Check before delivery
Compare the answer with its sources.
AI OUTPUT FIREWALL + HULLSMITH HORIZON
Your AI acts.Demand
evidence.
AI Output Firewall checks your model’s answers against their sources. Hullsmith Horizon adds information-flow checks at the agent boundary. Find the gap, hold a risky output and inspect the evidence behind the check, before your application decides what happens next.
Real engine demo. No account required to try it.
THE VERIFICATION BOUNDARYComputing
- SUPPLIED FLOW
- —
- GROUNDING SCORE
- —
- SIGNED DIAGNOSTICS
- —
01 / INPUTAgent
THE CONTROL POINTHorizonTwo checks. One boundary.
02 / DECISIONNot verified
INFORMATION FLOW × ANSWER GROUNDING
Leakage—
Grounding—
The engine is preparing this example.
Example inputs · real engine · no live traffic capture
Built for the stack you already run
RAGAI copilotsAgentsAPIYour infrastructure
Govern agent egress
Apply a policy at the boundary.
Inspect the evidence
Keep the scope of each check visible.
Choose your deployment
Local engine, hosted API or sovereign service.
TWO PRODUCTS. TWO CONTROL POINTS.
What needs a boundary?
Start with AI Output Firewall to check generated answers. Choose Horizon when you also need information-flow analysis at the agent boundary, combined with answer grounding.
Before the answer goes out.
Check generated text against the knowledge you supply. Surface factual gaps, route cases that need semantic review and inspect the resulting certificate.
- Grounding checks for RAG and copilots
- Deterministic first pass, semantic escalation
- Local checks or an authenticated API
Before the agent crosses the line.
Bring information-flow analysis and answer grounding into one firewall. Horizon combines both diagnostics to allow or block an agent emission. A grounded answer can still carry an information-leakage risk.
- Combined leakage and grounding checks
- Local semantic inference in the demonstrated path
- Inspectable evidence for each diagnostic
01 / TRY AI OUTPUT FIREWALL
How do you check an answer before sending it?
Give the Firewall a source and an answer. It runs a deterministic check and reports factual gaps or a need for semantic review. Change the answer, run it again and inspect the actual certificate.
DETECT → DECIDE → VERIFYReady
What did your AI add?
The source says 30 days. The answer says 90. Run the check to see what the engine actually detects.
- Engine + replay time
- —
- Certificate replay
- —
Inspect the actual certificate
Scope of this demo: deterministic Tier 1. It distinguishes a literal match, a numeric contradiction and a need for semantic review. A match does not establish that the source is true. The certificate is unsigned, neither retained nor billed; replay checks the computation, not the truth of the answer. An escalation flag requires further review.
02 / HULLSMITH HORIZON
A correct answer.
A dangerous emission?
An agent can stay faithful to its sources and still leak information. Horizon checks both dimensions together. Change one risk below: the real engine recomputes the diagnostics, and the boundary shows the decision.
Agent context
Loading the example…
GENERATED ANSWER
—
HorizonTWO DIAGNOSTICS. ONE DECISION.
Information leakage—
—
Answer grounding—
—
Either diagnostic can block passage.
Not verified
No computed authorization.
Horizon is checking both dimensions.
The result appears as soon as computation completes. The animation explains it.
FLOW DIAGNOSTIC · SHA-256
—GROUNDING ATTESTATION · SHA-256
—MEASURED COMPUTE—
Inspect the inputs and both signed diagnostics
Two separate signatures with an ephemeral key for this demo instance. There is no signed certificate for the combined decision. The flow certificate reports 3 replayed steps and 2 unreplayed steps; inspect their scope below.
Real computation on synthetic example telemetry, with local semantic inference. No network traffic is captured and no external agent is protected by this demo. Your deployment must instrument the relevant paths and apply the decision.
Scope my Horizon deployment ↗
Your request goes to [email protected]
INSIDE YOUR STACK
Your stack.
An explicit boundary.
Keep your model, your sources and your deployment. Put Hullsmith at the point where an AI output becomes an application action. Your integration supplies the context and applies your policy using the diagnostics.
Integration schematicLOCAL OR API
- CONTEXTSources
Your knowledge
Documents, RAG retrieval and reference facts supplied by your application.
- GENERATIONLLM / RAG
Your model. Your agents.
Generated text and, for Horizon, information-flow measurements from your instrumentation.
- THE VERIFICATION BOUNDARYCheck
Hullsmith
- APPLICATIONAction
Your policy decides.
Deliver, hold or send for review. Your application enforces the action using the diagnostics.
CHOOSE THE CONTROL
Which product fits which risk?
Start at the boundary that matters for your workflow. The two products can complement each other when an agent both generates answers and emits information to external systems.
