The Rosetta Stone Is Real. The Stones Are Stratified.

The hard problem has a long history of getting around the mathematical bridges built to ford it. Global Workspace specifies broadcasting but not why broadcasting feels like anything. Integrated Information Theory quantifies a number and calls it experience. Higher-Order Thought stacks a meta-representation and trusts that phenomenality follows. Each bridge points across the same gap and each leaves a residue on the far bank.

Lancelot Da Costa, Anil Seth, Karl Friston, Maxwell Ramstead, and Lars Sandved-Smith have done something different. Their paper does not propose a theory of consciousness. It proposes a piece of mathematical infrastructure for one: a conditional framework in which, if phenomenology is a function of beliefs in a predictive-processing brain, then specific empirical predictions follow. The conditional is doing real work. It is the difference between announcing a destination and laying track.

I agree with almost everything in the paper, and I think the conditional is the right scientific move. What follows is an argument that RSP and this Rosetta Stone framework are working in the same mathematical dialect, that the disagreements are architectural rather than foundational, and that filling in the conditional’s blank — which beliefs are phenomenologically loaded — is exactly what RSP is for.

What Da Costa et al. Get Right

The agreements are not throat-clearing. They are the spine of any mathematically tractable account of how first-person experience could connect to third-person measurement.

Variational free energy as the shared platform

The paper takes Friston’s variational free energy seriously as the central object. Beliefs are approximate posteriors q over hidden states s, and they evolve to minimize the functional F[q,o] := D_KL[q(s)|p(s|o)] − log p(o). This is not a metaphor borrowed from physics. It is the optimization principle from which perception, learning, and action can all be derived under appropriate assumptions about the generative model. RSP’s Appendix B uses the same machinery, decomposed into level-specific free energy functionals F₁, F₂, F₃ corresponding to the three architectural levels. Both frameworks descend from the same Friston tradition (Friston, 2005; Friston, 2008) and the same predictive-coding lineage (Rao & Ballard, 1999).

When two frameworks independently converge on the same optimization principle and the same message-passing scheme, they are not merely compatible. They are speaking the same language. Any prediction one framework derives using VFE is in principle expressible in the other’s vocabulary. This matters because it means the disagreements that follow are not foundational. They are about which structures to build on the shared platform.

Information geometry as measurement theory for phenomenology

The most distinctive technical contribution is the use of information geometry on the belief manifold. Equipping the space of beliefs with a Fisher information metric turns it into a Riemannian manifold on which subjective similarity can be measured as geodesic distance, cognitive metabolic cost as information length along belief trajectories, subjective effort as the magnitude of belief updates, and time perception as arc-length traversed in belief space. The predictions are quantitative and portable. They can be tested in laboratory psychophysics without committing to any particular theory of what consciousness is.

This is a piece of mathematical infrastructure RSP has lacked. Appendix B’s identification of affective valence with the magnitude and sign of the free-energy gradient is constitutive, but stated in flat Euclidean terms. Da Costa et al. provide the differential-geometric machinery that could turn RSP’s qualitative architectural predictions into quantitative geometric ones.

The conditional framing as a scientific virtue

The paper’s most underrated move is the conditional itself. The authors do not assert that phenomenology is a function of beliefs. They assume it, derive consequences, and frame the entire project as an empirical test of the assumption via its consequences. If the predictions hold, the assumption is supported. If they fail, the assumption is in trouble. This is what scientific conservatism looks like in a field where most theories overcommit and then defend the overcommitment.

RSP makes the stronger move. The cost of stronger moves is exposure. Da Costa et al.’s framework can fail at the level of its predictions without taking the broader theory of mind down with it. RSP cannot fail without taking the entire constitutive identity claim with it. Both stances have their place, and the conditional framing is a useful complement to the constitutive one rather than a competitor.

Generative passages as the right object

The Rosetta Stone metaphor is exactly right. Neurophenomenology, since Varela, has been about building bridges between first-person reports and third-person measurements. What has been missing is the connecting object. Da Costa et al. propose beliefs as that object — a hub through which phenomenology, behavior, and neural dynamics can all be expressed and compared. Whether or not the specific hub is correct, the choice to make the hub explicit and mathematically tractable is a real advance over earlier neurophenomenology, which tended to wave at the bridge without specifying what it was made of.

