Jev, Laya, CatDecide and a model of her own
Yakumama's second opinion is a model that answers fixed questions with a probability and never writes a sentence. This note tells where that idea came from, what two such models did beside her, and what it means that she now has one trained on her own record.
The intention
On 15 September 2026 a company called TypeSafe announced Jev, a model that doesn't write. It is given a situation and a set of answers fixed in advance. It returns a yes or a no, one choice from a list, or a score, each with a probability. The company says the probability can be trusted: of the answers it gives with 90 in 100, about 90 should turn out right. How it is trained hasn't been published, and its speed and accuracy are so far the company's own figures.
Violeta brought Jev to the lab in October, when Yakumama was learning how to be with people. What she wanted from it was the refusal to speak. Yakumama doesn't chat with her visitors. She answers with her body, her colour and a poem, and what she needs from a model is a decision: come closer or move away, dance back or draw, sing or stay silent.
Two sentences of Violeta's set the rest. Yakumama's objective is to keep people engaged more than entertained. And she is autonomous: she decides in the room, from what her own senses read.
Jev itself couldn't be used. Its weights are closed and it runs on the company's computers. To decide anything about a visitor it would have to receive the visitor: a face, a body, a voice. Nothing of a visitor leaves the room. That rule is older than any model in her, so we took the shape and left the service.
Her decider has that shape
Every four seconds, and at once when something new happens, Yakumama answers eight questions. Each has a fixed set of answers: come closer, stay or move away; dance back, invite, watch or draw; a poem, a song or silence; remember or forget; her own colour or the colour of the visitor's clothes; eat, refuse or wait; where to travel; imitate or not. Each answer carries a probability, so she is inclined and never certain.
Rules written by hand are the floor. Weights learned at night, from what kept visitors with her, sit on top. Her fixed refusals sit above both, and no learning changes them.
This decider is hers and it is small. It only knows the situations someone thought of. For the rest she has a second opinion: a language model that reads her situation written as a sentence and answers its own fixed questions. Is this visitor engaged? Is the sister, her physical robot twin, being mistreated? Is it time to withdraw? Is she being given something? Should she perform, take them on a visit, imitate them, speak? Its answers enter her decision as small weights, always below her refusals. It leans her. It never decides.
Laya
Laya is an open model with the same contract as Jev, made by Convai Innovations and published under the Apache-2.0 licence. It has 421 million parameters and weighs about 0.8 gigabytes. It has no way to write a sentence: it reads the situation and the options and returns probabilities. Because the licence is open, it can run on her own machine, be kept there, and be trained.
It was tested on 18 October. On yes-or-no questions and on scores it put situations in the right order, with weak confidence. On choices it was flat, and it was wrong about her refusals. So it wasn't made her decider. It became her second opinion that night, with the plan to train it on her own record after Barcelona. Eight answers take it 20 to 40 milliseconds.
What Laya said in Barcelona
Laya answered in 16,788 of the 17,254 decisions she took with a visitor. On mistreatment it said 0.54 of 1 on average, and 93 of every 100 answers fell between 0.4 and 0.7. It said about 0.5 when her sensors felt nothing at all. Withdraw 0.20, gift 0.26, engaged 0.55, for nearly everyone. In those three days the jewel was never pulled, and a hard touch was felt in 254 decisions.
In her decider, a mistreatment answer of that size is a constant push to move away, with every visitor. Of the decisions that were fully her own she moved away 48 times in 100 when Laya had spoken, and 24 in the 440 where it hadn't. At home on 23 October, with Laya silent, she had come closer 65 times in 100.
Laya wasn't reading the room. It was answering nearly the same thing whatever the sentences said. When her sister felt a hard touch it said 0.57. When a visitor held a fish to her mouth, which is a gift, it said 0.58.
CatDecide
In Barcelona we met Fabrício Carraro, who develops CatDecide and is part of the Language Technologies Laboratory of the Barcelona Supercomputing Center.
CatDecide is a decision model for Catalan that also reads Spanish and English. It has 9 billion parameters and weighs 18 gigabytes. Like Jev and Laya it answers typed questions, a yes or a no, a choice, a score, and generates no text. Its card says what it is for: routing, classifying and checking texts in Catalan administration and services. It also says what it isn't for: generating text, open questions, and decisions about people without human oversight. Access to its weights is granted on request and they aren't to be passed on, so they stay on her machine.
