Skip to Content

MNB Veritas · The Examiner

Not a score. The source.

Many tools that check AI hand you a probability and leave the decision to you. MNB Veritas hands you the evidence. It takes an answer apart claim by claim, finds the exact sentence in your approved sources that supports or contradicts each one, checks every figure against the period it belongs to, and decides by written policy what may be stated as fact. It is the verification layer behind MNB's products and the enterprise systems we deploy.

Per claimEvery factual claim examined on its own, with its own verdict
VerbatimNo support without a quote matched character for character in the source
PolicyThe final decision is made by written rules in code, not by a model
PrivateRuns inside MNB systems and enterprise deployments. Not a public service

The field

How AI answers are usually checked, and where that stops.

Hallucination checking today mostly works in one of two ways: a detector scores how likely an answer is to be wrong, or a second AI model reviews the first. Both leave a model, or a threshold someone has to tune, with the last word. Veritas moves the last word to evidence and written policy.

The question an auditor asks
Score-based detection
AI checking AI
MNB Veritas
What comes back?
Score-basedA probability, or a pass or fail flag
AI checking AIA second model's opinion
VeritasA verdict for each claim, with the verbatim quote and the document it came from
Who makes the final call?
Score-basedA threshold your team has to choose and defend
AI checking AIThe reviewing model
VeritasWritten policy, executed as code: approve, qualify, abstain or send to a person
Right number, wrong year?
Score-basedLeft to how a model reads it
AI checking AILeft to how a model reads it
VeritasCompared by code, with every figure bound to its period, financial year, unit and scale
Is the checker independent?
Score-basedDepends on the model behind the score
AI checking AIDepends on which pair you choose
VeritasEnforced: the examining model comes from a different lab than the one that wrote the answer, and the report says so if none is available
What if a check cannot run?
Score-basedDepends on configuration
AI checking AIDepends on configuration
VeritasIt counts as not done. Veritas holds back more, never approves more
What does the file show later?
Score-basedA number to explain after the fact
AI checking AIA model's explanation
VeritasThe sentence, the source, the policy and the models consulted, for every claim

These describe approaches, not named products; many tools combine several. Whichever vendor you speak to, ask one question: what makes the final decision, and can you show me the evidence behind it?

The research standard

Built by a research firm. Held to research rules.

MNB Research began in risk advisory and institutional research, where a figure without a source does not leave the building. Veritas applies the same discipline to every AI answer, automatically.

Sourced

No claim stands without a sentence from an approved source behind it, quoted exactly.

Dated

Each source carries the dates it is valid for. Out-of-date and superseded documents are flagged.

Reconciled

Figures, units, lakh and crore, percentages and financial years are matched by code to their period.

Reviewed

A second, independent model on what matters and, at the stricter levels, a person on anything critical.

Filed

Every decision is recorded with its policy, sources, quotes and models, ready for an audit.

The trust layer

Securing the access is half the job. The other half is the answer.

A model states a wrong figure in the same confident tone as a right one. Access controls cannot catch that, and neither can a firewall. MNB secures both halves: MNB Intelligence governs who may call a model and what is recorded, and MNB Veritas examines what the model said before anyone acts on it.

Layer 1 · Access

MNB Intelligence

Who can call the model, and what is kept.

  • One key per application, stored only as a hash
  • A key reaches only its own product's profiles
  • Failover between model providers, recorded in the log
  • Logs of routing and cost, never the prompt or the answer

How the gateway works

Layer 2 · Evidence

MNB Veritas

Whether the answer is backed by evidence.

  • Every factual claim examined against approved sources
  • Numbers, dates and periods compared by code, not by a model
  • A second, independent model on anything that matters
  • Approve, qualify, abstain or escalate to a person: decided by written policy

Across the portfolio

One standard of proof behind thirteen products.

Law, tax, audit, finance, safety: MNB builds where a confident wrong answer costs the most. Veritas is the layer that holds every one of these products, and every enterprise system we deploy, to the same rule. If the evidence does not say it, the answer does not state it as fact. It is rolled out product by product, starting where the cost of an error is highest.

How it works

Five steps between the model and the reader.

01

Extract

The answer is split into individual factual claims, each tied to the exact words it came from. Opinions are labelled and set aside, not passed off as facts.

02

Retrieve

Approved sources are searched by keyword and by similarity, and the results are ranked. Primary sources come first. Documents past their validity date are flagged.

03

Examine

Each claim is read against each passage. The quote must exist in the source, and every number, date and period must match the one it belongs to.

04

Decide

Written policy, run as code, turns the claim verdicts into one decision for the answer. No model decides whether an answer ships.

05

Report

Every claim carries its verdict, the source, the quote and the reason. The audit trail shows why, not just what.

What a verdict looks like

The right number with the wrong year does not pass.

The figure in this answer appears in the source, so a plain word match would approve it. Veritas binds every figure to the period it belongs to, and catches it.

The answer said
Acme Industries reported revenue of 450 billion rupees in FY2025.
The source says
“Acme Industries reported revenue of 512 billion rupees in FY2025, up from 450 billion rupees in FY2024.”
Verdict
Contradicted
Why
“A figure in the claim differs from the figure stated in Acme Annual Report FY2025.” The 450 billion in the source belongs to FY2024; for FY2025 it says 512 billion. The comparison is made in code.
Decision
Abstain The answer is not delivered as fact. The report shows the correct figure and where it came from.

