Part of ZenithAI · Intelligence Framework

A reasoning engine, not a wrapper.

Deploying an open-weight model is where most “private AI” stops. It is where ZenithAI starts. Between a user pressing Enter and an answer appearing, a sequence of deterministic stages decides how the question is understood, which evidence is allowed to answer it, and whether the answer may stand at all. This is the machinery that turns a capable model into intelligence an organisation can trust, audit and own.

See it reason on your material The full platform

The distinction that matters

The model is step zero. Everything above it is the product.

A capable open model is a commodity anyone can download. What decides whether an enterprise can rely on it is the reasoning layer around it: the rules that route a question, the evidence that grounds an answer, and the contracts that refuse a claim the system cannot back. ZenithAI is that layer, and it runs entirely on your own infrastructure.

Sec. 01 Design principles

Six ideas hold the whole engine together.

These are the principles a CIO can repeat to a board. Every stage that follows is one of these principles made concrete.

Principle 01

Layered defence

Precise, reviewable rules decide routing at the top; a semantic layer interprets paraphrase and multilingual phrasing beneath them; and independent evidence checks audit every answer at the bottom. A wrong answer has to defeat all three layers, not one.

Principle 02

Classify, never answer

Models may classify what a question is; they never author the consequence of a decision. Dedicated executors answer only from their own verified evidence — so a persuasive prompt cannot talk the system into the wrong action.

Principle 03

Deterministic precedence

The order in which a turn is decided is fixed and reviewable, not left to a model’s mood. The same question follows the same path every time. That repeatability is what makes the platform auditable.

Principle 04

Fail-closed by doctrine

Every learned component degrades to the deterministic safe path on any failure, timeout or malformed output. “Unavailable” becomes a safe word, never an outage and never a wrong answer.

Principle 05

Honesty as a runtime property

Not-found answers name what was searched. Answers that combine your documents with the web say so. A reply that claims research it never did is replaced before you see it, and truncated answers are flagged. Honesty is enforced by code paths, not by asking the model nicely.

Principle 06

Fully local and private

Models, embeddings, search and OCR all run on your own infrastructure, and your private workspace terms are stripped from any outbound query. Nothing about your documents leaves the platform.

Sec. 02 The pipeline

Ten stages between Enter and an answer.

The stages are not a rigid waterfall — grounded answering, web research and memory are alternative evidence sources chosen by the routing stages — but the precedence between them is fixed and deterministic. Four movements, ten stages.

Four movements, ten stages. Each stage below is a running capability of the platform.

IF-01Input & normalisationEstablishes who is asking, protects what they typed (secrets are scrubbed from every reply and log; your words are never rewritten), and safely ingests attachments — including scanned documents through local OCR. Pasted content is treated as data, never as an instruction.Operational

IF-02Deterministic policy gatesYour explicit wishes about web search are enforced by exact, reviewable rules, not a model’s judgment. “Answer only from my documents” is honoured for the session, and an instruction buried inside a pasted article can never flip that switch.Operational

IF-03Workspace policy layerA workspace answers from your documents by default, not the internet. When a question genuinely reaches beyond your documents, it answers from both the documents and the web, and says so honestly.Operational

IF-04Deterministic routersQuestions with exactly one right mechanism — list your files, look up a number, compare two documents — are recognised by precise structural rules and answered by dedicated executors, before any general model reasoning. Fast, exact and immune to prompt tricks.Operational

IF-05Grounded answering & retrievalAnswers are grounded in your documents with passage-level citations. Counts and totals run as real computations over the full data, not model guesses. When the answer isn’t in your documents, the platform says exactly that and names what it searched.Operational

IF-06Website knowledgeA public website becomes private, citable workspace knowledge — fetched, extracted and indexed on your own server, same-site and bounded by design. Website Knowledge →Operational

IF-07Web researchWhen a question needs the live web, the platform plans the search like a researcher — recognising freshness, expanding into complementary queries — and runs it through a self-hosted search stack, with your private workspace terms redacted from any outbound query.Operational

IF-08Memory & recallThe platform remembers what you told it in earlier conversations and brings it back when relevant — under strict relevance floors, so memory can never hijack an unrelated question, override your documents, or block a search it cannot actually replace.Operational

IF-09Evidence enforcementAn answer may only claim evidence it actually has. A reply that says it researched online on a turn where no search ran is replaced before you ever see it; an answer cut short by a budget is flagged, not hidden; and the underlying model vendor is never disclosed.Operational

IF-10Continuous self-evaluationThe platform grades its own routing on every turn, recording what the decision should have been against what actually happened. That evidence is what lets new capabilities be promoted safely — with a single-switch rollback if anything looks wrong.Always on

Every entry above decides real answers today. Capabilities still being calibrated are held in observation and are deliberately not listed here — the registry states only what is live.

