ungli, layers of NSMs (Narrow Small Models)
13 November 2026
From Koher's notes, where it was first published on 20 October 2026 and renamed on 13 November 2026.
Renamed on 13 November 2026. This method was published as Synapse on 20 October 2026. It is now ungli, after the Hindi ungli rakhna, to put a finger on the exact spot. That is all a model in it does. A model here is called narrow because its answer is small, whatever its own size, so the models it uses are now called NSMs (Narrow Small Models). It has a builder at ungli.syntax.prayasabhinav.net, and the kind of model it builds is set out as the NSM syntax.
ungli is a method by Prayas Abhinav for using AI in which no model writes a word. It came out on 20 October 2026, with code to copy. A job is broken into single steps, and each step that needs a model goes to an NSM (Narrow Small Model), which can answer only with a number or with nothing. Code checks every answer. A case a model cannot settle goes to a person, or to rules people wrote in advance. The words a person reads are written by people.
The method
The method page sets out why Koher and Split-Domain Cognition use small models, the steps, what small means and where the method does not fit. One problem stays open: the readers are built on text taken without its authors being asked.
Two worked examples
rai checks whether a description is complete. For each question on a form, its reader points at the phrase in the answer that answers it, or at nothing, and code says complete or not complete. RAM points at who will do each act in a piece of office writing, and code says <doer> karega or RAM bharose. Both readers have 33 million parameters and run on an ordinary computer's CPU.
The code
Four small files in plain Python. One lets a model's answer through only as the number of an item on a list. The same file holds the router, a table people write that says which layers handle each kind of case. One points into a text with the base that rai's and RAM's readers were trained from. One is semantic search: when a text is too long to read whole, it finds the passage closest in meaning to what a person asked about. One holds a model that scores labels to a fixed list. The code comes with two demonstrations, which run on an ordinary computer: twenty office notes, each marked by hand with who will do the work, and the same notes made long. These models are small enough to run in a web browser if needed.
Licence and source
The method is under CC BY 4.0 and the code under AGPL-3.0, at github.com/koherarchitecture/ungli. To cite it, use the repository's citation file. ungli is listed first on the tools page.