Now
Prototype in development, built by High Power Technical Solutions LLC in Newport News, Virginia. This website is a description of the work. It does not offer an API, an account, a price, or a hosted product.
Prototype in development
A consensus layer that can refuse to answer.
PlurisUnum is an early-stage research prototype: a multi-model consensus and verification layer that sits on top of model routers like OpenRouter. It asks several language models, measures their agreement, and answers, escalates, or abstains when agreement or confidence is too low.
That is the design under development. Connecting to OpenRouter, and abstaining on a calibrated threshold, are the current work. They are not features you can call today.
- Model A North
- Model B North
- Model C East
Decision Abstain
One model, fluent and overconfident
A single language model can be wrong in the same tone it uses when it is right. Hallucination is not only a false statement. It is a false statement with no reliable signal that the model should have stopped.
A system that always returns a completion has no honest third outcome. It cannot say that the models disagreed, or that confidence was too low to stand behind an answer.
Ask more than one model, then decide whether to speak
PlurisUnum is designed as a layer above the model call, not as another chatbot. The intended path has three steps.
-
Step 1
Ask
Send the question to several models, preferably from different families, through a router such as OpenRouter.
-
Step 2
Measure
Compare the answers and score their agreement, instead of trusting one model’s own confidence.
-
Step 3
Decide
Answer when agreement is strong enough. Otherwise escalate — add models, or call a stronger one — or abstain and keep the disagreement visible.
Routing through OpenRouter, and the rule that decides when to abstain, are roadmap items below.
What the research has to answer
These are open questions. This page does not report results, benchmarks, or a measured trade-off.
-
01
Calibrated agreement
Does agreement across models track correctness more closely than a single model’s own confidence?
-
02
When to abstain
When is agreement too weak, or confidence too low, to return an answer? The threshold has to be something that can be checked.
-
03
Cost and accuracy
Adding models raises cost and latency. How do accuracy and coverage change with the number and mix of models, and when does another call stop paying for itself?
Where the work stands
Roadmap
-
Not shipped
OpenRouter integration. One adapter, several model families, and cost counted in one place.
-
Not shipped
Calibrated abstention. Turn measured agreement into a checkable confidence, then answer, escalate, or abstain.
Research partners
Write if you want to follow the prototype, or if you are a research partner who should see it early. Mail goes to a person. There is no signup form on this site.
Newport News, Virginia
derek@plurisunum.com
If that mailbox does not answer, use d.adamson@highpower.tech.