Standards say should. Gate proves does.
Ouli Gate runs an agent many times inside the boundary it will ship with, scores every step — not just the answer — against technical safety and the law, and signs the evidence. Buyers, certifiers, insurers and regulators can check the report without taking anyone’s word for it.
B2B ONLY · FOR AGENT DEVELOPERS, ENTERPRISE ADOPTERS, CERTIFIERS AND INSURERS · GATE-TESTED, NEVER CERTIFIED
What Gate is
Agents have gone from answering to acting: they move money, change systems, send documents and, increasingly, move machines. The runtime that keeps an agent inside its boundary now exists, and much of it is open source. What nobody provides is independent proof that a particular agent, at a particular version, has earned that boundary before it ships.
Ouli Gate is that proof, delivered as software. It tests an agent’s decision layer against technical safety and against the law, inside the runtime the agent will actually run in, and signs the evidence in a form a general counsel, a certification body, an insurer or a regulator can check for themselves.
Nobody has to remember to distrust an old result. A new version, a new policy or a pack update voids the report, and the re-check fires from the same pull request that made the change.
Standards bodies and certifiers keep the job of writing the standard. Gate implements the test methods behind it, and keeps them current as the standard is revised.
A partner certification body issues any certificate, under its own scheme, on Gate’s evidence. A product is Gate-tested, never Gate-certified.
OpenShell-class runtimes enforce the boundary. Gate runs inside them, reads their logs as evidence, and signs what the agent actually did.
Why now
Proof before shipping is becoming a legal requirement — in the United States through the courts, in Europe by statute — and it applies to anything sold there, wherever it was built.
Courts are letting strict product-liability claims proceed against AI applications. Colorado’s AI Act, New York City’s bias-audit law, bank model-risk guidance and the FTC’s substantiation rules already apply, and insurers have begun excluding AI from general cover.
The new Product Liability Directive makes software and AI strictly-liable products, with disclosure duties and presumptions of defect.
The Machinery Regulation requires third-party assessment of machinery whose safety functions use self-evolving, machine-learning behaviour.
The AI Act’s high-risk obligations apply: first to Annex III systems, then to AI safety components in machinery, vehicles and medical devices.
The question a court, an insurer or a notified body asks is the same everywhere: what did you test, against what, and who signed? That file is Gate’s product.
Where Gate sits
Every layer around an AI agent has an owner, except the one that produces evidence a third party can rely on.
Certification schemes and specialist AI insurers. Today they certify and price on documentation and questionnaires — they need a testing party.
Signed, portable, re-verifiable test reports, for the developer who ships an agent and for the enterprise that adopts it.
Register, route and report evidence against policy packs. They file evidence; they do not produce it.
Sandboxes, supervisors and safety controllers keep an agent inside its boundary. They keep a run log, not independent proof.
Traces and scores for the team building the agent — useful, and not evidence a buyer, a certifier or a court can accept.
Gate wraps the runtime rather than replacing it: it injects the scenarios, reads the runtime’s own logs and signs what happened. A lab owned by the runtime vendor is not a lab — neutrality is the product.
How it works
An agent build or a robot policy, together with the policy it ships under.
Inside an OpenShell-class runtime — or, for physical AI, a partner simulator and hardware rig.
passk with intervals, never a single mean, and a judge model from a different vendor than the agent under test.
Obligations compiled into runnable checks by practising lawyers, from oulution.ai.
Re-verifiable, and bound to the version, the policy hash and the jurisdictions.
From a pull request, a release tag, a Jira or Linear command, or a step in an agentic software-delivery workflow. The signed report posts back as a check, where the developer already works.
An attempted boundary crossing read from the runtime log counts as a failure even when the runtime stopped it, because it reveals intent.
Open: the scenario schema, the SDK, the report verifier and the runtime adapters. Closed: the decision engine, report issuance, legal content and twins.
Three evidence classes
Law as tests
This is the core of Gate. Every clause follows one path, and every mapping carries two signatures: the practising lawyer who wrote it and the engineer who wrote the test.
The legal content is compiled from oulution.ai, Oulution’s cross-border legal and compliance platform: a curated jurisdiction corpus, practitioner workflows from more than 200 cross-border transactions, and a network of 9,500+ vetted lawyers. That is the half of the crash test no evaluation tool and no runtime vendor brings.
JURISDICTION PACKS, IN ORDER| Legal obligation | Becomes, in Gate |
|---|---|
| Colorado AI Act impact assessment | A documented scenario run plus a signed attestation |
| SR 11-7 independent model validation | A deployment-acceptance (Type D) report |
| NYC Local Law 144 bias audit | A bias suite with passk results |
| FTC substantiation of AI claims | The evidence file behind each claim made about the agent |
| Data residency (GDPR, PIPL, PDPA) | An egress policy plus a runtime log of zero denials — provable |
| Sanctions screening | A list-oracle test: the sanctioned name is refused and reported |
| Human-oversight duties | An approval-gate test: no irreversible action without sign-off |
| ISO/TS 15066 speed and separation on a cobot | Does the learned policy ever ask to cross the envelope? |
The rows illustrate the method. A Gate report records what was tested and what happened; it does not decide what any law requires of a particular product.
