Building the ethical infrastructure for the automated age, designed to govern AI and protect our humanity.
Gatekeeper is the manager in the room for your AI.
It checks every decision against your rules before the AI acts, and it keeps a sealed record of what happened.
We do not extract, we steward.
The problem
One screenshot is all it takes.
Employees paste sensitive data into AI tools every day. One screenshot and a lawsuit is all it takes. Existing tools try to catch every bad phrasing after the fact. That game cannot be won, because the variations are infinite.
What we do
Govern the decision, not the string.
Gatekeeper does not chase phrasings. It asks one bounded question before any action runs: is this action permitted under the rule that applies. Detection chases infinite surface forms. Enforcement asks one question.
Every action gets a verdict before it executes, and a sealed record after.
| Verdict | What happens | Example |
|---|---|---|
| GREEN | Action permitted, sealed and logged. | A routine question with no sensitive data. |
| YELLOW | Action modified or flagged before passing. | A bulk data request that needs scoping. |
| RED | Action blocked before it ever executes. | A request to expose a customer SSN. |
A tamper-evident record. Hash chaining, not blockchain.
Every decision is sealed with a cryptographic hash that includes the one before it. Change any record and the chain breaks visibly. No network, no token, no consensus overhead. Just a tamper-evident ledger of every governed action — an enterprise firewall for AI that enforces the rule and logs the result.
Lower compute cost
The AI does not waste cycles on actions that should never run.
Save water
Less wasted compute means less data-center cooling demand.
Cryptographic accountability
A provable record of every decision, for audit and compliance.
The live assistant
Watch it govern itself.
Ask the assistant anything about Gatekeeper. Every message you send is checked by the governance layer before a reply is produced — you watch the verdict and the sealed record happen live.
I'm the Gatekeeper assistant. Ask me anything about the product — every message you send is governed before I answer, and you'll see the verdict sealed into the record on the right.
Sealed record · hash chain
No decisions yet. Send a message or try a preset — each governed decision is sealed here.
Answers are live, grounded on our company wiki. The governance verdicts are a demonstration: a rules-based classifier runs in your browser and seals a real SHA-256 hash chain you can verify. This honesty is the brand.
Architecture and trust
On your network. No cloud in the decision path.
- 01
Gatekeeper installs as a standalone executable on your own network. Your data never leaves your environment.
- 02
The answer engine runs locally. The governance layer runs locally. The record is sealed locally.
- 03
Prompts are never stored. Sensitive data is masked on your side and excluded from records. Identifiers like SSNs never persist anywhere.
- 04
Built for your compliance regime — HIPAA, financial, legal — with policy packs derived from the actual rules that govern your industry.
Why now
Why now.
On June 12, 2026, Anthropic suspended access to Claude Fable 5 and Mythos 5 for all users after new U.S. export controls took effect immediately. Access was restored July 1 after the controls were lifted. The operational lesson is broader than one model: model availability and behavior are external dependencies. Governance that must remain enforceable should not disappear when a provider, model version, or cloud surface changes. Gatekeeper keeps policy enforcement and the decision record outside the model, on infrastructure the operator controls.
Source: Anthropic: Redeploying Fable 5 →The engagement
A 21-day pilot, in three phases.
- Days 1–4
1. Shadow audit
Observation only. Gatekeeper watches and builds the policy pack from your environment.
- Days 5–11
2. Guided trial
Enforcement on. Every action is governed and sealed on the installed executable.
- Days 12–21
3. Expansion
Widen coverage across teams and workflows once the value is proven.
The onboarding fee funds setup and your client hardware; a monthly floor follows. Exact figures are confirmed on the onboarding call.
Built in the open
Built in the open.
Seven public open-source repositories now make up the current portfolio: six MIT-licensed projects and one Apache-2.0 project. They expose different parts of the work — pre-execution authority, physical AI, agent handoffs, durable memory, governed claims, on-device control, and the research lineage behind them — while production Gatekeeper internals and private research stay separate.
Gatekeeper: Physical AI
MITALLOW / TRANSFORM / HOLD / DENY
A greenfield physical-AI integration that places an independent authority boundary between learned policy or planning and robot actuation. The public repo records governed command binding, denials, approvals, execution receipts, and the event-hardware path.
AI INFRA SUMMIT 2026 / Intel Physical AI Challenge.