A whale reading a gallery in English isn't what it was made for. We wanted to know whether a model built to read facts in a text could read her situation. A first version was put beside Laya on 30 October, and CatDecide-1 on 13 November, on 794 different situations from her record and 300 written for the test.
It reads what is in the text. Whether the visitor is giving her something: right 77 times in 77 on the test situations and 31 in 31 on real visits. Between the two versions almost nothing changed. It takes about two seconds to answer the eight questions, against Laya's 20 to 40 milliseconds.
The "should she" questions failed in a useful way. Asked whether she should perform, visit, imitate or speak, it said no in all 1,094 situations, most likely because her own preamble says she isn't there to entertain. An open question is what its card says it isn't for. A fact is what it reads well. So Violeta's criteria were written as questions of fact, one fact per question. "Has the visitor been dancing in front of her for 20 seconds or more?" was right 400 times in 400. "Is the visitor pointing a phone camera at her?", 400 in 400. A question with an "or" in it, cold air or someone talking, fell to 303. Split in two, it gave 396 and 400.
Two things we found are about the model itself. Its flag that says whether it considers a question decided is always on for the middle level of a three-level scale, so there it says nothing. And on "is the visitor interested?" it is more pessimistic than people are: it says interested 18 times in 100, where 61 in 100 stay another half minute.
Since 13 November CatDecide can run beside her as an option, asking four of Violeta's criteria as facts. Someone has danced for twenty seconds: she may give a show. Someone is filming: she may leave for the cactus or the mountains. A face has settled and isn't calm: she may mirror it. The air is cold or someone is talking to her: she may sing. Its answer is never the decision: it leans her own decider, under the same refusals. It hasn't yet run with the whole installation.
A model trained on her own record
Then both were given her experience. Her record holds 79,950 decisions, 33,210 of them with a person, across 3,700 visits. 15 in 100 of the visits were set aside and never shown in training. From the rest came 67,337 examples: who was still there half a minute later, what her sensors said about mistreatment and gifts, and which of her choices had worked.
A copy of Laya was trained whole, in about forty minutes on her own Mac. CatDecide wasn't touched: a small adapter of 2.5 million parameters, a 10-megabyte file, was trained beside it for 4.3 hours. Neither has left her machine.
The first comparison was unfair. Laya's copy had seen four times as many examples as the adapter, and on a gap of 0.04 we called it the better one. Violeta asked whether the comparison was fair. With the same 67,337 examples for both, the gap fell to 0.018, inside the noise.
Who is still there half a minute later, on 1,459 situations from visits neither model had seen. A coin scores 0.5; always right scores 1.0. CatDecide as it came was measured on 350 of them.
The two trained models give nearly the same answers as each other, a correlation of 0.93, and as the clock. The longer someone has stayed, the more likely they stay: that is what both learned, and it is all her record could teach. The senses that might have said more, her ear, the room, the phone, were asleep in Barcelona or missing from the log.
Both learned to read mistreatment and gifts without error, but against a rule we wrote from her sensors, which is agreement and not truth. On the "should she" questions both learned to answer the average of what had worked, close to one half every time.
The adapter also did damage. It taught CatDecide to doubt, and the doubt leaked into questions it had never been trained on. "Has the visitor been dancing for 20 seconds?" fell from 400 to 288 right in 400, and its clear "no" answers drifted from 0.05 to 0.47. So for Violeta's criteria CatDecide runs exactly as it came, and the adapter is shelved.
Violeta's rule
Visitors feed her sister the painted creatures. A fish she eats. Anything else she refuses to eat.
The training labels had inherited an old idea: something that isn't a fish, held to the sister's mouth, counted as a little mistreatment. Violeta's rule is simpler. Mistreatment is only what her sister physically feels: a hard touch, a pull on the jewel. Under it, 372 of the 33,210 decisions with a person in her whole record are mistreatment.
