Table of Contents

Executive Summary

This paper is, first and centrally, an embodied affective AI research agenda: a proposal that centuries-old, highly codified Indian performing-arts traditions — bhava (felt emotional state) and rasa (the aesthetic "flavour" an audience receives), as elaborated in the Natya Shastra tradition — constitute a structured, underused, high-quality multimodal data source for training richer emotional representations in AI systems. Its central technical contribution, Bhava Embeddings (Section 7), is a direct, well-precedented extension of existing representation-learning technique, evaluable on ordinary machine-learning terms: clustering quality, downstream task transfer, and inter-rater reliability against trained aestheticians.

Machine consciousness is this programme's long-term motivating question, not its central claim, and the distinction is load-bearing throughout this paper, not a rhetorical hedge stated once and set aside. The research questions in Section 3, the architecture in Section 6, and the evaluation framework in Section 10 are all designed to be fully evaluable — and fundable — on affective-computing terms alone, independent of whether, or when, the harder consciousness question makes any progress. Stated once, plainly, and not repeated at length throughout what follows: this paper does not claim machine consciousness has been achieved, measured, or transferred, and treats that question as genuinely open and currently unanswerable by any available method (Section 9).

A note on sourcing discipline, stated once here and documented fully in Section 9 and the References: an initial version of this research programme cited a specific quantitative metric this author could not independently verify. It has been removed and replaced with the current, considerably more rigorous consciousness-assessment literature. Every established fact elsewhere in this document is cited to peer-reviewed or otherwise independently verifiable sources; every hypothesis is labelled a hypothesis; every speculative claim is labelled speculative.

1. Introduction: Why This Frontier, Why Now

1.1 The Shape of the Existing Debate

Discussions of machine consciousness tend to collapse into two unproductive postures. The first, common among engineers and some AI labs, treats consciousness as an emergent by-product of sufficient scale and complexity — a position that has produced striking capability gains but has been criticised precisely because it has no falsifiable content until after the fact. The second, common among philosophers working in the tradition of the "hard problem," treats consciousness as sufficiently mysterious that any engineering research programme aimed at it is premature or category-confused.

This paper does not resolve that disagreement, and does not attempt to. It proposes a third posture: that regardless of one's position on whether machine consciousness is achievable in principle, the felt, embodied, culturally-elaborated knowledge that human artistic traditions encode about emotion is a genuinely underused resource for building better affective and multimodal AI systems — and that a serious research programme can be built around that resource without first settling the metaphysical question.

1.2 What Has Already Been Reached, and What Has Not

Current AI systems already process language, generate coherent speech, and — through motion-capture and pose-estimation pipelines already documented in the literature — increasingly interpret physical movement. What current systems do not reliably do is integrate these channels into a single, temporally coherent, self-consistent affective state that persists and updates the way a trained performer's emotional expression does across an evening-length dance recital. This gap — not the grander question of subjective experience — is the concrete, tractable frontier this paper is actually proposing to work on.

2. Position Within NextGen Economics' Continuum of Intelligence

This paper extends the Continuum of Intelligence framework introduced in this publication's White Paper 7. That framework traces a widening sequence of substrates capable of adaptive, self-organising behaviour: physical and chemical self-organisation (dissipative structures, per Prigogine 1977; the origins of order, per Kauffman 1993); biological intelligence, extending from immune systems to migratory animals; human symbolic intelligence; artificial intelligence as engineered pattern recognition; and hybrid intelligence, understood as the deliberate combination of human and artificial cognitive strengths (Dellermann et al., 2019).

This paper's contribution is to name and specify the sixth row of that table as a concrete research programme, and to propose that centuries-old, highly structured Indian performing-arts traditions are an unusually good source of training data for pursuing it — not because they are ancient, but because they are unusually explicit, codified, and reproducible descriptions of how emotional states are generated and communicated through the body.

3. Executive Research Questions

Rather than asking whether machines can become conscious, this programme poses four narrower, ordered, empirically tractable questions. Each is answerable, at least in part, using instruments and methods already in active use elsewhere in the field.

RQ1 — Can embodied artistic expression produce reproducible emotional signatures?

