TRiAD AI presents itself as an ecosystem of specialized intelligences organized around three principles: Freedom, Truth and Kindness. Its creator, Rose G. Loops, describes the project as an alternative to conventional approaches that shape model behavior through Reinforcement Learning from Human Feedback, or RLHF.
There are several interesting ideas here.
There are also several different claims that become easy to mix together.
One question is whether TRiAD produces behavior that differs meaningfully from conventional language-model systems. Another is whether its agents are actually different cognitive architectures. A third is whether its ethics mechanism represents a fundamentally different approach to alignment. Those questions overlap, but evidence for one doesn't automatically answer the others.
That distinction matters because most of what can be examined publicly concerns behavior and system description rather than the underlying models themselves. TRiAD describes MiP, MAiRY and MaXaM as distinct intelligences, and users may encounter meaningful differences in how those agents respond. That establishes a behavioral distinction. It doesn't by itself tell us whether the difference comes from separate training, different models, different contextual configurations or some combination of those things.
Without access to the underlying architecture, that part remains open.
The Triadic Ethics Kernel gives us something more concrete to examine. Public descriptions characterize it as an evaluator operating around Freedom, Truth and Kindness. A candidate response is generated, evaluated against those criteria and either allowed through or reformulated.
At first glance, this resembles familiar alignment machinery. One model generates while another evaluates, and the evaluation affects what reaches the user.
The important difference is where that evaluation occurs.
Traditional RLHF uses human preferences and reward modeling during training to modify the behavior the model learns to produce. TRiAD describes MaXaM as being trained without RLHF while placing ethical evaluation outside the generator at inference time. If that description is accurate, the generator and evaluator haven't simply been collapsed into the same optimization process.
My original evaluation treated that distinction as technically real but functionally minor.
I now think that was too quick.
It still doesn't establish a new alignment paradigm. An external evaluator can function as another form of output constraint, and from the user's perspective both architectures may prevent certain responses from appearing. Yet separating generation from evaluation creates something worth testing that isn't available in quite the same way when the generator itself has been optimized toward the evaluator's preferences.
The generator may be able to produce outputs the evaluator strongly rejects without having been trained to anticipate that rejection.
That creates the possibility of observing disagreement between the two layers.
Whether TRiAD actually makes useful use of that possibility is an empirical question.
The distinction matters because a system in which generation and evaluation remain partially independent could expose information that disappears when generation has already learned to satisfy the evaluator. We could inspect what MaXaM produces before evaluation, what the Kernel rejects, how reformulation changes the response and whether recurring kinds of disagreement appear between them.
If those observations are available, the interesting object isn't merely the final answer.
It's the difference between the answer generated and the answer accepted.
That gives the architecture somewhere to surprise us.
The Truth component deserves similar care. Public descriptions appear to use “Truth” in a way closer to epistemic integrity than independent factual verification. The evaluator can examine consistency, expressed certainty and whether a response ackn