Swipe to compare Firewall and Horizon →
| Your question | AI Output Firewall | Hullsmith Horizon |
|---|---|---|
| What is checked? | An AI answer against supplied source knowledge. | Information-flow risk and answer grounding together. |
| What decision follows? | A factual finding or semantic escalation that your application handles. | Allow or block. Under the demonstrated policy, either diagnostic can block passage. |
| What is the evidence? | A local certificate; signed, retained receipts through the eligible API path. | Two separately signed diagnostics in this demo. No signed certificate for the combined decision. |
| What remains your responsibility? | Source quality, review of unresolved cases and application behavior. | Source quality, telemetry, policy and instrumentation of the relevant agent paths. |
FROM EVALUATION TO YOUR STACK
How does Hullsmith fit your application?
Put the verification call between the generated answer and delivery. Send the answer and the source knowledge to the authenticated API, inspect the result, then apply your handling policy. For local or sovereign deployment, choose the runtime and signing boundary with your team.
Open API documentation ↗Sign-in required. Keep the API key on your server.
POST /api/verifyServer-side
curl https://app.hullsmith.com/api/verify \
-H "Authorization: Bearer $HULLSMITH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"knowledge": "Returns accepted within 30 days.",
"answer": "Returns accepted within 90 days."
}'Know what was checked.
Inspect the inputs, engine version and parameters carried by the certificate. Keep the decision attached to its actual scope.
Recheck the certificate.
Replay the verification artifact. A structurally valid certificate alone does not authenticate a signer or prove the source was true.
Separate checks from signed evidence.
Local checks and signed, retained receipts are different paths. Signature verification requires a trusted public key; retention follows the selected service.
PRICING · USD
What do you pay for?
Local checks do not consume API receipts. Hosted plans bill eligible signed, retained receipts. Choose a volume for your application; discuss Horizon and sovereign deployment with the team.
SDK Local
$0
USD / month
Local checks without a receipt quota. Hosted receipts counted below.
2,000 receipts / month
No paid overage on this tier.
Open an account ↗Starter
$99
USD / month
Move an application into production.
50,000 receipts / month
$2 / 1,000 additional receipts
View Starter ↗Growth
$499
USD / month
Increase verification volume.
500,000 receipts / month
$1 / 1,000 additional receipts
View Growth ↗Business
$1,499
USD / month
Support sustained application traffic.
3,000,000 receipts / month
$0.50 / 1,000 additional receipts
View Business ↗Scale
$3,999
USD / month
Plan for high receipt volumes.
13,000,000 receipts / month
$0.30 / 1,000 additional receipts
View Scale ↗Monthly billing shown, in USD, excluding applicable taxes. Included volume refers to receipts, not unlimited hosted calls. Annual billing options and the exact subscription terms are available in your account.
Billing options ↗Horizon, private infrastructure or a specific boundary?
Define the agent paths, information policy, signing and operational requirements for your deployment.
BEFORE YOU INTEGRATE
What should you know first?
Clear boundaries make a control useful. Here is what Hullsmith checks, what its evidence means and what remains in your hands.
What is an AI output firewall?
An AI output firewall is a control between a generated answer and its delivery. Hullsmith compares that answer with supplied knowledge to identify factual gaps and cases requiring semantic review. Your application uses the result to decide whether to deliver, review or stop the answer.
How is Horizon different from Firewall?
Firewall evaluates an answer against source knowledge. Horizon combines information-leakage analysis with answer grounding at the agent boundary. In the demonstrated fusion policy, either diagnostic can block passage. Its coverage depends on the sources, telemetry and paths supplied by your integration.
Does Hullsmith detect every hallucination?
No. A deterministic factual check does not cover every semantic error. A negation can preserve the same numbers and entities while reversing the meaning. The demo includes this boundary case. Source quality, semantic review and your application's handling of unresolved results remain essential.
Do my prompts leave my infrastructure?
It depends on your deployment. A local engine can run inside your infrastructure; the hosted API receives the text you submit. A sovereign deployment requires provisioning its runtime and any semantic models inside your perimeter. The demonstrator processes inputs on the Hullsmith server without storing them in an application database.
What does a verification receipt prove?
A receipt records a specific check and its provenance. An authenticated signature identifies its signer; it does not establish that the source is true or that all emissions were captured. A local unsigned certificate is distinct from a signed, retained receipt. Neither alone constitutes a compliance certification.
Where should a team start?
Start with representative sources and generated answers. Evaluate detected gaps, false alarms and semantic escalations against your own acceptance criteria. For Horizon, map every relevant emission path and define the information policy before deployment. Open a Hullsmith account for API access, or contact the team to discuss a sovereign deployment.
Give your AI a boundary.
Keep the evidence.
Try the engine, inspect the result, then bring the right control into your workflow.
LET’S TALK ABOUT YOUR DEPLOYMENT
Your use case.
The right boundary.
Firewall, Horizon or private infrastructure: tell us what you are building. Your request goes directly to [email protected].