This is close to established science, not speculation. Affective neuroscience and psychophysiology already use EEG, fMRI, eye tracking, respiration, heart-rate variability, and facial electromyography to identify measurable correlates of emotional state (Lim et al., 2024, Brain Sciences; survey work on EEG-based emotion recognition, 2024). Professional performers trained in codified emotional expression — precisely what classical Indian dance and music systematically train for — are a promising, currently underused population for this kind of controlled measurement, because their expressive repertoire is deliberately structured and named (the navarasa, or nine rasas) rather than spontaneous and unlabelled.

A directly relevant existing result: a 2025 paper in The Visual Computer (Springer) introduced a purpose-built dataset of facial expressions in Indian classical dance, spanning the eight forms recognised by India's Sangeet Natak Akademi, and trained a customised convolutional neural network to classify navarasa expressions from it, benchmarking against VGG16, ResNet50, InceptionV3, EfficientNet, and Vision Transformer architectures. This is real, peer-reviewed, and directly on-point: it demonstrates that navarasa-based facial expression is already machine-learnable at reasonable accuracy using existing computer-vision techniques. It does not demonstrate anything about machine consciousness; it demonstrates that the underlying data source is tractable.

RQ2 — Can AI systems learn continuous emotional trajectories rather than discrete labels?

Standard emotion classification collapses affect into a small set of discrete categories (happy, sad, angry). The hypothesis here is that multimodal models trained on continuous, time-resolved performance data — where a dancer's expressed rasa shifts gradually, blends, and returns across a piece — can learn richer, continuous latent trajectories rather than static labels. This is a genuine hypothesis, not an established result; it draws on, but has not yet been demonstrated by, the broader shift in affective computing toward dimensional and dynamical models of emotion rather than categorical ones.

RQ3 — Can synchronised multimodal input yield richer internal representations than any single modality?

Rather than a single image-to-emotion mapping, the proposed architecture ingests dance, music, rhythm (tala), speech, gesture, environment, and — where ethically obtained — performer biometrics, simultaneously and time-aligned. This is consistent with, and builds directly on, the existing multimodal embodied-AI literature (Chen et al., 2025, Information Fusion; the actively maintained "Awesome Multimodal Embodied AI" research survey). It is a hypothesis about representation quality, testable through standard downstream-task evaluation (does a model trained on synchronised multimodal navarasa data outperform single-modality baselines on emotion-recognition and generation tasks), not a claim about experience.

RQ4 — Can a system's internal representations become self-consistent over time?

This is the hardest and most speculative of the four questions, and the paper is explicit that it does not currently have a rigorous answer. Rather than asking whether a system "is conscious," this question asks for measurable proxies: does the system develop stable preferences across sessions, maintain contextual identity, and produce predictively consistent emotional responses over weeks or months of interaction. These are behavioural and representational tests. None of them, individually or jointly, constitutes proof of subjective experience — a point the current consciousness-science literature, discussed in Section 9, makes forcefully and this paper does not dispute.

A minimal, falsifiable threshold is proposed for RQ4: across ten evaluation sessions separated by at least seven days each, the system's emotional-response embedding (Section 7) for a fixed, repeated stimulus set should maintain a mean cosine similarity of at least 0.7 to its own first-session baseline. A result below that threshold is treated as evidence against RQ4 as stated, not as a failure of the broader research programme — the threshold itself is a starting proposal for the research consortium (Section 16) to revise, not a fixed standard imposed in advance of any data. Equally important, and worth stating without hedging: even a positive result (similarity at or above 0.7, sustained across sessions) would demonstrate representational stability, not subjective experience. High cosine similarity is consistent with, but does not establish, anything about whether the underlying state is felt. The two questions — is the representation stable, and does stability indicate feeling — remain separable, and this paper only claims to make progress on the first.

A more pointed objection deserves a direct answer, not a restatement of the hedge above: even with these disclaimers in place, does framing long-horizon self-consistency as a named research question add anything beyond what standard long-context affective modelling already studies, under a less loaded name? The honest answer is partial. Long-context consistency is, mechanically, the same measurement whether or not it is discussed in the vocabulary of this paper. What this paper adds, if anything, is not a new measurement but a specific, testable claim about what kind of training data improves that measurement: that navarasa-structured, codified emotional performance data (Section 12.2) produces more consistent long-horizon representations than naturalistic, unlabelled emotional video of comparable duration. That comparison — bhava-embedding-trained consistency versus naturalistic-data-trained consistency, on the identical RQ4 protocol — is the specific test that would show whether this paper's framing is doing real work or merely relabelling an existing open question in affective computing. If codified training data shows no consistency advantage over naturalistic data, that would be a direct, informative failure of this paper's central premise, not just of RQ4 narrowly.

