HullSmith Firewall
Catch the hallucination before your users do.
Three independent checks sit between any language model and whatever comes next — your users, your pipeline, your research record. Grounding, compression distance, claim verification. Deterministic math you can re-check. Not a classifier.
03 · The checks
CHECK 01
Grounding
PASS · anchored in source
BLOCK · confabulated
CHECK 02
Compression distance
PASS · useful middle band
BLOCK · parrot or untethered
CHECK 03
Claim verification
PASS · every claim supported
BLOCK · unsupported sentence
Walk the three checks with the engineer who runs them.
A G1 call is free. Bring a grounded and an ungrounded draft ; we run both through the checks, and put the verdicts in writing.
01 · Grounding
Anchored in source, or confabulated. Measured, not guessed.
02 · Distance
Useful middle band. Not a parrot. Not untethered.
03 · Claims
Every factual sentence checked against the source before it ships.
04 · Honest status
Production. Standalone SDK, or inside Connect.
3
Independent checks, on every answer
0
Training data — nothing to poison or retrain
100%
Auditable — re-check every verdict yourself
Cloud → Air-gap
Runs where your data lives
A firewall, not a vibe check.
Language models are fluent by default and correct only sometimes — and the fluent-but-wrong answer is the dangerous one, because it reads exactly like the right one. Firewall sits in front of that output and measures whether it's actually grounded in the source you gave it. The check is deterministic : the same answer against the same context always gets the same verdict.
Three checks. Every answer.
This page is document-grounding. Grounding, compression distance, claim verification — three independent checks of our own invention. Wrong answers get blocked or sent back before anyone downstream sees them.
The wall is HullSmith.
Five numbered gates, Horizon behind them, live on /products/maxor-shield. That is the L2.5 wall. This page is not a second HullSmith.
Wrap. Check. Decide.
Drop Firewall between your model and your users. One integration ; a verdict on every call.
01 — Wrap
Place Firewall in front of any model — OpenAI, Anthropic, a local Heisen engine, anything. Pass it the answer plus the source the answer was supposed to come from. No retraining, no model access required.
02 — Check
Three independent checks run in parallel against the source : is the answer grounded, is it appropriately novel, and is every factual claim actually supported. Deterministic math — no classifier, no training data.
03 — Decide
Get a verdict : pass, block, or return-for-revision — with the exact check that failed and why. You wire the policy : silent gate, human review, or auto-retry the model.
Three angles on the same question.
Each check is independent and measures a different failure mode. An answer only passes when all three agree — and any one of them can explain its own verdict in numbers you can reproduce.
Grounding
Does the answer actually follow from the source you provided — or did the model invent it? We measure how much of the output is anchored in the context versus confabulated on its own.
PASS · anchored in source
BLOCK · confabulated
Compression distance
Is the answer a verbatim parrot of the source (no value, and a leak risk) or untethered from it (a hallucination)? Only the grounded-but-useful middle band passes — measured, not guessed.
PASS · useful middle band
BLOCK · parrot or untethered
Claim verification
Every factual claim in the answer is checked against the source before it ships. A claim the source doesn't support is flagged with the exact sentence that failed.
PASS · every claim supported
BLOCK · unsupported sentence
Deterministic, honest, verifiable.
Built to be checked, not trusted on faith — and to drop into the stack you already run.
Drop-in firewall
Model-agnostic and provider-agnostic. Wrap any LLM, any RAG pipeline, any agent loop — Firewall only needs the answer and the source, not access to the model.
Bring-your-own-verifier
Don't trust the verdict — reproduce it. The math is open : re-check any decision yourself, on your own infrastructure, and get the same number.
Every block is explained
A blocked answer always says which check failed and on which sentence — actionable for a human reviewer, parseable for an auto-retry loop.
No training data, no telemetry
Nothing learned, nothing logged off-box by default. The whole check runs where your data lives — including fully air-gapped.
The failure modes that ship as facts.
Each of these reads like a correct answer. Firewall measures the failure directly, so a brand-new variant trips the same check — there's no signature list to keep updated.
Confabulation
Facts, figures, and quotes the model invented that appear nowhere in the source — the classic hallucination, caught before it ships.
Ungrounded claims
Assertions the provided context simply doesn't support — plausible, fluent, and unbacked.
Broken citations
Claims attributed to a source that doesn't actually say it — the footnote that doesn't check out.
Parroting & leakage
Verbatim regurgitation of the source that adds nothing and can leak material you never meant to surface.
Confident wrongness
A fluent, authoritative answer with no factual footing — the most dangerous output, because it reads like the right one.
Context drift
Answers that start grounded and wander off the source mid-response, smuggling unsupported claims into a credible opening.
Hosted, or entirely inside your perimeter.
Start in the cloud ; bring it in-house when your security review — or your data-residency rule — demands it.
Cloud
Fully managed on Canadian-resident infrastructure. The fastest way to put a firewall in front of your model — nothing to provision.
Self-hosted
Run the engine inside your own environment for Enterprise deployments — your keys, your infrastructure, your control plane. Answers never leave your network.
Air-gapped
Fully disconnected for the most sensitive work. No telemetry, no phone-home — the check runs where the data lives, because the math doesn't need the internet.
Research records and production answers alike.
Anyone who ships a language-model output that has to be right — whether that output goes to a customer, a regulator, or into the scientific record.
One deterministic philosophy, across the suite.
Firewall shares the Maxor posture : deterministic math over black-box models, verifiable over trust-us, sovereign-deployable by default. This page is document-grounding. The L2.5 wall is /products/maxor-shield — not a second HullSmith.
Explore the Maxor suiteAvailability
—
Status
Production
Deployment
Cloud · Self-hosted · Air-gap
Owner
Maxor Digital · IP held by Maxor Global LLC
Put the firewall in front of any model.
A G1 call is free. Bring a grounded draft and an ungrounded one. We run both through the three checks and put the verdicts in writing. Production. Standalone SDK, or inside Heisen and MaxorConnect.
HullSmith Firewall, answered
HullSmith Firewall is a deterministic hallucination firewall — it catches bad LLM output before it reaches your users. It runs three independent checks of Maxor's own invention, using defensible math you can audit rather than another ML classifier that has to be trained and trusted.
Ground measures whether a model's answer is actually supported by its inputs. When the three independent checks disagree with the claim, it flags the output as likely fabricated — before it ships — so a confident-but-wrong answer is caught at the boundary, not by your user.
An ML filter is itself a model that scores "how likely this is a hallucination" and drifts over time. Ground's checks are deterministic math — the same input yields the same verdict every run, with a reason you can audit. No classifier to retrain, no black box.
Between your LLM and your users: any agent or app can wrap its model output in Ground before it is shown or acted on. It runs inline, in real time — the firewall adds a check, not a step in your workflow.
Yes — deployment is hybrid: in the cloud or on your own hardware. Because the checks are deterministic and self-contained, they run identically wherever deployed, with no external model call.