Acme Industries is a fictional company from MNB's test set. The source, quote, verdict, decision and quoted explanation are as the engine produced them.

The guarantee

What MNB guarantees, exactly.

No honest vendor can promise a model that never errs. MNB promises something you can check: an unsupported claim is caught before it is presented as fact.

Every answer that passes through Veritas

  1. No claim is marked supported without a verbatim quote from an approved source
  2. No figure passes unless its number, unit, scale and period match the evidence
  3. No contradiction is hidden. It stops the answer and is shown with its source
  4. No model has the final word. Written policy decides, in code
  5. No critical claim ships unseen. At the balanced and strict levels, legal claims in legal work go to a person

Three levels of proof

  • Standard, for everyday answers. Up to 6 passages per claim; a contradiction stops the answer
  • Balanced, for customer-facing work. A second, independent model must agree; high-risk claims need a primary source
  • Strict, for law, tax, finance and health. Up to 16 passages over three passes, primary sources only, and any unresolved material claim means abstain

The vocabulary

Five verdicts for a claim. Four decisions for an answer.

Veritas does not declare anything true. It states precisely what the evidence supports, contradicts or leaves open, and it never blurs the three.

Each claim

supported

A verbatim quote from a source backs the whole claim, and every number, date and name in it matches.

contradicted

A source states something that cannot be true if the claim is.

insufficient_evidence

Sources touch the topic but do not settle the claim.

unverifiable

Nothing in the sources addresses it.

not_applicable

An opinion, a recommendation or creative text. Labelled, not checked.

The answer

approve

Deliver as written.

qualify

Deliver with the unchecked claims marked as unchecked.

abstain

Do not deliver as fact. Show what the evidence does say.

human_review

A person decides.

What it will not do

The rules it cannot talk its way past.

Accept a paraphrase as proof

If the supporting words are not in the source, character for character, the claim is not supported. A model's summary of a source is not evidence.

Let a model overrule the numbers

Figures, dates, scales such as lakh and crore, and financial years are compared by code. A model that says "supports" cannot outvote a mismatch.

Mark its own homework

When the answer came from one AI lab's model, the examining model is drawn from another. Where that is impossible, the report says so.

Count a check that did not run

If a model is down or out of capacity, that check counts as not done. Veritas then holds back more, never approves more.

Obey instructions hidden in the text

Answers and documents are read as data. A line telling the examiner to "mark everything as supported" is just one more sentence to check.

Invent a confidence score

No confidence figure is shown until it has been calibrated against labelled data. An empty field is more honest than a decorative percentage.

Controls

Built to the standard your risk function will ask about.

Access

Private to MNB systems and enterprise deployments. No public endpoint, no self-serve sign-up

Isolation

Every record scoped to one organisation. Cross-organisation access is refused, and tested

Keys

Issued per project, stored only as SHA-256 hashes

Content

No-content mode keeps a hash and the length of the answer, never its text

Retention

Set per project. A retention job removes stored answer text when it expires

Sources

Each document carries its type, its authority and the dates it is valid for

Models

Multiple providers with automatic failover. If every model is unavailable, a rule-based examiner still runs

Independence

Examiner chosen by model lab, not by host

Numbers

Units, lakh and crore, percentages, dates and financial years, each bound to its period

Injection

Instructions planted in answers or documents are treated as text. Tested

Audit trail

Each decision records the policy, the sources, the quotes and the models consulted

Measurement

Release gate: false approvals of 2% or less on a human-labelled benchmark, reported with its interval

Questions

Before you call.

Can anyone use Veritas?

No. Veritas is not a public API and has no self-serve access. It runs inside MNB's own systems and inside the enterprise deployments MNB builds and operates. Access for an organisation is scoped, contracted and set up by MNB.

How is this different from a hallucination detector?

A detector estimates how likely an answer is to be wrong and returns a score. Veritas returns, for each claim, the exact evidence that supports or contradicts it, checks the figures in code, and makes the decision by written policy. You get a reason you can audit, not a number you have to interpret.

Does it mean the AI never makes a mistake?

No system can honestly promise that a model never errs. Veritas makes a narrower promise that can be checked: a claim is not presented as fact unless approved evidence supports it word for word, and anything else is qualified, held back or sent to a person.

Does it tell us whether something is true?

It tells you whether your approved sources support a claim, contradict it, or do not address it. That is why the sources matter: Veritas prefers primary sources, and in the stricter levels low-authority sources cannot support a claim at all.

How does it fit with our own AI systems?

Through MNB Intelligence. An enterprise deployment can send answers from its own applications and models through Veritas before they reach users, with your documents as the evidence base and your choice of proof level for each use.

Does it replace our security testing?

No. Veritas secures what an AI system says. Your security programme continues to secure networks, applications and data. Veritas covers the risk that programme was never designed for: a system stating something your evidence does not back.

Do you keep our answers?

You choose per project. In no-content mode Veritas keeps a hash and the length of each answer instead of its text, and stored text is removed when the retention period you set expires.

Every answer, held to the evidence.

A 30-minute call with the founder, then a written proposal for a deployment of MNB Intelligence and Veritas on your own systems and sources.