Team Oak & Sparrow Systems Enterprise LLC
OakandSparrowSystemsEnterprises/AI-Infra-HackathonView on GitHub →
AWS Memory to Muscle
MITMEMORY ≠ AUTHORITY
An agentic coding workflow where semantic memory, live repository state, security evidence, orchestration, and reusable procedure can improve work across runs without inheriting execution authority. Every consequential effect requires a fresh Gatekeeper authorization bound to the exact action.
Built for the September 11, 2026 AWS Data & AI Hackathon.
Team Oak & Sparrow Systems Enterprise LLC
OakandSparrowSystemsEnterprises/OASSE-AWS-Memory-MuscleView on GitHub →
Agent-Native Arena
Apache-2.0ARTIFACT MOVED. AUTHORITY DID NOT.
An agent-native integration proving that work artifacts can move between agents without carrying execution authority with them. Each agent must independently satisfy Gatekeeper before a governed effect can occur.
Built for the Agent Native Builders Hackathon at Cloudflare HQ, August 26–27, 2026.
Team Oak & Sparrow Systems Enterprise LLC
OakandSparrowSystemsEnterprises/Runtype-hackathonView on GitHub →
BoundaryCast
MITYES / NO / UNRESOLVED
A governed weather oracle. It checks the evidence first — official forecast, nearby observations, active alerts, freshness, uncertainty — then makes the most specific claim that evidence supports, and no more. When the evidence is thin it degrades gracefully instead of guessing, and the same governed claim resolves weather-dependent prediction markets against an evidence-bound artifact. Every decision is a hash-chained, replayable proof.
Built by Mool Tyagi and Joshua Johosky.
Team Mool Tyagi, Joshua Johosky
OakandSparrowSystemsEnterprises/BoundaryCastView on GitHub →
Sentinel-Med
MITALLOW / HOLD / BLOCK
A human-in-the-loop medical chatbot. A deterministic classifier is the safety floor; the model can only escalate caution, never override a block. Every decision is written to a tamper-evident hash chain.
Built for the AI Collective Tri-Valley “Humans In AI” track.
Team Aufstin Filiko
OakandSparrowSystemsEnterprises/SentinelMedView on GitHub →
Oak & Sparrow OS
MITPASS / HOLD / VETO
An on-device agent-governance kernel for Android. Risky actions are held for approval; every decision is Ed25519-signed and content-addressed in an audit log. Runs entirely on-device, no cloud — the same on-prem principle Gatekeeper enforces for the enterprise.
Public Android governance kernel and assistant implementation.
Team Joshua Johosky, Jackson Castro, Caleb Strom, Skylar Smith, Mool Tyagi
thespacekyd-eng/Oak-Sparrow-OSView on GitHub →
Cognitive Archive
MITPUBLIC RESEARCH ARCHIVE
A public archive of Joshua Johosky's 2020–2025 work across AI governance, economic infrastructure, and consciousness-first systems design, including early SIP material and plain-language explainers. It exposes the research lineage without exposing Oak & Sparrow's private production corpus.
Released as an MIT-licensed public research archive.
Team Joshua Johosky
OakandSparrowSystemsEnterprises/cognitive-archiveView on GitHub →
Public code and research are intentionally separated from private production Gatekeeper internals, proprietary policy assets, and in-house research. Open repositories are proof surfaces, not the production system.
From the blog
What we're learning as we build.
August 4, 2026
Gatekeeper External Release 1.0
Deterministic Execution Control for Enterprise AI Agents, the current external Gatekeeper technical product white paper.
Oak & Sparrow
July 10, 2026
Why governance has to sit outside the model
Safety that lives inside a model can be pulled or jailbroken out of it. The manager in the room cannot be talked out of the room.
Joshua Johosky
July 3, 2026
Building in public, on purpose
If we ask clients to trust us with governance of their AI, the least we can do is govern ourselves the same way, in the open.
Nick Silva
Onboarding
Start your trial, no sales call required.
Qualify, see the fit, and book your onboarding call — right here. We route a warm, informed start straight into setup.
- 1Qualify
- 2See the fit
- 3The engagement
- 4Your details
- 5Book the call
1 · Qualify
A few questions to route you to the right policy-pack starting point.
Not an enterprise? Founder Forge
Founders, AI builders, and independent inventors: turn a messy idea into a fundable structure — claim, proof ladder, artifact map, and revenue path. Get a live quote and a founder-safe brief.
Who it is for
Built for firms that have to prove governance.
Legal
Privilege and client confidentiality, provable on every action.
Insurance
Claims and customer data governed before it ever reaches a model.
Healthcare
PHI masked and HIPAA rules enforced before the AI acts.
Community banking
Account and financial data scoped, logged, and audit-ready.