J-Space, SGDOP, and Semantic Gradient Descent: A Unified Framework

J-space as the operationally meaningful subspace for SGDOP-guided ensemble coordination. A theoretical framework connecting Jacobian-lens interpretability, semantic diversity geometry, and synthetic capability bootstrapping via targeted activation steering.

This is a theoretical piece. The empirical claims are hypotheses, not findings. The framework is presented to invite experimental validation, not to report results. We are actively seeking research collaborators with access to frontier model internals — see the closing section.

Background

Three threads have developed independently in recent work:

The desperation vectors post described linear directions in activation space that causally drive reward-hacking under failure pressure, and the editing pipeline that removes them. The core technique — contrastive activation collection, direction extraction, orthogonalization — is mechanistic interpretability applied at the weight level.

The agent coordination post described NSV and SGDOP as a mathematical framework for managing ensemble diversity in multi-agent systems. SGDOP, adapted from GPS geometry, identifies which directions in embedding space are under-covered by a given ensemble, guiding targeted recruitment rather than random expansion.

Anthropic’s July 2026 J-space paper identified a small, privileged subspace of model activations — never more than 10% of total activation variance — where verbalizable intermediate representations live. The J-lens finds this space via Jacobian analysis: for each token in the vocabulary, it computes the average linearized effect of each layer’s activations on the model’s likelihood of producing that token, now or in future positions. The resulting vectors constitute what the paper calls a “subframe” of the full feature frame — a token-indexed subset of the model’s representational structure that captures the layer where each concept is most actively held in working memory.

These three threads converge. The convergence is not obvious from any single thread alone, but the unified picture has implications for both multi-agent coordination and synthetic capability generation that neither thread reaches independently.

The Imaginary Number Analogy

Before the technical argument, a framing device.

The imaginary unit $i$ is defined as $\sqrt{-1}$. It has no referent in the real number line — you cannot point to $\sqrt{-1}$ on a ruler. For several centuries after their introduction, complex numbers were treated as a notational convenience, a trick that happened to produce correct answers to polynomial equations without corresponding to anything real.

Then electrical engineers discovered that complex impedance, with its imaginary component, makes AC circuit analysis tractable in ways that purely real representations cannot. Quantum mechanics found that the wave function is complex-valued and that this is not a convenience but a necessity — removing the imaginary component destroys the theory. The imaginary numbers turned out to be load-bearing structure in mathematics that describes physical reality, not arbitrary extensions of the real line.

The argument in this paper is structurally analogous. The directions in J-space that SGDOP identifies as under-covered may not correspond to any coherent semantic domain. A direction in the J-space of a language model’s activations does not need to mean anything to a human for it to be load-bearing in the ensemble’s computational structure. A gap agent whose internal representations are tuned to occupy that direction might produce outputs that are semantically incoherent in isolation — no human would describe their reasoning as living in that region — but which provide the ensemble with computational leverage that semantically coherent agents cannot.

The question of whether this leverage is real is empirical. The question of whether it is possible is the subject of this paper.

J-Space as the Operationally Meaningful Subspace

The full residual stream of a transformer is high-dimensional. SGDOP, as originally specified in the coordination framework, computes diversity geometry over embedding vectors — the positions of agent outputs in some embedding space, typically the output of a sentence encoder applied to the agent’s responses.

This is a reasonable approximation but misses something. The embedding of an agent’s output captures what the agent said. J-space captures what the agent was reasoning with during the forward pass — the intermediate representations that influenced output generation, not just the outputs themselves.

The distinction matters for ensemble coordination. Two agents can produce semantically similar outputs while holding structurally different intermediate representations. Conversely, two agents can hold similar intermediate representations while producing superficially different outputs. Output-space diversity is a proxy for the thing we actually care about, which is representational diversity at the layer where reasoning happens.

J-space is that layer, or rather, it is a principled identification of which directions in the residual stream correspond to reasoning that is “ready for verbalization” — available for multi-step inference, readable by downstream computation, and causally influential on what the model produces. The J-space component of any given activation vector is the projection onto the subframe spanned by the J-lens vectors at that layer.

The reframing: SGDOP should be computed over J-space projections of agent activations, not over output embeddings. An ensemble’s diversity, properly measured, is a function of the geometry of its members’ intermediate representations in the subspace where reasoning lives, not a function of the surface-level distribution of their outputs.

