When you assess someone, all you ever see is the result.
AIdeaText makes visible how the argument and the spoken defence were built.
Modelling how people think, not just what they produce.
We analyse written and spoken argumentation, and show how it matures from one draft to the next. We do not measure reasoning in full—that is larger and more complex: what we measure is the trace it leaves in the text.
Anyone who has to assess another person—a teacher, an investor, a hiring manager—has nothing to go on but finished work: deliverables that may be faked, that may be flawed, that may have been shaped by circumstances no one can see.
With generative AI, a correct result no longer proves that anything was understood.
And institutions apply a single evaluative logic to disciplines that demand quite different kinds of reasoning: deductive in engineering, inductive in the health sciences, abductive in the social sciences, analogical in the humanities.
The kind of reasoning is not a label next to the graph: it is the shape of the graph.
Four reference texts, each written with one kind of reasoning, run through the same pipeline the application uses. The detector gets all four right. This is a demonstration that the shapes can be told apart—validation at scale is part of what we want to do with you.
Deductive · star
A star. One concept holds everything else up—here it is “aprendizaje”, learning—and the rest hangs off it. The rings are not decoration: the inner one sits one step from the core, the outer one two.
Inductive · funnel
A funnel. At the bottom, the individual cases, each with few connections; at the top, the handful of ideas that gather them up. Height is the number of connections, and it is labelled down the left.
Abductive · bridges
Rings. At the centre, the concepts the explanation passes through: the bridges. This is a text that jumps from one region to another to explain something that did not fit.
Analogical · strands developed side by side
Lobes. Strands the text develops in parallel, each with its own place. And look at the pink lines crossing from one to the other: that is the analogy.
On a small screen what you can tell apart is the shape, which is what the detector measures; the concept names become legible when you enlarge the image.
Dual-process theories of cognition · multimodal discourse analysis · computational knowledge representation through semantic networks.
The map does not rearrange itself.
There is no force-directed solver. How far a concept sits from the centre is a measurement of that concept—steps to the core, connections, betweenness—and its angle follows from its name.
A concept moves only when what it measures changes. On the test corpus, with the graph left untouched, all 299 concepts keep their exact coordinates.
That is why comparing draft 3 against draft 4 means something. The movement is the diagnosis.
Four drafts of one real text, truncated to the length it would have in each week of the term: the sequence is simulated. The measurements are not—density, depth and the position of every concept are computed by the same code the application runs.
What you get, twice over
We apply for the grant with you.
The dossier is already written: a comparative IRB framework, a cross-site validity argument, a data management plan, compute sizing, a cost model and an institutional ownership model. What is missing is your institution, not the paperwork.
And you deploy it before anyone else.
Assessment of written and spoken reasoning through semantic graphs: nobody else in Latin America has it running.
What it does not do
It will not tell you whether the reasoning is sound. It shows you how the text was built: so that the student can see what they actually did, and the teacher has something concrete to ask about.
Where it is being applied
Same infrastructure, same architecture, same analytical design. What changes is how it is applied.
Law
The problem it addresses
Assessing legal reasoning, case-law analysis and legal drafting beyond the memorisation of statutes
Target client
Law firms, law schools, judicial training institutes
Education
The problem it addresses
Assessing critical thinking and the framing of ideas in essays and long-form exams, at secondary and university level
Public policy evaluation, civil service examinations, document processing
Target client
Government ministries, state governments, schools of public administration, international organisations
See how it works
What AIdeaText has achieved
Incubated at the IFE Living Lab, Tecnológico de Monterrey
NVIDIA Inception
Tools Competition—second round
Claude Startup Program—first tier
Pilot at a Peruvian university
Who runs it
Manuel Vargas Alegría. Former National Administrator of SIAGIE (2013-2015), the school information system of the Ministry of Education of Peru: 7 million students and 350,000 users. More than a decade as a specialist and consultant for that same ministry. Doctoral candidate; his thesis is on the binding cognitive agent.
Three ways to deploy it
The analysis is the same in all three. What changes is where the data lives and who runs the platform.
On our servers
We run the platform and your institution uses it. It is the option with the least work for your IT department: there is nothing to install and nothing to maintain. And it is the only one of the three for which the storage architecture is already written, object type by object type and with a retention policy for each.
Who it fits
An institution that wants to start soon and would rather not stand up infrastructure just to try it out.
On your institution’s servers
The platform is deployed inside the infrastructure your institution already runs, and your own team operates it.
Who it fits
An institution with its own IT department and a written policy on where its students’ data has to live.
Hybrid
Part of it lives at your institution and part of it we run. The split is decided case by case, not in one block.
Who it fits
An institution that needs to separate what cannot leave the building from what can, and would rather not give a single answer for everything.
What has to be decided, in all three
Where your students’ data lives.
Who runs the platform day to day.
Who updates it, and on what schedule.
What does not change is the data schema. Between your institution and the platform sits an adapter layer that translates its academic calendar, its grading scale and its virtual classroom, without touching the core schema.
The institutional relationship is a separate decision: client, research partner or co-owner are three levels of involvement in the project, with different rights over the data, over publication and over the governance of the platform.
There are no figures here, and that is not caution: the technical dossier is written for each institution, because one institution’s does not fit the next. What can be said in advance is this; the rest is the first meeting.