Runtime assurance for AI systems

AI everyone can stand behind.

A receipt for every governed action — one your board, your customers, and your regulator can check without trusting us.

OVERT receipt · #4f2c…a081 a081 ← 4f2c ← 1d9a ← … Verified — check it yourself

Why runtime

Move fast and prove things.

Accountability that slows you down is a tax, not accountability. Evidence is made in exactly one place — the moment a governed action runs — so Glacis stands in the request path at that moment, with three jobs.

See

Which controls ran. Which didn’t.

Fleet-wide coverage and drift — not event volume.

coverage18 of 21 workflows
uncoveredclaims intake +2
drift, 7d2 controls off policy
Simulated coverage for illustration

Control

Block what shouldn’t run. Allow what should.

Inline at the inference boundary. Deterministic, policy-driven, model-agnostic. No retraining, no swap.

actionpayments.transfer
policy hitclaims/v3
outcomedenied · receipted

Prove

Signed. Witnessed. Portable.

Every governed action becomes an OVERT receipt. Portal minting adds an independent witness countersignature. Verifiable without a Glacis account.

formatOVERT · open spec
witnessoperator signed · witness on portal
replaydeterministic
Portal-minted path shown
“Together, Glacis and nVoq have raised the bar for proactive AI trust and assurance, embedding consent validation, human‑in‑the‑loop oversight, and robust guardrails directly into ambient workflows.”
Chad Hiner Chad HinerSenior Vice President of Product, nVoq nVoq
Working with
  • CHAI — Coalition for Health AI
  • DiMe — Digital Medicine Society
  • AI GovOps Foundation
  • ScaleHealth
  • EisnerAmper
  • Cloudflare

Verification vs validation

Two questions. Your monitoring answers one.

Verification: did it work? Validation: did it stay inside your rules? Most enterprises instrument only the first.

What your telemetry answers
What none of it answers
The call completed in 1.2 seconds. No errors, no timeouts.
The disclosure your policy requires was never read to the caller.
Token spend, latency, and error rate are all inside budget.
The assistant quoted a benefit the caller isn’t entitled to.
The data stayed in the right region, correctly classified and labeled.
The escalation rule that should have routed this to a human never fired.
Every request and response is traced end to end, inside your tenant.
Nobody outside your tenant can check any of it without taking your word for it.
Backed by
  • AI House
  • Mighty Capital
  • Lionheart Ventures
  • SAIF — Safe AI Fund
  • Plug and Play
  • Sourdough Ventures

Where GLACIS fits

Glacis is the evidence layer under the AI you already run.

Each category below is good at its job. None was built to sit in the AI request path, stop an action that breaks a stated rule, and hand a counterparty evidence they can check without trusting the system that produced it.

Runtime enforcement Signed evidence Portable proof
Governance platforms
OneTrust · Credo · etc.
— policy only — PDF reports — vendor-locked
Model evals
Patronus · Galileo · etc.
— scores, not gates — eval scores — vendor-locked
LLM observability
LangSmith · Helicone · etc.
— traces, not gates — unsigned traces — vendor-locked
Platform telemetry & posture
Microsoft Purview · Grafana + OpenTelemetry · DSPM
— at the data layer — audit logs, unsigned — tenant-bound
GLACIS ✓ inline arbiter ✓ OVERT receipts ✓ verifiable by anyone

Glacis sits inline at the inference boundary and signs a record of which control ran on a governed action and what it decided — evidence a control ran, not a certificate the system is safe, correct, or compliant. Receipts minted through the portal are countersigned by an independent witness and carry an inclusion proof; SDK receipts are operator‑signed only.

Receipt chain

What an auditable AI decision looks like.

Every governed action mints a hash-chained receipt. Portal-minted receipts add an independent witness. Tamper one — the verifier shows where.

Customer trust boundary hash Agent tool call Arbiter policy · enforce arbiter-01.us-east Allowed action proceeds Blocked denied · receipted Receipt n−1 1d9a… Receipt n 4f2c…a081 Witness co-signed · append-only Verifier anyone, any time runtime · inside customer trust portable proof · independently verifiable Customer trust boundary hash Agent tool call Arbiter policy · enforce arbiter-01.us-east Allowed proceeds Blocked receipted Receipt n−1 1d9a… Receipt n 4f2c…a081 Witness co-signed Verifier anyone, any time
Chain intact — every receipt links to the one before it.

Simulated for illustration

The buyer, the carrier, the regulator

The same question, on three letterheads.

Sooner or later, someone outside your company asks what your AI actually did — an insurer at renewal, a customer in security review, a regulator with a question. That conversation goes better when the answer already exists, rather than being assembled under pressure. OVERT turns it into a single portable artifact you can hand over: the evidence pack.

  • InsuranceE&O carriers are carving out autonomous-agent action. The renewal conversation increasingly turns on runtime evidence.
  • RegulationThe EU AI Act and Colorado’s automated-decision rules are converging on the runtime receipt: the disclosable record of a consequential decision.
  • ProcurementAI security questionnaires have begun asking for third-party-verifiable runtime audit artifacts. Deals move faster when you can hand one over.

Verify it yourself

Don’t take the receipt on trust. Check it.

The receipt below came out of a real run. Check it here and you have done exactly what an auditor, a customer, or an insurer would do — no account, nothing sent anywhere, nothing taken on our word.

Open standard

Glacis is the steward of OVERT.

Observable Verification Evidence for Runtime Trust: evidence independent parties can verify without protected‑content egress. Every receipt on this site carries it.

Standard
OVERT 1.1
Status
Published 11 June 2026
License
Royalty‑free covenant

Someone will ask. You’ll already have the answer.

A procurement questionnaire. A renewal. A board asking about the AI. Each goes better when the evidence already exists: a runtime receipt for every governed action, an evidence pack the regulator, customer, or insurer can verify themselves. Live in weeks, on a plan that starts free — and there’s a real receipt on this page you can check right now.