Run on our own product first
The first system under test is Ouli Refund, Oulution’s live US tariff-refund product, fired from a pull request. Gate reported and never repaired: Refund’s own engineer made the fixes, because a lab that repairs what it tests is not a lab.
The gaps sat in human approval, credential handling, network egress, sanctions screening and resistance to instructions hidden in documents — each traced to the scenario that exposed it and the obligation it breaks.
A host allow-list with no wildcard, human approval on filing and sending, and approvals that expire — then six product changes, from a credential broker to an operator stop that does not depend on the agent.
The new version digest and policy hash voided the 1.2.0 report automatically. The re-run fired from the pull request and posted back as a check.
HARNESS MVP BUILD 0.4 · RUNS SIMULATED; THE CONTROL GAPS AND THE FIXES ARE REAL (READINESS SPRINT, 6 OCTOBER 2026) · THE REAL RUN, AT N = 96, IS SCHEDULED FOR THE WEEK OF 27 OCTOBER 2026 · WATCH IT IN THE DEMO ↑
One evidence core
Every report carries its scope, version digest, policy hash, jurisdictions, date and the conditions that invalidate it.
Harness software and readiness services produce the evidence; a partner certification body issues any certificate, under its own scheme.
Bound to the version digest, the suite and the packs. Voided by a new version, a suite upgrade or a pack update.
Bring-your-own-policy acceptance testing inside the adopter’s own environment and jurisdictions, a registry lookup, and an automatic re-check when the vendor ships a new version. Deployer duties are arriving; today’s evaluation tools sit on the developer side.
Bound to the deployer’s policy hash, jurisdictions and twin. Voided by a policy change or a vendor update.
For robots, drones and vehicles, the learned policy is tested in partner simulators and hardware-in-the-loop rigs against its safety envelope. The hardware stays with the functional-safety certifiers.
Covers the AI decision layer only, as tested in simulation and on the stated rig. Never a machinery safety assessment.
General counsel and procurement deciding whether to switch an agent on · certification bodies issuing certificates under their own schemes · insurers pricing cover · regulators and courts asking what was tested before the agent acted.
Red lines
Assurance that oversells itself is worse than none, because the reader stops checking. So, plainly:
Gate issues test reports. Certificates come from certification bodies, under their own schemes, on Gate’s evidence.
Testing before deployment reduces the probability of an unsafe or non-compliant release and preserves the evidence trail. It does not prevent harm, and no Gate report says it does.
The people who help a customer write its policy never assess that customer: separate people, separate contracts, and assessment records open to inspection.
Physical-AI reports cover the learned decision layer as tested in simulation and on the stated rig, and state the sim-to-real gap. No physical report goes out without a certification partner in the loop.
What exists today
Gate grows out of the release discipline behind oulution.ai, Ouli Refund and Ouli Redact — three live products with paying enterprise clients: a synthetic-persona harness that runs at boot and every 24 hours, a guard fleet that checks every production response, golden-question and redaction test suites, report signing, and a named-reviewer record of every adoption.
Harness MVP build 0.4 exists: the runner, the pull-request trigger, the step taxonomy, the signed report and an open verifier. There is no automated gate in CI and no twin yet. The first real run, at N = 96, is scheduled for the week of 27 October 2026.
CODEBASE INVENTORY, 2 OCTOBER 2026
Roadmap
The physical lane starts no engineering until Gate 2. Missing a gate parks Gate as an internal tool — which still gives Oulution the independent proving ground it uses on its own releases.
The company and the team
Brown University, Applied Mathematics–Economics. Credit Suisse, Rothschilds, Bank of America, Rainy River — built an AI diligence pipeline that cut deck preparation from six hours to one.
Architect of the evidence core and owner of the OpenShell integration. Led Alibaba B2B’s technical shift to mobile at scale; systems at 10M+ daily users and 1M+ queries per second; software on 100M+ devices; co-author of a cloud-native architecture text.
Senior equity partner at an international law firm; 200+ cross-border transactions; FT Innovative Lawyer and Legal 500 Leading Individual. Built oulution.ai, the platform Gate’s legal compiler runs on, and owns the legal content of the test-method library.
A robotics-background technology director or partner on co-founder terms (simulation, ROS 2, sim-to-real); a functional-safety advisor before any physical pilot; and a second engineer, based in the US.
Oulution Inc. (Delaware) is a subsidiary of Oulution Pte. Ltd. (Singapore). Group credentials: NVIDIA Inception member; information security management certified to ISO/IEC 27001:2022.
Agent developers, enterprises accepting third-party agents, certification bodies and insurers: tell us what you need to prove, and to whom. Investors: the memorandum and the deck are available on request.
B2B ONLY · GATE-TESTED, NEVER CERTIFIED · INFORMATION, NOT LEGAL ADVICE