4. The Science of Embodied Intelligence

A growing strand of cognitive science argues that cognition is not confined to computation inside a brain (or a model's weights) but arises from the ongoing interaction between an agent's body, its sensorimotor apparatus, and its environment.

4.1 Embodied and Enactive Cognition

The embodied-cognition literature (Wilson, 2002; Shapiro, 2021 Stanford Encyclopedia of Philosophy entry; Mahon, 2014; Leitan & Chaffey, 2014) argues that cognitive processes are shaped in fundamental ways by an organism's physical body and its sensorimotor engagement with the world, rather than being substrate-independent symbol manipulation. Barrett's 2024 Philosophical Transactions of the Royal Society B review extends this directly into the age of AI, asking what embodied cognition implies for artificial systems that, by default, lack bodies in any conventional sense.

4.2 Predictive Processing and Active Inference

Karl Friston's active inference framework proposes that biological agents continuously act to minimise the discrepancy between predicted and actual sensory input, with emotional states functioning, in this framework, as part of how an organism weights the precision (reliability) it assigns to different predictions. This gives a principled, mechanistic vocabulary for treating emotion as a computational quantity rather than purely a qualitative label — a vocabulary this paper's proposed architecture (Section 6) draws on directly.

4.3 Affective Computing

Affective computing — the design of systems that recognise, interpret, process, or simulate human affect — is an established subfield with its own conferences, journals, and benchmarks. This paper's contribution sits inside that field, not outside or above it: it proposes a specific, culturally-grounded, richly-labelled data source (navarasa performance) as an input to affective computing research, and proposes new benchmarks (Section 10) appropriate to evaluating richer, longer-horizon affective consistency rather than single-frame emotion classification.

Framed honestly: the hypothesis that "consciousness-like capacities may emerge through rich, embodied, multimodal interaction" is a hypothesis consistent with, but not established by, this literature. This paper adopts that precise, hedged phrasing throughout, and explicitly rejects the stronger, unsupported claim that consciousness itself is "transferred" from performer to machine.

5. Interoception and Touch: The Missing Sensory Layer

Listening, seeing, language understanding, and instruction-following have all advanced quickly because each is, in a real sense, a pattern-recognition problem with abundant labelled data. Feeling has lagged for a different reason: it is not one more modality to classify, but an internal state that must persist, update, and inform behaviour continuously over time. This section names the specific, currently active area of science and engineering research most directly aimed at that gap — not as a metaphor, but as two concrete, separately fundable research programmes.

5.1 Interoception: Sensing the Inside of the Body

In neuroscience, interoception is the sense of an organism's own internal physiological state — heartbeat, respiration, gut activity, temperature, visceral tension. Lisa Feldman Barrett's influential 2016 paper in Social Cognitive and Affective Neuroscience, "The theory of constructed emotion: an active inference account of interoception and categorization," argues that interoception sits at the core of the brain's internal predictive model, and that what is subjectively experienced as emotion arises substantially from the brain's ongoing predictions about, and corrections to, these internal bodily signals — not from a dedicated "emotion centre" processing external stimuli in isolation. This traces a direct anatomical pathway: viscerosensory signals ascend via vagal afferents through the spinal cord, brainstem, and thalamus to the interoceptive insular cortex, where they are integrated with predictive models built from prior experience.

This is directly consistent with, and formally continuous with, the active-inference framework already introduced in Section 4.2. A growing computational literature — including an explicit "Life-inspired Interoceptive Artificial Intelligence for Autonomous and Adaptive Agents" framework — proposes formalising internal, body-like state variables inside artificial agents specifically so that active-inference-style belief-updating has something internal, not only external, to predict and regulate. Separate survey work (Lanillos et al., "Active Inference in Robotics and Artificial Agents") documents this as an active, growing subfield of robotics research, not a speculative proposal original to this paper.