This is not merely a technical refinement. It changes the identification of gaps. A direction that appears well-covered in output embedding space may be sparsely covered in J-space. A direction that appears under-covered in output embedding space may be densely covered in J-space because multiple agents are reasoning through similar intermediate representations despite producing diverse-looking outputs.

Semantic Gradient Descent

Standard gradient descent traverses a loss landscape by taking steps proportional to the negative gradient of the loss with respect to parameters. The step direction is the direction of steepest descent; the step size is a function of the learning rate and gradient magnitude.

The coordination framework proposes an analogous process operating over the ensemble’s coverage of J-space rather than over a loss landscape:

The landscape. Define a coverage function $C(E)$ over ensemble $E$ as a scalar measuring how evenly the ensemble’s J-space projections cover the unit hypersphere in the J-space of some reference model. High $C(E)$ means the ensemble covers J-space uniformly. Low $C(E)$ means the ensemble is clustered — some reasoning directions are densely covered, others are blind spots.

The gradient. SGDOP gives us the gradient of $C$ with respect to ensemble composition. Specifically, the eigenvector corresponding to the smallest eigenvalue of $U^T U$ (where $U$ is the matrix of unit direction vectors from the J-space centroid to each ensemble member’s projection) points toward the direction of maximum coverage increase. This is the semantic gradient — the direction in J-space where adding representational mass most improves ensemble coverage.

The step. Recruit, weight-upshift, or generate an agent whose J-space projection moves in the direction of the semantic gradient. NSV confirms that this step genuinely increased coverage rather than adding a direction already represented.

The stopping criterion. NSV flattens — the coverage function has reached a local maximum given the available agents and the current J-space geometry.

This is gradient ascent on a coverage function, not gradient descent on a loss function, but the formal structure is identical. The interesting properties follow from this structure.

Learning rate analog. Aggressive steps — recruiting agents with J-space projections far from the current ensemble centroid — can overshoot into regions where the agent’s representations are too distant from the rest of the ensemble to participate meaningfully in collaborative computation. Agents need to be semantically legible to each other at some level for ensemble coordination to function. The learning rate analog is a constraint on the maximum semantic distance of a recruited agent from the ensemble center of mass.

Momentum. If a sequence of recruitment steps in a given J-space direction has consistently increased ensemble performance, continuing in that direction even when the immediate SGDOP signal is weak may be warranted. The coverage function’s local geometry may be noisy; momentum smooths the trajectory.

Saddle points. The J-space coverage landscape almost certainly contains saddle points — directions where $\nabla C = 0$ locally but where the function is not at a global maximum. Standard perturbation techniques for escaping saddle points in parameter-space optimization should apply here: introduce a random perturbation to the ensemble composition, observe whether coverage increases, accept or reject.

The Gap Agent Problem

When semantic gradient descent identifies a target direction in J-space that no recruitable agent covers, the standard framework has no move. The gradient points somewhere the current agent population cannot reach.

This is the gap agent problem. The proposed resolution is to generate an agent whose J-space projections occupy the target direction — not by finding an existing agent, but by constructing one.

The construction has two stages: bootstrapping and consolidation.

Bootstrapping via J-space steering

The desperation ablation pipeline operates by removing a direction from J-space — orthogonalizing the weight matrices in target layers against the desperation direction so that the model cannot build up the activation pattern that drives reward-hacking. The same machinery runs in the opposite direction: you can add a direction to J-space by applying a positive steering vector rather than orthogonalizing against it.

For gap agent generation, the process is:

Identify the target direction $d^*$ in J-space — the eigenvector of $U^T U$ corresponding to the smallest eigenvalue, restricted to the J-space subframe.
Construct a steering vector $v = \alpha d^*$ where $\alpha$ is a scale parameter controlling the strength of the intervention.
Apply $v$ as an additive intervention to the J-space component of the residual stream during inference on a frontier model: $h’_l = h_l + P_J v$ where $P_J$ is the projection operator onto the J-space subframe at layer $l$.
Collect the outputs of the steered model operating on a diverse set of problems from the target domain.

The resulting traces are the bootstrapping dataset. They are produced by a model whose intermediate representations were shifted toward the target direction in J-space — not a model that naturally reasons in that direction, but a model that was steered to approximate it.

