The early twentieth-century Japanese philosopher Kitaro Nishida, the central figure of what became the Kyoto School, spent much of his life on a deceptively simple question: where does a thing's meaning reside? His answer was that meaning is not contained in the thing itself but in the place that holds it. He called this place basho, and built from it a logic in which the field comes first. The situation, the context, the predicate is not something a subject acquires after the fact. It is what lets the subject be determinate at all.
It is an old idea, from a tradition far from machine learning, and it turns out to describe, with uncomfortable precision, the central mistake in how we build language models.
Context as content
Ask how a transformer treats context and the answer is: as content. Everything that conditions a model's behaviour, the system instruction, the role, the standing constraints, is poured into the same context window as the ordinary material of the conversation, as one more stretch of tokens competing for the same attention. The field that is supposed to hold the conversation is placed inside the conversation, as an item within it.
Said plainly, the conditioning field is being stored as another piece of conditioned content. And because attention is flat, that misfiling has consequences. As the exchange grows, the governing conditions are crowded by the very content they were meant to govern. What should have been the place that holds the meaning becomes just another object in the pile.
The logic of place, applied
Nishida's correction is to reverse the order. The place determines what stands in it. The field constitutes the subject, rather than the subject reaching out to attend to the field. Applied to a language model this is not mysticism; it is an architectural instruction. Context should not be a payload the model attends to. It should be the field that conditions what the model can generate, operating from a level above the content rather than from within the same flat space.
It helps to see the levels. There is the content, what is being discussed. There is the role or manner, the space in which that content is held. And there is a further level still, the governing conditions, the norms and boundaries that condition everything beneath them without themselves being just another thing on the table. A structured cognition keeps these levels distinct, the higher conditioning the lower. A transformer collapses all three onto one plane and lets them compete. That collapse is not a tuning problem to be smoothed away with scale. It is structural incoherence: putting the thing that should govern into the very place where it can be governed.
The conditioning field cannot be safely stored as another piece of conditioned content.
Two kinds of competition
This gives a clean way to separate two failures that are usually run together. The first is content competing with other content: the model has too much to attend to and loses the relevant fact somewhere in the middle of it. That is a retrieval problem, real but reasonably well understood. The second is content competing with a governing condition: a value, a role boundary, a safety constraint, gradually out-weighed by the ordinary material it was meant to hold in check. That second kind is not a retrieval problem at all. It is an alignment problem, and it is the one almost no one names, because once you have decided that context is content, the two look identical. The logic of place pulls them apart. A field and an object are not the same kind of thing, and they fail in different ways.
What it asks us to build
Treating context as place rather than content is a design programme, not a metaphor. It points towards architectures that give the governing field its own standing rather than a share of the same attention budget: levels with separate budgets, periodic re-grounding of the conditions partway through a long interaction, and constraint that runs in both directions, the field shaping the content from above while the content informs the field from below, settling across more than one pass instead of resolving in a single sweep. There is real precedent for the machinery. Part-whole architectures such as Hinton's work on representing hierarchy, predictive-coding models, equilibrium models that iterate to a settled state: the components exist. What has been missing is the reason to assemble them this way, and the recognition that the reason is alignment, not efficiency.
A note of discipline is owed here, and Nishida would have insisted on it. Philosophy earns its place in an engineering argument only if the engineering would be different without it. The test is plain: remove the philosophy, and does the architecture change? Here it does. The logic of place yields a concrete and falsifiable prediction, that a system which treats context as a governing field, in bounded and multiple passes, will hold its alignment under growing context better than a flat, single-pass system, and by more than improved retrieval alone can explain. If that prediction fails, the idea was decorative and should be dropped. We do not think it will fail.
Place over position
The window is the wrong unit. A position in a sequence tells you where a token sits; it does not tell you what holds it. Nishida's lesson, arrived at a century early and from an entirely different inquiry, is that meaning belongs to the place, and that if you want a system to behave appropriately you must design the place that conditions it, not merely fill the window it reads from. That is what we mean, at Haku Labs, by context architecture: building the field, not just supplying the content.