle control cycle each budget their full decision path in single-digit milliseconds — an oversight layer that arrives after the action window closes is an audit log, not a control [VC-C].

Architectures that serialize a second model into the decision path — the LLM-judge pattern — inherit a full model inference per governed decision, which places them orders of magnitude outside those budgets. The gap is structural rather than an optimization target: a serialized judge cannot be faster than the model it runs. MRM-CFS avoids the serialization entirely. Because its 828 monitors are minimal-resource, supervision does not duplicate the governed system, does not sit in the critical path as a bottleneck, and requires no control coupling with the model it governs [VC-C].

Parallel supervision: governed AI system with 828-monitor ensemble running alongside, feeding a permission gate that outputs act or minimal-risk fallback GOVERNED AI SYSTEMcritical path · unimpeded200 MONITORING ENSEMBLE828 minimal-resource models210 GATE3.6 ms220 ACT MINIMAL-RISK FALLBACK230
FIG. 1Governed system (200) runs unimpeded; the monitoring ensemble (210) observes in parallel and the permission gate (220) blocks, permits, or triggers minimal-risk fallback (230) at 3.6 ms.
Table 1 · Parallel supervision vs. the serialized-judge pattern
PropertySerialized judge (LLM-as-judge)MRM-CFS parallel supervision
Position in decision pathInside — every decision waits on a second inferenceAlongside — governed system never waits on a monitor
Latency classFull model inference per decision3.6 ms blocking
Compute profileDuplicates a frontier-class model828 minimal-resource monitors
Failure postureUndefined without integrator workMinimal-risk fallback — claimed behavior
Coupling to governed modelPrompt/control coupling requiredNon-intrusive; no control coupling

§ 2 · THE MECHANISMHow Parallel Supervision Works

The behavioral monitoring ensemble

The 828 monitors observe the governed system's behavior non-intrusively and in parallel, each covering a narrow behavioral surface at minimal compute. Breadth comes from the ensemble, not from any single heavyweight overseer — which is what keeps the layer inside a 3.6-millisecond blocking budget [VC-C].

The permission-to-act gate

At the moment of execution, the gate evaluates the candidate action against runtime constraints and the HCF2-SG trust level. Permission is granted, or it is not, before the action occurs — inline blocking rather than after-the-fact alerting [VC-C].

Minimal-risk fallback as claimed behavior

When constraints cannot be satisfied, the system falls back to minimal-risk behavior — and that fallback is claimed in the filing rather than left to the integrator. For functional-safety arguments in the ISO 26262, IEC 61508, and DO-178C families, a claimed fallback is the difference between a documented safety property and an implementation hope; the same design posture aligns with the technical-robustness expectations of the EU AI Act's Article 15 [6] and the MANAGE function of the NIST AI RMF [2][3].

§ 3 · IN THE STACKWhere MRM-CFS Sits in the Portfolio

MRM-CFS consumes HCF2-SG trust levels and supplies the execution gate that Cyber-SG applies inside security enforcement paths. A companion standalone filing claims the monitoring ensemble as an independent article of manufacture, so an enterprise can adopt observability first without committing to the full permission-to-act governance model — and upgrade later without replacing the monitors [VC-C]. Explore the hub at HCF2-SG, the security application at Cyber-SG, and the deployment platform at SecureAgent ACA.

§ 4 · QUESTIONSFrequently Asked Questions

What is execution governance for AI?

Execution governance is the runtime judgment between a trusted decision and a permitted action. MRM-CFS enforces constraints at the moment of execution through a permission-to-act gate operating at 3.6 milliseconds, supervised by 828 parallel behavioral monitors. It answers a different question from trust scoring: not "is this decision sound?" but "may it happen now, here, under these constraints?" — with minimal-risk fallback when the answer is no.

Why parallel monitors instead of an LLM judge?

A serialized judge adds a full model inference to every governed decision, which structurally cannot meet millisecond budgets — it is at least as slow as the model it runs. Parallel minimal-resource monitors observe without sitting in the critical path, so the governed system never waits on supervision. That is what makes 3.6 ms blocking possible inside trading loops, network enforcement, and control cycles.

What happens when constraints cannot be satisfied?

The system falls back to minimal-risk behavior, and that fallback is claimed in the patent filing rather than delegated to the integrator. For safety cases built on ISO 26262, IEC 61508, or DO-178C, a claimed fallback converts an implementation detail into a documented, licensable safety property — validated across 11,197 test executions.

CONTACTTalk to VectorCertain

Every figure on this page traces to sealed, hash-verified validation artifacts — including findings that contradicted our own published estimates, which we recorded rather than reconciled away. Technical briefings are available for enterprises, evaluators, and standards bodies.

Request the technical briefing

REFERENCES

  1. NIST. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. nvlpubs.nist.gov
  2. NIST. AI Risk Management Framework (program page). nist.gov/itl/ai-risk-management-framework
  3. European Parliament and Council. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex. eur-lex.europa.eu
First-party validation artifacts (VectorCertain, sealed and hash-verified):

[VC-C] Patent portfolio and platform engineering baseline — 77-claim hub filing, stack integration claims, 36,181-test regression suite; portfolio documentation, January 2026. Public URLs will replace artifact names when the corresponding research pages publish.

Join the waitlist

Our signup form is temporarily offline while we perform maintenance. Nothing is lost — reach us directly and we'll add you by hand.

PLACEHOLDER · Tally.so form returns here · engineering ticket open

Email us to join