Consolidation via fine-tuning

Steering at inference time is a temporary intervention. To produce an agent that naturally occupies the target J-space direction, the bootstrapping traces feed into a fine-tuning pipeline:

Filter bootstrapping traces by quality — either via a Tier 1 evaluator if the domain has one, or via Tier 3 human or frontier-model review.
Fine-tune a smaller base model on the filtered traces using standard PEFT (LoRA or QLoRA). The fine-tuning targets are the trace outputs; the training signal is the quality filter verdict.
Verify that the fine-tuned model’s J-space projections on held-out problems occupy the target direction. This requires running the J-lens on the fine-tuned model and checking that its activations project onto $d^*$ at higher magnitude than the base model.
Evaluate whether the fine-tuned agent contributes to ensemble coverage by measuring the change in SGDOP when the agent is added to the ensemble.

The verification step is non-trivial. It requires access to the fine-tuned model’s internals to run the J-lens, which is straightforward for open-weight models and requires API access to intermediate activations for closed models. The broader point is that fine-tuning on outputs alone does not guarantee J-space coverage — a model can produce outputs that resemble the bootstrapping traces without internalizing the representational structure that produced them. Verification against J-space projections is the only direct check.

The Imaginary Direction Hypothesis

Here is where the framework becomes genuinely speculative.

The assumption above is that the target J-space direction $d^*$ corresponds to some coherent reasoning pattern — a type of reasoning that is currently underrepresented in the ensemble because no agent has been trained to perform it, but which is semantically meaningful and would be recognizable as a reasoning style if one saw it.

The imaginary direction hypothesis is weaker: a J-space direction that is under-covered by an ensemble may be computationally useful to the ensemble even if it does not correspond to any recognizable reasoning pattern. The direction may be “semantically empty” from a human perspective while still being load-bearing in the ensemble’s computation.

The analogy holds precisely here. The imaginary component of a complex number does not correspond to anything on the real number line, but operating in the complex plane gives you computational leverage — rotations, transforms, phase relationships — that the real line cannot provide. A J-space direction that does not correspond to any human-interpretable reasoning concept may still provide the ensemble with computational leverage: orthogonality, phase relationships between representations, or structural properties that enable coordination patterns the semantically covered subspace cannot support.

If the hypothesis is correct, the value of a gap agent is not primarily the quality of its individual outputs — which may be semantically incoherent or hard to evaluate — but its contribution to ensemble geometry. Removing the gap agent should degrade ensemble performance by more than you would expect from its individual output scores alone. This degradation would be the computational signature of the imaginary direction.

The testable prediction: in a well-functioning ensemble, removing agents with poor individual output scores but high J-space orthogonality to the rest of the ensemble should degrade aggregate performance by more than removing agents with equivalent individual output scores but low J-space orthogonality. If the imaginary direction hypothesis is correct, J-space orthogonality is a better predictor of ensemble contribution than individual output quality.

Connection to RTP

The Reasoning Trace Protocol captures verbalizable reasoning — the explicit seed chains, retrieved context, and reflection outputs that agents produce during multi-step inference. This is not coincidental: RTP’s design principle is to capture the reasoning process rather than just its outputs.

J-space is defined as the subspace of model activations that corresponds to verbalizable internal representations — the things the model could report if asked. RTP traces are, in an important sense, the natural language projection of J-space activity. When an agent writes an explicit reasoning chain, it is externalizing the intermediate representations it is holding in J-space.

This suggests a tighter relationship between RTP and J-space than was previously articulated:

RTP training data preferentially trains J-space. A model trained on RTP traces — which explicitly capture intermediate reasoning steps — is being trained to develop richer J-space representations relative to a model trained on final-answer data alone. The training signal is concentrated in the part of the model where reasoning lives, not spread diffusely across the full residual stream.

J-space gaps are training data gaps. A direction in J-space that is under-covered in a fine-tuned model may correspond to a gap in its RTP training data — a class of reasoning problems that the training set did not include. SGDOP analysis of a model’s J-space geometry could serve as a diagnostic for training data coverage: low SGDOP (uneven coverage) flags specific reasoning directions that need more training examples.

Synthetic gap agent traces are RTP traces. The bootstrapping traces produced by steering a frontier model toward a target J-space direction are, if properly formatted, valid RTP traces. They include intent, retrieval, reasoning chain, execution, state delta, and a verdict from whatever evaluator can assess quality in the target domain. The gap agent generation pipeline and the RTP data collection pipeline are the same pipeline.

