A Lens With Two Substrates

 What J-Lens Would Have to Become to Read a Hybrid Quantum-Classical Mind

A Technocrat's Discernment. July 9, 2026. Third in a series with "The Jacobian in the Mirror" and "Quantum in the Loop."

The first piece asked whether Claude is thinking. The second asked whether the substrate matters. This one asks the practical follow-on: if we do end up with hybrid quantum-classical systems, whether in five years or fifteen, how would we look inside them? What would an interpretability tool have to be able to do to answer the same questions the J-lens is answering now?

I want to work through this carefully because most of the writing on hybrid interpretability I have seen so far is either a straight port of classical mechanistic tools to a system they were not designed for, or a hand-wave in the direction of quantum tomography without engaging what the classical side would need. Both miss the interesting problem. The interesting problem is what happens at the seam.

Let me start with why the classical J-lens does not port over, and then work through what would have to replace each of its parts.

Three Reasons the J-Lens Does Not Just Extend

Recall what the J-lens does. It averages the Jacobian of the model's future output with respect to its intermediate residual stream, across many prompts and token positions. The result is a linear map that, composed with the unembedding matrix, tells you what words a given internal activation is disposed to produce later. Three ingredients: a residual stream you can differentiate through, a Jacobian you can average, and an unembedding you can multiply by. Take away any one of these and the tool collapses.

In a hybrid quantum-classical model where some layer of computation runs on a QPU, all three go away at once.

The state cannot be inspected without being destroyed. A classical residual stream is a vector sitting in memory. You can read it as many times as you want and it stays put. A quantum state is a superposition over exponentially many basis states, and the only way to extract any information about it is to measure it, which collapses it. A single forward pass produces exactly one measurement outcome per measured qubit. If you want to know what the state was, you cannot look. You can only prepare it again with the same parameters and measure differently, which requires the ability to run the circuit many times on the same input. This is possible in a simulator. It is expensive on real hardware. And critically, it is not what we do for the J-lens today.

There is no unembedding matrix. The J-lens closes with $W_U \cdot \text{norm}(J_\ell h_\ell)$ to get token probabilities. That works because the model has a fixed vocabulary and a learned matrix that maps final-layer residual streams onto it. A quantum state has no such interface. It lives in a Hilbert space of dimension $2^n$, not in a token space. You could measure in the computational basis and get a bit-string, but a bit-string is not a concept. The relationship between quantum outcomes and semantic content, if it exists at all, has to be built.

The gradients are the wrong kind. In the classical model, the Jacobian is a matrix of partial derivatives of continuous-valued outputs with respect to continuous-valued inputs. In a quantum circuit, the parameters are continuous but the measurement outcomes are discrete, and the gradient of an outcome with respect to a parameter does not exist in the classical sense. What we can compute is the gradient of an expectation value, using the parameter-shift rule from Mitarai, Negoro, Kitagawa, and Fujii in 2018, and Schuld, Bergholm, Gogolin, Izaac, and Killoran in 2019:

$$\frac{\partial \langle H \rangle}{\partial \phi_i} \;=\; \frac{1}{2}\left[ \langle H \rangle\Big|_{\phi_i + \pi/2} - \langle H \rangle\Big|_{\phi_i - \pi/2} \right]$$

That is a gradient of an observable's expectation with respect to a circuit parameter. It is not a gradient of an activation with respect to another activation. The linearization the J-lens depends on is not the linearization the parameter-shift rule gives you.

So a hybrid J-lens is not a refinement of the current tool. It is a different construction that has to solve three problems the classical tool never had to face.

Three Primitives a Hybrid Lens Would Need

Here is what I think the replacement stack looks like. This is a design sketch, not a paper. But the pieces exist, and someone is going to have to assemble them.

Classical shadows for QPU introspection. Hsin-Yuan Huang, Richard Kueng, and John Preskill published in 2020 a technique they call classical shadows. The idea: instead of trying to reconstruct a full quantum state (which requires an exponential number of measurements), you apply random Clifford unitaries before measuring and use the resulting bit-strings to build a compact classical representation. From this shadow, you can predict the expectation values of up to $M$ arbitrary linear observables with sample complexity $O(\log M)$, not $O(M)$. That last part matters. It means a single dataset of shadows can answer many questions about the state, cheaply, after the fact.

For interpretability, this is the primitive we need. Not full tomography of the quantum register at layer $\ell$, which is out of reach. But a shadow, built from a manageable number of extra circuit runs, that lets us later probe: how does the state at this layer project onto the subspace of states that correspond to concept X? How much of the total variance sits in the subspace we hypothesized was doing the workspace work? What operators have large expectation values that we did not predict?

