Gatekeeper: The AI Governance Layer

A pre-execution kernel that filters AI output and users' input against any relevant laws or regulations.

Project Details

This is our best attempt to put brakes on a system that is rapidly evolving faster than we can understand it. LLM's currently do not show the work or provide receipts for it's work. Drift is real, hallucinations are real, and the people putting sensitive data into thee machines are real.
This started as a base question, "How do we protect the kids?" Our solution was to bake the actual laws written by humans and approved by judges would be the best way to accomplish this end goal. To this end we have successfully launched our flagship product Gatekeeper. This is AI Governance infrastructure with MITM, Brower extensions and plug in ready to go. Pre-input, pre-execution is our lane, we sit between when the user sends a command to the AI and the actual time the AI takes to start the work. In our model the AI must present a Json file of what it wants to do before it does it, and that is equal for both gates.
This happens in sub milliseconds per input.
I coded the laws into VsCode, used the laws to make tenants and from there wrote the O.W.L Ontology. Using this Ontology we created a rule engine which is Flask API driving an OWL reasoner; allows us to actually use the code we wrote.
Joshua Johosky and Nick Silva are the Founders along with Mool Tyagi the CTO.
We have established an advisory board consisting of two GTM's, a planetary scientist and a financial advisor.

We exist to help anyone using AI understand the ways to communicate while adhering to the laws written by elected officials.

Theory of Impact

The dominant AI risk model treats catastrophe as a function of capability. We treat it as a function of unaccountable execution: a system acting on the world faster than any human or institution can verify what it is about to do. Capability keeps climbing; the verification gap is what turns capability into harm.

Gatekeeper narrows that gap at the one point where it can still be closed, before execution. Every command an AI receives must first emit a structured declaration of what it intends to do. That declaration is checked against an explicit, human-authored rule set, encoded from actual law, and resolved deterministically to permit, block, or escalate, in 25 to 300 microseconds, before the action begins. The same gate applies whether the AI is answering a child or moving money.

This reduces x-risk along three lines. First, it makes AI action legible: the system produces a receipt for every decision, the rule that fired and the verdict, so behavior can be audited rather than inferred. Second, it makes governance deterministic rather than probabilistic: the same input always yields the same verdict, so the control does not itself hallucinate. Third, it is : the governing layer sits outside the model and outlives any single one, so it continues to apply as models grow more capable, which is precisely the regime where post-hoc monitoring fails.

People

Joshua Johosky

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$275,000

Ideal

$612,000

What this ask is

We exist to help anyone using AI understand the ways to communicate while adhering to the laws written by elected officials.

How the money will be spent

At the minimum ($275,000), the raise takes Oak & Sparrow from a pre-capital LLC to an operating company: Delaware C-Corp conversion and the legal and corporate groundwork to stand up as a fundable entity, compute and infrastructure to run the Gatekeeper engine and sealed artifact ledger in production, the reasoner and validation tooling that keeps the governance layer verifiable, and the runway for the founding team to carry the work through conversion and into an institutional seed round. This is the amount that makes the company real and the product deployable.

At the ideal ($612,000), the same foundation is funded and then accelerated: continued development of the deterministic governance LLM for economic-entity oversight, expansion of the encoded policy packs beyond the initial California disclosure domain into additional regulated verticals, the security and reproducibility hardening required before institutional and banking deployment, and the pilot-to-production engineering that lets multiple clients onboard in parallel rather than one at a time. The additional capital buys speed and breadth: it turns a working prototype into a scaled governance platform on the timeline the market is moving.

Discussion

MMark Aiken16d

This is a really interesting idea, although the amount requested seem larger than the current grant window.

I might have missed it, but when you say you bake the actual laws into the system - which laws are you using? Laws from all around the world? How are you managing questions of jurisdictions, where something might be permissible in one jurisdiction but not another? How are you drawing redlines around permitted / nonpermitted content? Eg the anarchist cookbook might be legal as free speech in some jurisdictions and restricted under terrorism laws in. Would you go based off the local laws wherever you are? or is there also some kind of moral / extra-legal filter built in to make judgements on objectionable content?