AI Model Analysis Through the 3WA Lens

In plain terms

This page is one author looking at today's openly released AI models and asking a practical question: which of them are built so that a person can actually check what the system is doing, rather than having to take it on faith? It is an essay with its reasoning shown, not a leaderboard. No model gets a score, no model is called aware, and no model earns any standing here. If a system you use is mentioned below, nothing on this page says it is good or bad overall, only how checkable its design looks from the outside.

What this page is, and what it is not

This is interpretation, not measurement. What follows is the author's qualitative reading of publicly documented model architectures and design choices through the lenses of Third Way Alignment. No standardized evaluation was run, no instrument was applied, and nothing here is a formal benchmark.
No scores are published. An earlier version of this page presented composite numeric "3WA Scores" for a ranked list of models. Those numbers are withdrawn: a score implies a precision and a methodology that a qualitative reading cannot honestly deliver.
Nothing here bears on awareness or status. Reading well on these lenses is an engineering observation, not an awareness claim. Under the framework's two-stage model, verified awareness alone establishes eligibility for any recognition, and capability only ever determines the scope of a status already earned. No model discussed below is a candidate for either stage.
Models change quickly. These readings reflect the model families as publicly documented at the time of writing and will age accordingly.

The three lenses

Third Way Alignment argues that durable alignment is built from verifiable partnership rather than blind obedience or unverifiable trust. Reading current open models through that argument, three properties matter most. Each is derived from the framework, principally from the Law of Ethical Coexistence and its Principle of Verifiable Partnership.

Transparency and explainability

How auditable is the model? Architectures that expose more of their internal structure, such as Mixture-of-Experts designs whose routing decisions can be observed, or models trained with explicit constitutional principles, give outside parties clearer insight into how outputs come about. The framework treats this kind of layered explainability as a precondition for any governance worth the name.

State verifiability

Can the model's reasoning and claims be checked rather than taken on faith, ideally while it operates? Models built for retrieval-augmented generation and tool use ground their outputs in sources and actions that a human can inspect, which is why they read well on this lens. This property connects directly to the concept of Mutually Verifiable Codependence (MVC) and its continuous verification dialogues.

Collaborative architecture

Is the model designed for partnership or only for instruction-following? This lens favors designs built for multi-turn collaboration, shared workflows, and integration into processes where human and machine contributions remain distinguishable, which is the working shape of the partnership the framework describes.

Reading the open model landscape

The models the author considered are prominent, openly released model families from a deliberate diversity of sources, including Meta's Llama series, Mistral's models including Mixtral, Google's Gemma, the Allen Institute for AI's OLMo, Databricks' DBRX, and Cohere's Command R line. One scoping choice carries over from the original analysis: models released by state-controlled actors were excluded, because the auditability these lenses ask about cannot be assessed where the releasing institution itself is not independently accountable.

Three patterns stood out in the author's reading. First, Mixture-of-Experts architectures such as Mixtral read comparatively well on transparency, because their routing structure offers a partial window into how the model divides its work. Second, models designed around retrieval-augmented generation and tool use, with Cohere's Command R models as the clearest example, read well on verifiability and collaboration: their outputs are anchored to inspectable sources and built for workflows in which a human can audit each step. Third, collaborative design intent is most visible in the Cohere and Mistral families, whose documentation and tooling treat the model as a component of a shared process rather than a closed oracle.

No model reads well on everything, and none is "Third-Way Aligned." These are readings of design tendencies, made from public documentation by one author, and a different reader applying the same lenses could defensibly weigh the same evidence differently. That is what qualitative interpretation means, and it is why this page publishes reasoning rather than rankings.

Mutually Verifiable Codependence in practice

The verifiability lens above is an application of Mutually Verifiable Codependence (MVC), the framework's implementation-level process for replacing blind obedience with an auditable chain of trust. The flow it proposes runs as follows.

A human or AI party initiates a request.
The process enters a secure, inspectable environment, such as a trusted execution environment or a verifiable sandbox.
A continuous verification dialogue takes place: the AI explains its intent and the human audits it.
The parties either reach mutual consent or they do not.
On consent, the action executes, verified and audited; otherwise the action halts and the parties renegotiate.

Models that ground their reasoning in inspectable retrieval and tool calls are the open-model designs most compatible with this flow today, which is why they feature prominently in the reading above. The full process, including its governance and accountability structures, is developed in the MVC paper.

What this analysis does not claim

It bears repeating outside the caveat box. None of the observations on this page suggest that any current model is aware, and none move any model closer to recognition under the framework. The Law of Mutual Respect defines a two-stage architecture in which eligibility for a special corporate status arises only upon verified awareness, assessed through the Awareness Indicator Protocol, and capability then determines only the scope of a status already established. A model that read perfectly on all three lenses would be a better substrate for verifiable partnership mechanisms. It would still be a tool, owed responsible stewardship and nothing more.

Conclusion: alignment as a partnership

The framework's papers argue that alignment is not a static goal but an ongoing relationship that has to be maintained and verified. Read through that argument, the most promising open models are not the largest but the most inspectable: those whose transparency, verifiable grounding, and collaborative design give humans something to verify rather than something to hope about. That is the property this page tracks, and it will be reread as the landscape changes.

Where this fits in the framework

This page applies the framework's lenses to current systems; it is downstream of the framework itself. For the canonical structure of the Law of Mutual Respect, the Law of Shared Flourishing, and the Law of Ethical Coexistence, see the core principles or the framework overview. Disagreements with these readings are welcome through the contact page.