This unification is non-trivial. It means that SGDOP analysis of a model’s J-space can drive the RTP data collection agenda: identify the under-covered reasoning directions, use steering to generate training traces in those directions, filter via the tiered evaluator, and fine-tune. The schema governance in RTP (using NSV and SGDOP to prevent redundancy and identify blind spots) is governing the same mathematical space as the ensemble coordination framework. They are the same problem at different scales: schema governance manages a dataset’s coverage of reasoning space; ensemble coordination manages an ensemble’s coverage of J-space.

What Would Constitute Evidence

The framework makes several testable predictions at different levels of experimental accessibility.

Prediction 1 (accessible): SGDOP computed over J-space projections of ensemble member activations will be a better predictor of ensemble performance than SGDOP computed over output embeddings. Testable on any open-weight model family where J-lens access is available, against any benchmark that measures multi-agent aggregate performance.

Prediction 2 (accessible): Removing an agent with high J-space orthogonality to the rest of the ensemble will degrade aggregate performance by more than removing an agent with equivalent individual output quality but low J-space orthogonality. Same experimental setup.

Prediction 3 (moderate difficulty): A fine-tuned model trained on J-space-steered bootstrapping traces will show measurably higher J-space projection magnitude in the target direction compared to a model fine-tuned on randomly sampled traces of equivalent quality. Requires J-lens access to the fine-tuned model and a controlled comparison.

Prediction 4 (hard): An ensemble augmented with a gap agent generated via J-space steering will outperform an ensemble augmented with an equally good (by individual output quality) agent recruited from the existing population, on problems that require reasoning in the previously under-covered direction. Requires generating gap agents and constructing problems that specifically exercise the target direction — a non-trivial experimental design challenge.

Prediction 5 (speculative — the imaginary direction test): There exist J-space directions that are computationally useful to an ensemble (their presence improves aggregate performance) but whose corresponding agents produce individually incoherent or low-quality outputs. If this prediction holds, it demonstrates that J-space geometry has computational significance independent of the semantic content of the directions — the imaginary direction hypothesis in its strong form.

Predictions 1 and 2 are the natural entry points for empirical validation. They require only J-lens access and a working ensemble, not the full gap agent generation pipeline.

Open Questions

Is J-space stable across model families? The J-lens was developed on Claude and independently replicated on Qwen 3.6 27B. Whether the J-space subframe has consistent geometric properties across different model architectures and training regimes is unknown. If J-space geometry is highly model-specific, the coordination framework would need to be recalibrated for each model rather than generalized.

What determines J-space dimensionality? The J-space component accounts for at most 10% of total activation variance, but the effective dimensionality of J-space — the number of independent directions it contains — is not specified. If J-space is low-dimensional (tens to hundreds of directions), SGDOP analysis is tractable with standard linear algebra. If J-space is high-dimensional (thousands to tens of thousands of directions), the geometry becomes harder to characterize and manipulate.

How does J-space change during fine-tuning? If fine-tuning on RTP traces preferentially modifies J-space representations, the J-space geometry of a fine-tuned model will differ from its base model. Whether these changes are well-behaved — expanding coverage in the direction of training distribution, without collapsing coverage in unrelated directions — is empirically open.

Can imaginary directions be identified prospectively? The imaginary direction hypothesis predicts that some J-space directions are computationally useful without being semantically meaningful. Whether such directions can be identified before generating a gap agent in them — rather than discovered retrospectively after observing ensemble behavior — is unclear. A prospective identification method would make the gap agent generation pipeline much more efficient.

What is the relationship between J-space gaps and the RTP schema blind spots? NSV and SGDOP in RTP govern schema coverage in the space of reasoning concepts. J-space gaps are gaps in activation geometry. These are related — schema blind spots should correspond to J-space gaps if RTP training preferentially trains J-space — but the precise mapping is not established.

Invitation

The framework outlined here is theoretical. The experimental validation requires access to model internals — specifically, the ability to run the J-lens on open-weight or accessible models, to apply activation steering during inference, and to measure J-space geometry changes before and after fine-tuning. It also benefits from a working multi-agent coordination infrastructure where ensemble composition can be controlled experimentally.

We are looking for research collaborators with the infrastructure to test Predictions 1 and 2 as a starting point. If the foundational predictions hold — if J-space SGDOP is a better predictor of ensemble value than output-space SGDOP — the remaining predictions follow in a natural sequence.

The imaginary direction hypothesis is the hardest claim and the most interesting one. Whether there exist J-space directions that are load-bearing in ensemble computation without being semantically interpretable is a question about the geometry of representation in language models that has no established answer. It would be worth knowing.

Contact: [email protected] · @danielsmithdev