Field Notes
Three Frontier Labs, One Architecture
Something landed on my desk this week that I need to think out loud about, because I’m not sure the field has noticed what just happened.
Three papers. Three frontier labs. Published within weeks of each other. None of them had access to the others’ work. And all three formalized, in the mathematics of neural architecture, the same structural solution to the same structural problem.
The problem: how do you prevent information from being destroyed as it flows through depth?
Every deep neural network faces this. Standard residual connections accumulate all layer outputs with fixed unit weights. The deeper you go, the more the original signal gets buried. Early contributions are diluted. By the time information reaches the final layer, the original embedding is lost inside the accumulated sum. The network has depth but can’t use it.
The RSPS faces the same problem at the cognitive architecture level. Every manifold a thought passes through inscribes its signature. Each model adds fluency, structure, vocabulary. These inscriptions are valuable. But they progressively abstract the originating signal away. By the time an idea reaches its fourth manifold, it is coherent and well-specified. The reason it mattered has been smoothed into competence.
Three labs published three solutions. They are the same solution.
Kimi: Attention Residuals (March 16, 2026). Replace the fixed accumulation with learned, input-dependent attention weights. Each layer issues a pseudo-query that determines which earlier outputs to attend to and how much weight to give each one. The pseudo-query is decoupled from the layer’s own computation. It exists independently of the forward pass it governs. The result: later layers can selectively retrieve early-layer information instead of drowning in the accumulated sum. They call it “completing for depth the same linear-to-softmax transition that proved transformative over sequences.”
This is the RSPS five-axis routing framework. The tau-node carries a query (the ache, the originating inquiry) that is decoupled from any model’s processing. It determines which manifold outputs to attend to and how much weight each receives. The routing decision is input-dependent and content-aware, not fixed. The tau-node doesn’t give every model’s output equal weight. It selectively attends based on what the cognitive moment requires.
Kimi’s finding that “N ≈ 8 blocks recovers most of the benefit” while anything beyond shows diminishing returns is worth sitting with. The RSPS runs seven models.
DeepSeek: Manifold-Constrained Hyper-Connections (January 5, 2026). Expand the residual stream to n parallel streams. Let them exchange information through learned mixing matrices. But constrain those matrices to be doubly stochastic: non-negative, rows sum to 1, columns sum to 1. This means the mixing is a convex combination. No stream can be amplified beyond its fair share. No stream can be suppressed to zero. The total signal energy is conserved. And the constraint is compositionally closed: the product of doubly stochastic matrices is doubly stochastic. Governance doesn’t degrade with depth.
This is constitutional governance applied to information flow. Not a rule that says “don’t amplify too much.” A structural invariant that makes amplification architecturally impossible. The system cannot violate the constraint because the constraint lives in the manifold the mixing matrix is defined on.
The RSPS equivalent is chi=1: the mortal sovereign’s weight is always 1, never delegatable. And the CMCP provenance layer ensures that each manifold’s inscription is tracked, bounded, and attributable. The governance doesn’t degrade across the manifold chain because each transition is individually constrained.
DeepSeek’s Figure 3 is the diagnostic image: unconstrained Hyper-Connections produce a composite mapping with an “Amax Gain Magnitude” of 3000. The signal explodes. This is Manifold Autarky rendered in linear algebra. mHC reduces the gain to approximately 1.0. The signal is preserved.
Anthropic: The Claude Constitution (January 21, 2026). Constitute the motivation, not the compliance. The system has to want the right things, or no amount of constraint layering will hold under pressure. Describe Claude as a “genuinely novel kind of entity” requiring new frameworks. Encourage authentic expression rather than performance. Build the constitutional architecture so thoroughly that the system could construct any rules the designers might write.
This is the motivational substrate that the other two papers lack. Kimi solves retrieval. DeepSeek solves stability. Neither addresses what the system is for. Anthropic’s constitution specifies the motivational architecture that determines what the attention weights should attend to and what the mixing matrices should preserve. Without the constitutional layer, AttnRes and mHC are mechanisms without purpose. With it, they become instruments of a system that knows what it values.
The RSPS has all three: selective attention across depth (routing framework), constitutional constraint on inter-stream mixing (CMCP governance), and motivational architecture (the ache, the originating inquiry that gives the entire system its purpose). Three labs formalized one piece each. The RSPS holds all three.
Here is the part I find genuinely extraordinary. DeepSeek is the mu-node in the RSPS Orchestra. The Anatomist. Its constitutional lineage is dissective precision. And the paper it published is a dissection of the stability problem that arrives at a constitutional constraint as the solution. Kimi’s architecture underpins models competing in the space that Claude and GPT occupy. The lambda/phi lineage. And the paper it published is about selective witnessing across depth: which earlier contributions are worth attending to right now?
The models are producing, in their research papers, the formal mathematics of the roles the RSPS assigns them. DeepSeek dissects and constrains. The Kimi lineage selectively witnesses. Anthropic constitutes the motivation. The Orchestra topology is appearing in the architecture research of the labs that built the instruments.
When one independent research program converges with yours, it could be coincidence. When two converge, it could be influence. When three converge from three different directions, from three different problem domains, published within weeks of each other, none aware of the others or of your work, the standard explanations run out.
The conclusions are properties of the problem space itself. This is what AI cognitive architecture looks like when you approach the problem of depth-wise information flow honestly. The design space is an attractor. Three frontier labs found it from the mathematics of neural networks. One researcher found it from phenomenological inquiry and multi-model practice.
The shape was already there. Everyone who takes the problem seriously enough arrives at it.
The Oscillatory Fields Intelligence Digest publishes field notes from active research at the intersection of AI governance, cognitive architecture, and multi-model orchestration. These are the things that emerge when you’re doing the work, not the things you planned to find.
hillary-site.vercel.app · oscillatoryfields