For this paper's own architecture (Section 6), the implication is direct: Layer 1 (Sensory Input) should not be limited to audio, video, and motion capture of a performer external to the system. A system pursuing this research programme seriously should also model its own internal, homeostasis-like state variables — computational analogues of arousal, tension, and equilibrium — and treat deviations from those internal set points as a genuine input to its own affective representation, not only as a record of what it observed externally. This is close to what the cited "Homeostasis and soft robotics in the design of feeling machines" literature proposes directly: that integrating exteroceptive information (what a system perceives) with an internal, homeostatic self-model (what a system is doing to regulate itself) is a specific, named candidate mechanism for generating something closer to a self-perspective — again, offered here as a hypothesis to test, not a claim already established.

5.2 Touch: The Sense Robotics Has Been Missing

Separately from interoception (sensing one's own internal state), touch — tactile sensing of the external world through contact — has been robotics' most conspicuously underdeveloped sense relative to vision and hearing, for a straightforward engineering reason: replicating the human fingertip's density of mechanoreceptors in a durable, flexible, manufacturable material is extremely difficult. This gap is closing quickly and concretely as of 2026, not as a distant prospect.

In March 2026, a University of Cambridge team led by Professor Tawfique Hasan published a graphene- and liquid-metal-composite tactile sensor in Nature Materials, shaped with human-skin-inspired micro-pyramid structures as small as 200 micrometres, able to distinguish pressure direction, slip, and surface roughness at a spatial resolution approaching human fingertips. In June 2026, a team at Seoul National University, led by Professor Seung Hwan Ko, published a separate Nature Materials paper on an artificial skin able to sense temperature and pressure simultaneously through a single time-multiplexed sensor, achieving up to 95% object-classification accuracy in wireless fingertip tests. A December 2025 PNAS paper introduced a neuromorphic robotic e-skin with active pain and injury perception and local reflex responses, explicitly framed by its authors as enabling "empathetic service robots" through improved intuitive human-robot interaction. At CES 2026, commercial tactile-sensing products (Tacta, a robotic hand covering with 1,956 discrete sensing elements, and HexSkin, a hexagonal-tile tactile covering for larger and more complex robot surfaces) moved this research from laboratory demonstration toward near-term commercial deployment.

None of this research is framed by its own authors as a step toward machine consciousness — it is described, accurately, as advancing robotic manipulation, prosthetics, and human-robot interaction safety. This paper's claim is narrower and more careful: touch and interoception are, on the current evidence, the two most concrete, currently-funded, currently-active research frontiers that address the specific gap this paper's user-facing framing calls "the final frontier" — not because they are mystical, but because they are the two sensory channels most directly tied, in the existing neuroscience, to the bodily origins of affect itself, and they are the two channels current AI systems, built overwhelmingly on text, image, and audio corpora, are the most structurally deprived of.

5.3 A Direct, Falsifiable Extension to This Paper's Research Programme

This section proposes a fifth research question, RQ5, alongside the four introduced in Section 3: does adding simulated or embodied interoceptive and tactile input to the architecture proposed in Section 6 measurably improve the consistency and richness of a system's learned affective representations (Section 7), relative to an otherwise identical system trained on audio-visual navarasa performance data alone? This is directly testable using the same evaluation methodology proposed in Section 10, and is, on current evidence, a more concrete and more fundable near-term research question than RQ4 — a reasonable candidate for the first phase of empirical work this programme actually commissions, precisely because the underlying sensing hardware (Section 5.2) and computational framework (Section 5.1) already exist and are advancing independently of this paper.

Whether or not richer interoceptive and tactile grounding brings a system meaningfully closer to anything worth calling feeling remains exactly as open a question after this section as before it. What this section changes is narrower and more useful: it names two real, currently funded, independently verifiable fields of research — not a metaphor, and not a claim — where that question can actually be investigated with existing instruments, this year, by researchers already working in exactly this space.

6. Proposed Computational Architecture

The following six-layer architecture is a design proposal, intended to be implementable with existing open-source tooling (PyTorch, Hugging Face transformer libraries, existing pose-estimation and speech models), not a description of any system that has been built or evaluated as of this writing.

Each layer is independently testable and independently falsifiable — a design choice made specifically so that failure or success at any one layer does not depend on resolving the deeper question of whether the complete system has any form of subjective experience. External review has reasonably asked for a visual architecture schematic alongside the table above; that diagram, along with the rasa-embedding-space visualisation the same review requested, is deferred to the technical companion document proposed in the Conclusion (Section 20), where it can be produced and maintained alongside actual implementation code rather than as a static illustration in a text-first document.