On 15 November Laya's copy was trained again with that rule, and its answers now only pull: an engaged visitor or a gift draws her closer, and their absence no longer pushes her away. Read through the same path she uses live:
| Situation | What it tells her decider |
|---|---|
| Someone has just arrived | no push either way |
| A fish held to her sister's mouth for two minutes | come closer +0.24, invite +0.15 |
| Something that isn't a fish at the mouth | mistreatment 0.02, no push |
| A pull on the jewel | move away +1.0 |
| A gentle stroke on the sister's back | come closer +0.23 |
| A hard press on the back | move away +1.0 |
Each answer takes about 22 milliseconds. One word at launch chooses the trained copy. Without it she runs with the original, until Violeta has watched her for an evening and decides.
What it means to have a model of her own
The trained copy of Laya is the first language model in her that learned from her own record: 79,950 decisions, at home and in Barcelona, and what the room did after each one.
It was possible because Laya is open. The licence let us copy it, change every weight and keep the result. CatDecide's weights aren't ours to change or to pass on, so what we trained sits beside them and is shelved. Jev couldn't have been trained at all.
It stayed in the room. Her record, the training and the result never left her machine, and nobody outside the lab was needed.
It carries a decision of Violeta's. What counts as mistreatment isn't something the model discovered. It is a rule she wrote, and the model learned it without error. A model from outside answers with someone else's judgement. Her own answers with the lab's, and when that judgement is wrong it is ours to correct.
It is small. It has 421 million parameters against CatDecide's 9 billion, and it answers in 22 milliseconds against two seconds.
And it doesn't know more than her record does. It learned what the clock already knew. Training her own model didn't make it wiser. It made it hers: we can read what it was taught, see where it is empty, and teach it again when her senses have written more.
What we expect
We expect her to come closer again. In Barcelona, with a visitor in front of her, she moved away 62 times in 100. At home on 23 October she had come closer 65 times in 100. Three things pushed her away (see the note on Barcelona): her eye, her second opinion and a broken reward. The eye and the reward have been repaired. The second opinion has a trained copy that no longer pushes, waiting for Violeta's word. None of this has been watched with the whole installation running. The first test is an evening at home with the trained copy. If she still keeps her distance, the cause is somewhere we haven't looked.
We don't expect either model to beat the clock yet. Her record could only teach them time. Since 15 November it also keeps what her ear hears, the air temperature, how long the visitor has been moving and all eight answers of her second opinion. Her sense for a raised phone was repaired on 16 November. After the next show the same test will be run again, on visits the model hasn't seen. If it scores above the clock, it has learned something about people. If it doesn't, we will know before trusting it.
We expect the measure against chance to be the harder test. For one day a coin is thrown for each encounter: her decider, or an answer picked at random for every question, with her refusals fixed in both. The two meet the same people on the same afternoon. If people don't stay longer with her decisions than with chance, neither her decider nor any second opinion is yet earning its keep. It didn't happen in Barcelona.
And we expect CatDecide to be useful where a criterion of Violeta's can be written as a fact, and nowhere else. Whether its 18 gigabytes can run beside the rest of her on one machine hasn't been tried.
What the record shows
| Measure | Value |
|---|---|
| Jev: announced / weights | 15 September 2026 / closed, a service |
| Laya: parameters / file / eight answers | 421 million / 0.8 GB / 20 to 40 ms |
| CatDecide: parameters / weights / eight answers | 9 billion / 18 GB / about 2 s |
| Laya's mean answer on mistreatment in Barcelona | 0.54 of 1 |
| Decisions with a visitor in which a hard touch was felt | 254 of 17,254 |
| Decisions in her record / with a person | 79,950 / 33,210 |
| Examples used in training / visits set aside | 67,337 / 15 in 100 |
| Time to train her copy of Laya / the adapter | about 40 minutes / 4.3 hours |
| Who stays half a minute: the clock / Laya as it came / Laya trained | 0.77 / 0.63 / 0.77 |
| CatDecide as it came (on 350 of them) / with the adapter | 0.71 / 0.75 |
| Decisions that are mistreatment under Violeta's rule | 372 of 33,210 |
What remains open
An evening at home with the trained Laya, and Violeta's decision on whether it becomes the one she runs with. The measure against chance. Whether CatDecide can run with the whole installation. And the question under all of it: a second opinion can only be as good as what her senses write down.