Sec. 03 Precedence

The exact order a turn is decided.

This is the sequence every message runs through. Each gate either resolves the turn or hands it to the next — a fixed order, not a model choosing what to do. It is the clearest picture of how much has to go right, deterministically, before an answer is allowed to form.

#	Gate in the turn	Outcome if it fires
01	Authentication, rate limiting, session load, workspace re-authorisation	Access established, or the request refused
02	Search policy resolved — honour “search” / “don’t search”	Web permission set for the turn or session
03	Guarded spellcheck and secret redaction	Text normalised, never rewritten; secrets scrubbed
04	Retry-feedback evaluation	Feedback on a previous answer is handled
05	Call confirmation and call-status questions	Call-flow reply from the recorded outcome
06	Connected-mailbox operations	Mailbox reply
07	Workspace manifest — file-listing questions	Deterministic file list, never paraphrased
08	Workspace knowledge injection — stuff, retrieve or map-reduce	Document context assembled
09	Workspace web decision and search plan	Search allowed, denied, or a clarification returned
10	Recall gate, then web search, then recall injection	Evidence assembled from the right sources
11	Attachment build and missing-document guard	Guided refusal if a named document is absent
12	Document-generation intent	A Word, PowerPoint or Excel artifact
13	Product-identity and fragment guards	Deterministic reply or a clarification
14	Grounded answering — compute-first, table search, cited answer, widen-on-miss	A cited, grounded answer
15	Budget fit and output-room reservation, then the model call	The answer is generated
16	Research-honesty guard and truncation flags	The answer is corrected or flagged
17	Self-observation snapshot, then persist	The decision is recorded for review

Seventeen gates, in this order, on every turn — and the streaming path maintains the same order, with one safety exception: an armed call confirmation outranks everything.

Sec. 04 Guardrails, enforced in code

The answer has to earn the right to stand.

Most AI failures are not wrong facts — they are confident claims the system cannot back. ZenithAI closes that gap after the model has spoken: before an answer reaches the user, it is checked against what actually happened on the turn.

These are not prompt instructions a clever message can talk around. They are code paths the answer must pass.

01Evidence contracts — every claimed source of an answer must be provably present. A citation to evidence that was never retrieved invalidates the answer.
02Fabricated-research replacement — if a reply claims it searched the web on a turn where no search ran, the answer is replaced with an honest failure before it is shown or saved.
03Vendor never disclosed — the assistant never introduces itself as its underlying model or maker; the tier answers as ZenithAI, on your own terms.
04Truncation honesty — an answer that hit its output ceiling is marked as truncated, never passed off as complete.
05Injection-safe evidence — text from documents, websites and the web enters as untrusted evidence; instructions hidden inside it cannot trigger tools or change how the assistant behaves.

Sec. 05 Plain-language glossary

The vocabulary, without the jargon.

A short reference for the terms a technical evaluation will raise.

Term	What it means
Workspace	A private, owner-scoped collection of documents, facts and a persona; a workspace answers from local knowledge by default.
Grounded answer	An answer whose every factual claim cites a real piece of retrieved evidence — not a free-form generation.
Evidence contract	The set of conditions that must be true for a route’s answer to count as verified before it is delivered.
Deterministic routing	Choosing how to answer by fixed, reviewable rules rather than by a model’s judgment.
Fail-closed	On any failure or timeout, the system falls back to the safe deterministic path rather than guessing.
Recall	Cross-session memory over the owner’s own earlier chats, brought back only when it clears strict relevance floors.
Compute-first	Answering a “how many / total” question by running an exact calculation over the full data, before any sampling.
Intelligence tiers	Instant, Fast, Smart and Genius — the platform’s own names for four levels of reasoning, run on locally hosted open-weight models sized to your infrastructure.

Models are examples, not commitments — the tiers run leading open-weight families selected for your deployment. Models & infrastructure →

The next step

See the reasoning on your own material.

A demonstration takes under an hour. Bring your own documents and questions, and watch the engine route, ground, cite and refuse — on infrastructure you control.

Request an Enterprise Demonstration Explore the platform