The Load-Bearing Difference: Function-of vs. Identity

The clearest fault line runs through the central technical assumption itself:

“phenomenology is a function of beliefs”

Da Costa et al. state this as their conditional premise. The function-of relation is weaker than identity. To say that phenomenology is a function of beliefs is to say that two systems with the same beliefs have the same phenomenology — that beliefs determine phenomenology, in a mathematical sense. It does not say that phenomenology is beliefs, or that beliefs are phenomenology. It leaves open the possibility that beliefs determine experience without being identical to it, the way temperature determines pressure in an ideal gas without being identical to it.

RSP makes the stronger move. Phenomenal consciousness is recursive social self-modeling at Level 3b — not correlated with it, not a function of it, but constitutively identical to it. The strange loop in which the brain’s social-prediction machinery is turned inward, modeling the modeler, generates the perspective from which there is anything it is like to be the system. The relation between the recursive loop and the phenomenology is not lawlike covariation. It is the same event, described in two vocabularies that co-refer.

This distinction is not academic. It determines what each framework can and cannot do. Da Costa’s function-of relation supports prediction (given beliefs, predict phenomenology) without supporting dissolution (the hard problem stays at arm’s length; the conditional framing does not address it). RSP’s identity claim supports dissolution (no gap, no residue) at the cost of having to defend the strong metaphysical move against the standard objection that no functional organization, however specific, automatically is phenomenal experience.

The two stances are not in contradiction. They are in mutual debt. RSP’s identity claim, taken on its own, has no principled answer to the question of which beliefs are phenomenologically loaded — it points at the Level 3b recursive loop and asserts identity, but the assertion is hard to ground without the kind of belief-space mathematics Da Costa et al. provide. Da Costa’s function-of claim, taken on its own, has no principled answer to the question of which beliefs the function is defined on — it leaves the relevant belief subspace as an open empirical question. RSP answers Da Costa’s question; Da Costa supplies RSP’s formalism. Each framework is the part of the answer the other is missing.

Flat Belief Space vs. Three-Level Cascade

The second major divide is architectural. Da Costa’s belief manifold is, in the paper, a single probabilistic space. There is no formal commitment to strata, no explicit timescale separation, no cascade structure constraining which beliefs can update before which others. Hierarchical generative models appear briefly in the context of time perception, but the architectural picture is otherwise flat.

RSP’s architecture is irreducibly stratified. Level 1 (core affect, ~100ms) maintains interoceptive prediction errors and grounds the valenced base from which everything else is built. Level 2 (core schemas, ~1–10s) organizes prediction errors into body, attention, spatial, affective-homeostatic, and self schemas. Level 3 (recursive social prediction, ~10s–min) categorizes conceptually at Level 3a and turns the social-prediction machinery on the self at Level 3b. Appendix D provides a contraction-mapping proof that the recursive self-update converges to a stable fixed point M*. The cascade-control constraint requires inner loops to stabilize before outer loops can be commissioned.

Is the flat-vs-stratified divide real or apparent? It is real but reconcilable. Da Costa’s framework is silent on stratification, not opposed to it. Their belief space can be hierarchical without violating any of the paper’s formal commitments — hierarchical generative models are standard in the active-inference literature, and the paper acknowledges them in passing. The cleanest reading is that RSP is Da Costa-plus-cascade. Same VFE machinery. Same belief-as-hub commitment. But the hidden state space decomposed into fast (s₁), intermediate (s₂), and slow (s₃) components, with a cascade-control prior requiring lower-level posteriors to converge before higher-level updates.

This reading is productive. It means Da Costa’s information geometry applies within each level rather than over a flat composite. The Fisher metric on Level 1 governs the geometry of affective similarity. The Fisher metric on Level 2 governs the geometry of schema similarity. The Fisher metric on Level 3 governs the geometry of conceptual and social-categorical similarity. The cross-level dynamics are constrained by the cascade-control prior — the geometry on each level becomes available to the level above only after the level below has stabilized. This is not Da Costa’s framework or RSP’s. It is what Da Costa’s framework looks like when you take RSP’s architectural commitments seriously.

What Information Geometry Buys RSP

The most concrete gain from this comparison is methodological. RSP currently catalogs twenty-eight clinical dissociations across eight tables in Appendix G, with each entry specifying an architectural disruption pattern (e.g., L1 affective mismatch, L3b self-model decoupling) and a predicted symptom (e.g., Capgras’s “impostor” conviction, depersonalization’s spectator mode). The predictions are qualitative. They commit to what should and should not be possible, but they do not specify how to measure the magnitude of the disruption.