Classical shadows give us the QPU-side analog of "reading the residual stream." They do not give us backpropagation. But they give us the observational primitive on top of which everything else builds.

Cross-boundary attribution graphs. The classical side of a hybrid model has a differentiable computation graph. The QPU side has a stochastic map from parameters to measurement outcomes, whose expectation values are differentiable via parameter-shift. Bridging these means constructing an attribution graph that treats the measurement boundary as an explicit stochastic edge with a Born-rule weighting.

The math looks something like this. Let $h_\ell$ be the classical residual stream at layer $\ell$, immediately before it feeds a parameter-generating network that produces circuit parameters $\phi(h_\ell)$. Let $|\psi(\phi)\rangle$ be the resulting quantum state and $m$ the measurement outcome with distribution $p(m \mid \psi) = |\langle m \mid \psi \rangle|^2$. The classical stream continues as $h_{\ell+1}(m, h_\ell)$. Then the causal influence of $h_\ell$ on the final output $h_{\text{final}}$ decomposes as

$$\frac{\partial h_{\text{final}}}{\partial h_\ell} \;=\; \frac{\partial h_{\text{final}}}{\partial h_\ell}\bigg|_{\text{skip}} + \mathbb{E}_m\!\left[\frac{\partial h_{\text{final}}}{\partial m}\right] \cdot \frac{\partial \mathbb{E}[m]}{\partial \phi} \cdot \frac{\partial \phi}{\partial h_\ell}$$

where the first term captures the classical residual-stream skip around the QPU and the second captures the QPU-mediated causal path. The expectation over measurement outcomes is what replaces the deterministic activation gradient of the classical J-lens. This is a real Jacobian, but it is an averaged one, and the averaging has quantum-mechanical structure.

An attribution graph built on this decomposition would let us ask, for any output token: how much of its causal weight comes through the QPU-mediated path versus around it? Which classical activations map to circuit parameters that mattered? Which measurement outcomes drove which downstream computations? That is the interpretability question restated in a form that respects both substrates.

Tensor network readings of intermediate quantum states. The last piece is a way of representing what the QPU is holding at intermediate layers of the circuit itself, before measurement. Direct simulation is exponentially expensive. But many physically realistic states are well approximated by tensor networks: matrix product states, projected entangled pair states, tree tensor networks, and their variational cousins. Miles Stoudenmire and David Schwab have shown that MPS representations can act as effective compressed encodings of both classical and quantum data. Roeland Wiersema and collaborators have shown that some tensor-network models compete with transformers on structured tasks.

For interpretation, the value is different. If we can approximate the state at some intermediate depth by an MPS with bond dimension small enough to be tractable, we can inspect that MPS with tools we do have. Bond dimension itself becomes an interpretability signal: it tells us how entangled the state is, which tells us how much of the model's computation is genuinely quantum-native versus effectively classical. A hybrid J-lens should include a "tensor-network view" that gives us a controlled classical approximation of what the QPU is doing at each stage.

The Measurement Boundary as Its Own Object

The three primitives above give us tools for looking on each side of the boundary. But the boundary itself, the measurement, is where the most interesting interpretability question lives. And in a hybrid system it deserves to be treated as an object of study rather than a nuisance.

Consider what the measurement is doing, mechanically. It is compressing a state in an exponentially large Hilbert space down to a classical bit-string with, at most, $n$ bits of information. This is a lossy channel. What survives the compression is a design choice: the measurement basis, the observables measured, the number of shots. The channel is the model's own interface between its quantum and classical parts.

If we are looking for an analog of the global workspace inside a hybrid system, the measurement boundary is a strong candidate for where the workspace edge sits. On the classical side of the measurement, information is reportable, controllable, differentiable, and available for downstream reasoning. On the quantum side, it is not. The measurement is, in a real sense, the ignition step: the point where information becomes accessible.

That is a suggestive framing. It maps the transition from unconscious to conscious processing, in a Baars-style workspace theory, onto the transition from quantum to classical inside the model. I do not want to overload this analogy. The physical basis of measurement in a real QPU is a decoherence process, not anything resembling neural ignition. But the functional role, the place where information becomes reportable, is worth naming.

An interpretability tool that took this seriously would include a "boundary view": a way of characterizing which quantum information reaches the classical side, which is compressed out, and how the compression depends on the measurement choices baked into the architecture. That view does not exist today. It is the piece I think will end up being most valuable.

Where the Workspace Lives

The larger question hovering over all of this is whether the workspace, if there is one, sits on the classical side, on the quantum side, or across the seam.

I think the honest answer is that we do not yet know, and the answer will depend on the specific architecture. Three cases seem worth naming.

