Skip to content

The Control Platform

Section 17 argued the denominator must be counted at the point of decision. This section is about the machine that counts it: Glacis’s Control Platform, whose runtime core is called the Arbiter.

The Arbiter is a reverse proxy at the inference boundary — the line where an inference request enters production. Customer applications keep calling OpenAI- or Anthropic-shaped endpoints exactly as before; the requests flow through the proxy, which applies policy controls in real time and signs an attestation for every decision. GROUNDED · Labs No SDK rewrite, no new API to learn — the integration story is “change the base URL.”

Two engineering properties are load-bearing for everything the insurance side later claims: GROUNDED · Labs

  • Deterministic. Policy evaluation on the hot path is rule-based, not model-based — the same input yields the same decision, which is what makes a decision attestable and a trigger contractually definable.
  • Offline-by-construction. The runtime makes no LLM calls of its own. The thing producing the evidence cannot itself drift.

The one-line pitch for the whole section, from the draft brochure: every existing AI-insurance product is built on an assessment of what a system was designed to do; Inference Re is built on a continuous record of what it actually does in production — and the ability to control and prove it. GROUNDED · Brochure Pre-deployment assessments stop at the model. The Control Platform starts where the model’s output meets the world.

Redrawn from the draft brochure’s Figure I: GROUNDED · Brochure

AI vendor model "what it was DESIGNED to do" ↑ pre-deployment assessments stop here inference request enters production GLACIS CONTROL PLATFORM ↓ Inference Re starts here SCAN ENFORCE NOTARIZE attests every inference against the declared baseline blocks scope exceedance before it reaches the user signs the OVERT attestation record Compliant output → the end user OVERT record → attestation record — Inference Re trigger data VERIFIED INFERENCE POPULATION "the denominator — the missing unit of exposure"

Read the three verbs as the three insurance functions from section 17: GROUNDED · Brochure

Verb What happens Insurance meaning
Scan Every inference is attested against the declared behavioral baseline Measure — the denominator increments
Enforce Scope exceedance is blocked before output reaches the user Control — loss frequency is actively reduced
Notarize An OVERT attestation record is signed for the decision Verify — the evidence becomes independently checkable

Enforce deserves emphasis because it changes the category: the platform “does not only observe.” Blocking non-compliant output pre-delivery makes this a risk-control layer, not just an evidence layer — the same distinction Volume 1 drew between documenting a risk and improving it (see risk selection). The next page unpacks it. GROUNDED · Brochure

Discipline from day one: the platform produces evidence for governed paths only. Traffic that never routes through the Arbiter is invisible to it, and the canonical scope claim — taught verbatim in chain of custody and drilled in the claim taxonomy — says exactly that. Receipts “are silent about un-mediated paths.” Never let a diagram like Fig. I tempt you into “all your AI traffic.” GROUNDED · Labs

01 Baselines and drift

Drill this page →13 bank questions stand behind what you just read. Check it while it’s warm.