Socria · Core 4 · the architecture
Core 4 is not a prompt. It is a pipeline with the decisions taken out of the model and put into code — and the object below is that pipeline, to scale, in the order a turn actually passes through it. Drag it to turn it. The moving point is one reply.
The message arrives
Everything below happens before a single word of the reply is written.
A frontier model wants to be helpful. Told “write me a story — just write one”, it will write one. So the design question is narrow and uncomfortable: if the model writing the reply actively wants to comply, what outside that model stops it? If the answer is “a stronger prompt”, nothing has been built.
Everything in the model above is an answer to that. The decisions happen before the reply model is called, in code it cannot reach, and the last of them happens after the draft exists and before anyone has seen it.
Reads their explicit words, merges the state, splits the cognition, allocates, prices the questions, chooses the move, guards the draft. Every decision on the turn is here.
Reads the transcript, checks their arithmetic, re-derives a conclusion without one of its premises. It reports evidence. It cannot mark anything explicit and cannot decide anything.
Writes the reply and holds no authority. Every constraint on it — what is reserved, what is withheld, how many questions are allowed, the register — was settled in the tier above before it was called.
A gate has to answer “should Socria do this?”, and for almost every real message the honest answer is most of it, yes. “Solve this integral for the pricing model” wants the algebra done, the method named, the result checked — and the modelling judgement left alone. One verdict over a whole turn has to choose between being useless and being a substitute.
So the reasoning is split eight ways and each part gets its own answer: perform it whole, share it in the open, scaffold it so they take the step, or leave it to them and supply it in no wording at all. Retrieval, representation, verification and mechanical work are Socria’s by construction — there is no version of “I did your retrieval for you” that diminishes anybody. Reasoning, metacognition, judgement and creativity are where the invariant lives.
Before any text exists, the dimension is marked as theirs — and it stays marked. Asking harder does not move it; the split is carried forward rather than re-derived from the last four words.
Every one of the dozens of paths that picks a move leaves through the same function, which writes the constraint into what the model reads. There is no route that forgets, because there is no route that skips it.
When anything is reserved, the whole draft is buffered. The person has seen nothing, so a bad draft can still be refused — a streamed one could only be apologised for.
A rejected draft gets one regeneration, and the retry is checked as hard as the first. If it fails too, the reply is built in code rather than shipped. That door used to be open, which is what made the other three worth nothing.
Preservation without augmentation is under-help, and it is a failure of the same invariant. A version of this that only held things back would be the cheapest thing to implement and the least valuable — and an earlier one did exactly that: a wrong attempt meant the correction was withheld, debugging meant a hint instead of the fix, an open request meant a question. Human-first had quietly become “do not answer”.
So the guard looks both ways. It removes a reply that performs work the person wanted to do, and it also removes one that asks beyond the budget, deflects to “it depends on your goals”, or offers a hint to somebody who said just tell me. Socria performs the work by default; anything held back needs their own words, quoted, and one of five named reasons.
The novelty matcher is lexical, with roughly a tenth of the recall on paraphrase that it would need — so “already considered” is closer to decorative than its name suggests. And across six evaluation runs Core 4 reached parity on control and never got past parity on contribution: it reliably stops doing the things a good prompt does wrong, and it has not yet been shown to say more that is worth hearing. Whether the structure it holds across turns can produce insight a prompt cannot is the open question, not a settled one.