The record
Deployed infrastructure, not a proposal.
MAIA BRAIN is in use at several multinational corporations in the oil and gas, finance and technology sectors, carrying everyday business workloads across narrow and wide domains as one general-purpose system (MAIA BRAIN technical brief, Section I). We do not name customers, by choice, for client privacy. The same architecture, with no retraining, reads a legal document estate in one deployment and repairs programs in another. Against frontier neural models it wins by multiples, not margins.
One measured deployment
A legal document estate of 1,620 documents, read end to end.
Its provenance was established rather than assumed. Every object was reconciled across the index, the store and a checksum pass before a single question was asked.
-
1,620 of 1,620
checksums matched
No object without a record, no record without an object, no size mismatch, no empty file
-
278,768
facts admitted
Each one carrying where it came from
-
1 ms
median response
Across 61 timed capabilities. 52 of the 61 returned in under 1.5 seconds
MAIA BRAIN technical brief, Section I.
-
What that is worth, in business terms.
A document estate that size is normally a reading project with a date on it, staffed by people whose time is the most expensive in the building, and the answer is only as good as which parts they got to. Here the whole thing is answerable, in milliseconds, with the page behind every answer. The fortnight of somebody emailing four departments does not happen.
-
And over code.
The same construction has been carried end to end over a single repository of 10.4 million source lines of mixed C++, Python and Java, from which the analyser admitted 2.31 million schemes, one per procedure, at roughly 41 thousand lines per second in a single pass. Spine Is All You Need.
Five ways of doing the same work
Where the separation actually is.
Task success on work whose difficulty is keeping track of state and procedure rather than recognising something. Higher is better.
| What it tests | A transcript | Retrieval | A one-million-token context | A structured store | MAIA BRAIN |
|---|---|---|---|---|---|
| Planning over changing state | 4.3 | 5.1 | 4.8 | 31.6 | 98.4 |
| Hard constraint reasoning | 42.5 | 43.9 | 44.7 | 61.2 | 98.6 |
| Tracking a value through a long document | 99.6 | 99.4 | 99.7 | 99.5 | 99.8 |
| Holding state across a long task | 20.5 | 24.8 | 22.1 | 36.3 | 98.5 |
| Multi-step work with tools | 34.0 | 36.2 | 35.4 | 52.7 | 98.2 |
98.2 to 99.8, against as low as 4.3. That tells you precisely which of your problems this is for, which is a more useful thing to know than a single headline number.
Success against cost
2.5 times the task success, at around a thirtieth of the cost.
The same work, done five ways. The first row is a frontier model carrying the task alone, which is what most of the market is selling. Cost is relative to that row.
| Configuration | Relative cost | Task success |
|---|---|---|
| A frontier model alone, with no store | 1.000 | 40.2 |
| MAIA, with frontier models at both edges | 1.040 | 98.7 |
| MAIA, with an open 8B model fine-tuned for the job | 0.031 | 98.5 |
| MAIA, with an open 3B model fine-tuned for the job | 0.012 | 96.8 |
| MAIA, with that same 8B model off the shelf | 0.031 | 71.4 |
-
Read the third row against the first.
2.5 times the success rate, at around a thirtieth of the inference cost and of the compute and energy behind it. That is what stops a deployment being scoped down to a sample. At frontier rates the economics only work on a narrow slice. That is why so many pilots succeed and so few scale.
-
And read the last row too.
The same open model, off the shelf rather than fitted to the job, loses twenty-seven points. The scarce thing is not the size of the model. It is a model trained for the work it is being asked to do. Which is also why bespoke here means something: the fit is the result, not the packaging.
Why it holds as the question gets harder
There is no middle of the context to get lost in.
A fact admitted on arrival has an address rather than a position. Retrieving it is a lookup whose cost does not depend on how much other material surrounds it, because the store has no middle.
So eight facts gathered at once is eight lookups. The flatness is the point. The height alone would be unremarkable.
Plot answers correct against how many references a question has to follow, and by the fourth step the separation is an order of magnitude. By the fifth it exceeds one.
Those are exactly the questions a business cannot currently answer: the ones where the evidence sits in four systems and each answer depends on the last.
The two numbers people ask for
99% accuracy. 14 days to live.
-
99%
The commercial figure, and it is a rounding of measurements rather than a claim they do not support: 98.7 with frontier models at both edges, 98.5 with a fine-tuned open 8B model, against 40.2 for a frontier model working alone. Accuracy is how much of a job gets completed. Determinism is separate, and it is the thing that makes the answer checkable: the same inputs and the same rules reach the same conclusion, and the conclusion shows its work.
-
14 days
From decision to live, in three phases, with read-only credentials at database level throughout the initial stages and a full audit trail from day one. Foundation and connectivity, days 1 to 4. Validation and testing, days 5 to 8, where your own experts check the reasoning. Production from day 9. It is 14 days rather than a research timeline because fitting MAIA to a new organisation needs no retraining. That is the fact everything else rests on.
Where the figures come from
Two papers, and you can read both.
-
Spine Is All You Need for Going Further
The full account and the experimental specification. The source of the benchmark and cost tables above.
-
MAIA Brain: Language Models at the Edges of a Governed Symbolic Core
The companion brief. The deployment record, the architecture, and the cost and edge-model table.
Ask for either by name and we will send it. A claim you can check is the only kind worth making, and it is the thing we are ultimately selling.
The next step
Make us run it on your own data.
The fastest way to test a record like this is to hand us a question you cannot currently answer and watch. Bring one that spans four systems, and we will show you what MAIA reads, what it concludes and under what authority it acted.