Syndicate

Inspiration

Highly accurate AI models are expensive, and the people who rely on them the most run the same kinds of questions over and over. Every query to a frontier model adds up in token fees, and sensitive documents have to leave your machine. We wondered if a group of small models could do what one big model does, and whether they could catch each other’s mistakes the way people in a good debate do.

What it does

Synthesis is a local AI system that gathers several small models into a council-style debate. Given a prompt, each model thinks it through, then works to convince or disprove the others’ theories, which sharpens the accuracy of models like Gemma. The council settles on one answer once 80% of its members agree.

To add even more accuracy, Synthesis includes a lie detector that studies each model’s weights and neuron activations to tell whether it is lying on a question with an absolute right or wrong answer. When it catches a lie, it shows the council the data along with a confidence percentage, and the other models question the drifting model until it finds its way back on track.

The council works only from your own files, so contracts, filings, and compliance material never leave the machine. A council.md file learns from the council’s mistakes and your preferences, and a lesson is saved only after the council reasons about it and decides it’s worth keeping.

How we built it

Syndicate runs entirely on an ASUS Ascent GX10, whose large unified memory lets several models stay loaded at once. We serve the models locally with llama.cpp and built a routing layer on top that sets up each debate with 2 to 6 seats, runs the rounds, tracks agreement, and enforces a time limit. For the lie detector, we read the models’ internal activations during a question and compare how they look when the model is telling the truth versus lying. The interface is a normal chat box, with a Table View that draws the council around a half-circle table and a Chat View that shows what the models say to each other. It has light and dark themes and no login, since it runs on your own machine.

Challenges we ran into

Syndicate runs entirely on an ASUS Ascent GX10, whose large unified memory lets several models stay loaded at once. We serve the models locally with llama.cpp and built a routing layer on top that sets up each debate with 3 to 7 seats, runs the rounds, tracks agreement, and enforces a time limit. For the lie detector, we read the models’ internal activations during a question and compare how they look when the model is telling the truth versus lying. The interface is a normal chat box, with a Table View that draws the council around a half-circle table and a Chat View that shows what the models say to each other. It has light and dark themes and no login, since it runs on your own machine.

Getting small models to challenge each other instead of agreeing too quickly or repeating the same wrong answer. Keeping debates short enough to finish in a few minutes while still giving the models room to change their minds. Deciding when the council is “done,” which led us to the 80% agreement rule with a time limit. Making the lie detector’s output simple enough to be useful, which is why it reports a single confidence percentage. Stopping council.md from filling up with noise by making the council vet every lesson first

Accomplishments that we're proud of

We built a fully local system where small models work together and check each other, with no per-token fees and no data leaving the machine. We’re also proud of connecting the lie detector to the debate, so the council gets a confidence signal as evidence instead of just trusting whichever model sounds the most sure. And the Table View makes a normally hidden process easy to follow, so you can watch the council argue.

What we learned

Disagreement is a useful signal. Models are better at spotting each other’s mistakes than their own, and a confidence number gives the council something concrete to push back on. We also learned that smaller models can do a lot more together than alone when the structure around them is right, and that how you design the debate matters as much as which models you use.

What's next for Syndicate

We want to test the lie detector against many more known-answer questions to see how reliable it is, support more models and bigger councils, and speed up the debates. We also plan to expand the document research side for privacy-sensitive work like legal, financial, and compliance review, so teams can get dependable answers without frontier-model costs.