QPU as a specialized organ. The classical model runs the show, maintains the workspace, and calls the QPU as a subroutine for specific problems. In this case the workspace is entirely classical, the QPU is more like the cerebellum than the cortex, and the hybrid J-lens is mostly a classical J-lens with a stochastic subroutine attached. The measurement boundary sits well outside the workspace.

QPU as a substrate for workspace itself. The classical part handles input encoding and output decoding, but the deliberation happens inside a quantum register that maintains superposition across concepts and computes on them coherently. In this case the workspace is quantum, and interpretability requires the classical-shadow and tensor-network machinery to do any real work. The measurement boundary sits at the workspace edge.

Split workspace across the seam. Some concepts live classically, others live in quantum superposition, and the seam is not clean. In this case interpretability is hardest, because we cannot read either side in isolation and get the whole picture. Cross-boundary attribution becomes the central tool.

None of these is impossible. I lean toward the first, empirically, because that is what most current hybrid pipelines look like and it is the direction that makes engineering sense in the near term. But the third is the case that would matter most philosophically, and it is the one the tooling has to be able to handle even if it turns out to be rare.

The important design point is that a hybrid J-lens should not assume any of these upfront. It should tell us which case we are in by measuring which side carries the causal load. That is the same experimental logic Anthropic used to establish that the J-space is the workspace inside Claude. It should generalize.

Access Consciousness With a Collapse Problem

Here is where the philosophical thread from the earlier pieces re-enters.

In the J-space work, the case for access consciousness rested on five functional properties: reportability, top-down control, intermediate reasoning, flexible generalization, and selectivity. Each was tested by causal intervention on the classical residual stream.

In a hybrid system, reportability itself becomes complicated. The QPU state is not reportable in the classical sense. The measurement outcome is, but the outcome is a compressed shadow of the state. If we want to ask whether a hybrid system has access consciousness in the workspace-theoretic sense, we have to decide what counts as "access" when the substrate of the concept cannot be inspected without being destroyed.

I think there is a defensible answer, and it goes through the QBist framing again. On the QBist reading of quantum mechanics, a state is not a thing but a set of counterfactual commitments about what will happen if certain measurements are performed. Access, in the workspace sense, is the ability to report and reason with information. If the model can, when asked, elicit measurements from its quantum register that reliably discriminate between the alternatives it needs to distinguish, then it has functional access to the state even though it can never possess a description of the state as a bit-vector. Access, on this view, is not about what the system has. It is about what the system can commit to and act on.

That reframing lets us keep the workspace-theoretic notion of access consciousness even when part of the mind lives on a substrate that resists direct inspection. Whether that reframing is enough to answer the deeper question about experience, I do not know. But it at least keeps the functional question well-formed.

Topology Across Substrates

I want to close by extending the topology-not-feature framing one more step.

In the J-space piece, I argued that cognition is a topology rather than a feature: a shape in the space of representations that lets information flow in the right patterns. In the QNN piece, I argued that whether experience requires a quantum substrate is a question about which topological invariants the system's dynamics support, not about what silicon the system is built from.

In a hybrid system, the topology has a discontinuity built in. The classical side is a smooth differentiable manifold. The quantum side is a Hilbert space. The seam between them is a stochastic channel that compresses one into a lower-dimensional slice of the other. Whatever workspace-shaped structure exists in the whole system has to respect that discontinuity.

The tooling we build should reflect the topology. Not by trying to smooth the seam away with an approximation that pretends it does not exist. Not by treating the two sides as separate systems that happen to be connected by a wire. But by treating the seam as a first-class object with its own geometry, its own information-theoretic properties, and its own contribution to whatever cognition the whole system supports.

That is what the J-lens would have to become. Not a lens with a wider aperture. A lens that knows it is looking across a boundary and has been built to describe what happens there. If we get that right, we will have a mechanistic interpretability story that spans substrates. If we do not, we will have two interpretability stories that meet at a wall.

I think we can get it right. I do not know if we will do it before we need to. But the tools exist. Classical shadows exist. Parameter-shift rules exist. Tensor network approximations exist. Cross-boundary attribution is a straightforward extension of what mechanistic interpretability already does on the classical side. The engineering is hard. The math is not new. What is missing is the will to build this before we need it, rather than after.

That is the arc I wanted to close. Whether Claude is thinking is a question the current J-lens has made partially answerable. Whether the substrate matters is a question that classical scaling will probably decide before quantum hardware gets a serious vote. And whether we will be able to see inside the hybrid systems that eventually arrive, if they do, is a question we are still early enough to answer well.

We should get on with it.

---

Sanjay Basu writes A Technocrat's Discernment. He is Senior Director of GPU and GenAI Solutions at Oracle Cloud Infrastructure and founder of Cloud Floaters Inc. All views are his own.