Blockhouse

Blockhouse: AI Sandboxing With a Hard Stop.

The Blockhouse sandbox isolates AI workloads on hosts you control. Models get the files, runtime paths, and communication channel you approve. Attempts to step outside the sandbox are blocked, halted, and recorded.

Evidence

Evidence, Not Promises

Sandbox status Each run records whether the sandbox is active and healthy.

Approved access The report shows what the model is allowed to read, write, and call.

Monitored communication All model traffic moves through one audited channel: Porthole.

Halt record Blocked or halted behavior leaves evidence your team can review.

Clear limits If a host cannot provide the expected protection, Blockhouse says so clearly.

Integrity

Seal Once, Enforce Always

Model seal Approved model files are sealed before serving.

Refuse drift If model files change after approval, serving is refused.

Recheck on demand Teams can recheck a runtime before or during operations.

Content identity Approved artifacts stay tied to what was reviewed.

Reviewable records Evidence is written for audit review, not hidden in an opaque process.

Sandbox

Works Where Your Models Run

Model sandbox The model runtime is placed inside a controlled operating boundary.

Porthole access Users and apps reach the model through one monitored endpoint.

Key separation Runtime credentials and audit controls stay outside the model process.

Operating record Setup, access, and runtime decisions are recorded for review.

Control

For Hosts You Control

Your infrastructure Run the sandbox on hosts selected and managed by your team.

Resource control Keep model serving inside defined operating limits.

Isolated runtime Keep the model process separated from systems it does not need.

Separated roles Preparation, launch, and serving are handled as separate operating responsibilities.

Engines

One Sandbox, Multiple Serving Engines

vLLM Default GPU serving engine.

SGLang Throughput-oriented sibling engine.

llama.cpp GGUF models and CPU-capable serving lane.

Serving plan Each runtime choice is captured before the model is served.

Same boundary The same sandbox and monitored channel apply across supported engines.

Boundaries

What Blockhouse Does and Does Not Claim

Supported hosts Blockhouse runs on supported Linux environments selected during deployment.

Real protection The sandbox reports what protection the host can actually provide.

Honest evidence If a deployment cannot support a stronger claim, the evidence says so.

Reviewable source Security teams can review source and release materials during evaluation.

Known limits Blockhouse is a sandbox for AI workloads, not a replacement for trusted infrastructure and sound operating practice.

Next Step

Do you want to see the source code? We can do that.

Request Source Access