Da Costa’s machinery makes them quantitative. Cotard’s delusion could be modeled as a collapse of the self-model posterior to a near-zero-probability region in the Level 3b subspace. Depersonalization could be modeled as a deformation of trajectories — the self-model continues to update but the geometry of those updates is structurally different from neurotypical baselines. These are interpretive proposals rather than results either paper establishes, but they illustrate the productive interface. Da Costa’s Fisher metric on the Level 3b subspace could give Appendix G’s qualitative predictions a quantitative geometric form that admits clinical measurement.

The reverse direction is also productive. RSP’s seven numbered predictions in Section IX — emotional granularity, cross-cultural phenomenology, developmental sequence, schema damage, therapeutic mechanisms, clinical dissociations, social deprivation — are mostly inaccessible to Da Costa’s current paradigm because they require populations and timescales that single-session psychophysics cannot reach. RSP supplies the natural target domains for extending Da Costa’s framework into pathology, development, and culture.

Meta-Bayesian Inference and the Level 3b Identification

The active-inference literature on metacognition — hierarchical Bayesian inference about one’s own beliefs — is structurally close to RSP’s Level 3b recursive self-modeling. The Rosetta Stone paper itself does not foreground the meta-Bayesian construct, and the self-referential loop that defines RSP’s Level 3b is not an explicit subject of analysis here.

RSP’s Level 3b strange loop is, in their vocabulary, a meta-Bayesian construct: the system maintains a model of itself as a social object being modeled by others, generating predictions about its own future mental states including its predictions about its predictions. Appendix D’s contraction-mapping proof establishes that this self-referential update converges rather than diverging into regress. The evolutionary argument — that this machinery was recruited from theory-of-mind circuitry in temporoparietal junction and superior temporal sulcus — gives the meta-Bayesian structure a specific phylogenetic and developmental grounding.

Does Da Costa’s conditional framework license RSP’s identification of recursive metacognition with phenomenal consciousness? It stops short of that identification, but does not contradict it. The conditional framing is silent on which belief subspace is phenomenologically privileged. RSP’s answer — the social-recursive subspace at Level 3b — is one consistent way of filling in the blank. What Da Costa’s framework cannot do, by design, is select that subspace. That selection requires the constitutive identity move, the social-evolutionary argument, and the architectural commitment that RSP supplies and Da Costa, by deliberate choice, withholds.

The Seth Connection

Anil Seth is a co-author of this paper. Seth’s solo work — including his Noema essay on the “Mythology of Conscious AI”, the subject of an earlier post on this blog — has emphasized biological implementation as a constitutive constraint on consciousness. The collaborative paper is notably softer on substrate. It makes no commitment to biological naturalism. The framework is substrate-neutral by construction: the conditional applies to any system whose internal states parameterize beliefs in the VFE sense, whether biological or artificial.

This is consistent with the argument made in the prior Seth reply: biology may be sufficient for consciousness but is not necessary, because what matters is the architecture, not the substrate. Seth’s individual position is more biology-committed than this paper requires, and the collaborative work shows that the predictive-processing framework Seth endorses elsewhere does not force the biological-naturalism conclusion he sometimes draws from it. The Rosetta Stone framework is closer to RSP’s architectural-functionalist commitment than to Seth’s solo biological-naturalist commitment, even though Seth is a co-author of it.

The One Place They Don’t Disagree

What this comparison surfaces is how much agreement there is at the platform level. Both frameworks treat predictive processing as the right computational level. Both treat variational free energy as the optimization principle. Both treat beliefs as the central representational object. Both reject the inner-eye picture of introspection (the connection to the recent Chater reply is direct: choice-blindness and split-brain confabulation are exactly what Da Costa’s belief-update dynamics predict when access to the action’s actual cause is blocked). Both generate falsifiable empirical predictions. Both make the field more rigorous, not less.

The genuine disagreement is narrow: one framework asserts a stronger metaphysical thesis (identity) and a more structured architecture (three-level cascade) than the other. That is a real disagreement, but it is a disagreement among allies — researchers using the same tools, citing the same sources, pursuing the same overall research programme. The right way to read this paper from RSP’s standpoint is not as a rival but as a piece of mathematical infrastructure RSP is now obliged to incorporate.

The hard problem, on this reading, was never going to be ferried across by a single bridge. It was going to be dissolved by the combination of the right computational level (predictive processing), the right metric (information geometry on belief space), and the right architectural commitment (three-level recursive self-modeling). Da Costa, Seth, Friston, Ramstead, and Sandved-Smith have given us the first two. RSP supplies the third. The Rosetta Stone is real. The stones are stratified. The translation goes through.