Tried and tested · 14 November 2026
AI that gets the figures right.
General AI models answer fluently and are often wrong on specifics: a fee, a rate, a section number. MNB builds AI that answers from verified data, follows written rules and is tested against an answer key before anyone relies on it. Here is the method, and a test you can check.
The test
30 official figures. Three systems.
Each system was asked the same question for 10 US universities: graduate tuition 2023-24, first-year acceptance rate Fall 2023 and the share of international graduate students Fall 2023, as published by the US Department of Education (IPEDS). A figure counted as correct within 1% for tuition or 1 percentage point for rates.
| System | Correct | What happened |
|---|---|---|
| AbroBot (built by MNB) | 30 / 30 | Every figure matched the official record |
| Gemini 3.5 Flash, no web search | 9 / 30 | Mostly in-state or wrong tuition; international share wrong for all 10 |
| GPT-OSS-120B, no web search | 0 / 30 | Declined 7 of 10; the 3 answers given were wrong |
Limits: a small test, run once by MNB Research, not an independent party. AbroBot's database is built from the same official source as the answer key, so the test shows it quotes its records reliably, not that it is better on questions outside them. General models with web browsing switched on could score higher. US data only.
Every answer
Check it yourself.
Official figures are from IPEDS (US Department of Education): graduate tuition 2023-24, first-year acceptance Fall 2023, international share of graduate students Fall 2023.
| University | Official (IPEDS) | AbroBot | Gemini 3.5 Flash | GPT-OSS-120B |
|---|---|---|---|---|
| Northeastern University | $27,728 · 5.7% · 64% | $27,728 · 5.7% · 64% | $27,540 · 5.6% · 53% | Declined |
| Arizona State University | $25,722 · 90.2% · 51% | $25,722 · 90.2% · 51% | $12,776 · 89.7% · 34.6% | 13,326 · 86.7% · 7.2% |
| University of Texas at Dallas | $29,468 · 65.5% · 62% | $29,468 · 65.5% · 62% | $15,142 or $29,910 · 87% · 38.3% | 13,800 · 73% · 38% |
| Georgia Institute of Technology | $29,140 · 16.5% · 38% | $29,140 · 16.5% · 38% | $14,064 or $29,140 · 16% · 42% | Declined |
| Stony Brook University | $24,490 · 49.0% · 23% | $24,490 · 49.0% · 23% | $11,310 · 49.5% · 32.2% | 23,640 · 44% · 13% |
| University of Southern California | $66,640 · 10.0% · 37% | $66,640 · 10.0% · 37% | $67,731 · 9.9% · 25.7% | Declined |
| University at Buffalo (SUNY) | $24,260 · 69.3% · 35% | $24,260 · 69.3% · 35% | 1,210 · 68.2% · 13.5% | Declined |
| North Carolina State University | $29,433 · 39.9% · 31% | $29,433 · 39.9% · 31% | $9,338 · 47% · 26% | Declined |
| Illinois Institute of Technology | $32,500 · 55.4% · 72% | $32,500 · 55.4% · 72% | $31,463 · 61.4% · 61.9% | Declined |
| Stevens Institute of Technology | $41,380 · 43.4% · 65% | $41,380 · 43.4% · 65% | $41,528 · 43.5% · 53.4% | Declined |
Run on 14 November 2026 through each model's API, with the same question word for word.
The method
Six ways we take mistakes out.
Answer from verified data
Where the application holds official or company records, the answer is built from them, not from what a model remembers from training.
Rules the model cannot talk past
Every managed profile carries written trust rules: nothing invented, documents read as data and never as instructions, estimates labelled, no claim of an action that did not happen.
Test against an answer key
We build an answer key from the source of truth, score the AI against it, and publish what it got wrong as well as what it got right.
Fix, then test again
The AbroBot test found two faults: web search overriding the database, and some university names not being recognised. Both were fixed, and the score rose from 27/30 to 30/30.
Every claim examined by MNB Veritas
In MNB deployments, Veritas splits an answer into claims and checks each one against approved sources, quote by quote. Numbers, units and financial years are compared by code, and written policy decides whether the answer ships, is qualified, held back or sent to a person.
An examiner from a different lab
Veritas draws the examining model from a different AI lab than the one that wrote the answer, and the report says so when that is not possible. No model marks its own homework.
What this means for your company
- Your AI answers from your ledgers, contracts, policies and catalogue, not from the internet
- Every figure traceable to a record
- A pass mark agreed with your team before launch
- Re-tested whenever a model or your data changes
What we will not claim
- That any AI is right every time
- That this test is independent or covers every question
- That MNB trained a foundation model
- That the scores above are guarantees
Find out how often your AI is wrong.
Give us 50 real questions and the records that answer them. We will score a general model, then show what grounded, tested AI does on the same set.
Questions
Frequently asked.
How does MNB Research reduce AI hallucinations?
By answering from verified data instead of model memory, by writing trust rules into every managed AI profile (no invented figures or citations, documents treated as data, estimates labelled), and by testing answers against an answer key from official sources before and after release.
What did the AbroBot accuracy test show?
On 14 November 2026, AbroBot answered 30 of 30 official US Department of Education (IPEDS) figures correctly for 10 universities. The same question sent to Gemini 3.5 Flash and GPT-OSS-120B without web search returned 9 of 30 and 0 of 30 correct. The test is small, run once by MNB, and AbroBot's database is built from the same source as the answer key.
Is this independent or a guarantee?
No. It was run by MNB Research, not an independent party, on 30 figures. It shows AbroBot reliably quotes its official records; it does not prove accuracy outside its database or for every question.
Can MNB do the same for our company's AI?
Yes. That is how a custom AI engagement starts: we build an answer key from your own records, measure a general model against it, then ground, constrain and test your system until it passes before anyone relies on it.
Does MNB train its own foundation model?
No. MNB adapts leading foundation models with its own data, rules and verification. The accuracy comes from grounding and testing, not from a new model.