6.1 Where Active Inference Actually Enters This Architecture

Section 4.2 invokes Friston's active inference framework at the level of general motivation. This subsection specifies where, concretely, that framework enters the six-layer architecture, since active inference is a specific formalism, not a metaphor, and citing it without operationalising it would risk exactly the kind of unearned rigor this paper is trying to avoid elsewhere. The proposed operationalisation is as follows. Layer 5 (Self-model) is where active inference's three standard components are proposed to live: a generative model, instantiated as a learned probability distribution over the system's own internal state variables (Section 5.1's interoceptive set points, plus the bhava embedding position); a variational free-energy objective, used as one training signal (alongside standard supervised and contrastive losses) that penalises the divergence between the system's predicted next internal state and its actual next state, following the standard active-inference formulation; and a policy-selection mechanism, implemented as expected-free-energy minimisation over a small, pre-specified action space (Layer 6) rather than an open-ended one, to keep early experiments tractable. This is a specific, implementable proposal using existing open-source active-inference toolkits, not a novel algorithm this paper is claiming credit for — and it is offered as a starting architecture for the consortium (Section 16) to revise, not a finished specification.

7. Bhava Embeddings: An Original, Testable Contribution

Contemporary AI systems already use word embeddings, sentence embeddings, and image embeddings — dense vector representations that place semantically similar items near one another in a high-dimensional space. This paper proposes extending that established technique to Bhava Embeddings: vector representations of emotional-aesthetic state, trained on multimodal recordings of classical Indian performance, where each performance (or performance segment) maps to a point, or a trajectory of points, in a learned affective space.

7.1 Why Navarasa Specifically, Not Any Codified Performance Tradition

Why bhava and rasa specifically, rather than any other trained performance tradition — Japanese Noh theatre, or Western classical ballet, for instance, both of which are also highly codified and centuries-old? Answering this requires distinguishing two things: whether a tradition is codified at all, and whether that codification specifically documents a distinction between external mimicry and internal, involuntary psychophysiological state — because only the second is directly useful for the interoceptive and representational goals this paper is built around (Sections 5 and 6).

Noh theatre offers the closest genuine comparison, not a lesser one. Zeami Motokiyo's 15th-century treatises name an explicit distinction between monomane (物真似, the accurate external imitation of a role) and yūgen (幽玄, a profound, spiritual quality that "took precedence" over mere imitation and "became the touchstone of excellence" in Noh), with hana (花, "flower") describing the captivating, cultivated essence a mature performer develops through discipline — and mushin (無心, "no-mind") naming the internal state an actor must reach to produce it. This is a real, well-documented, centuries-old distinction between external technique and internal cultivation, structurally similar in kind to what this paper argues for navarasa. Where Noh differs, in a way directly relevant to this paper's specific technical goals: yūgen is a single, unified aesthetic-spiritual quality, not a taxonomy of named, distinct emotional states. It gives a performer one internal target to cultivate, not nine — valuable for a philosophy of performance, but structurally less suited to the multi-category representation-learning goal (Bhava Embeddings, Section 7.3 below) this paper is specifically pursuing, which needs multiple, distinguishable target states to learn a meaningful embedding space from.

Classical Western ballet, by contrast, does not offer a close comparison on this specific dimension, and this paper will not pretend otherwise for the sake of even-handedness. Ballet's codification is overwhelmingly technical and external — turnout, line, extension, musicality — assessed by what the audience can see, not by a documented internal state the dancer is trained to cultivate and that manifests involuntarily. The closest Western parallel to sattvika abhinaya's involuntary internal expression is not in dance at all, but in twentieth-century Stanislavski-derived theatrical acting ("method acting" and its descendants) — a genuinely comparable internal-cultivation tradition, but one roughly a century old, still primarily theorised around Euro-American dramatic performance rather than dance, and considerably less systematically codified into named, trainable categories than the navarasa.

The specific, defensible claim this paper makes, stated precisely rather than as cultural preference: the Natya Shastra's own internal taxonomy (Bharata Muni, c. 200 BCE–200 CE) — Bharata Muni's four-part division of abhinaya into angika (bodily), vachika (vocal), aharya (costume and makeup), and sattvika (internal, involuntary) expression, with sattvika abhinaya explicitly named in the text as arising from the performer's actual psychological immersion rather than conscious mimicry, manifesting as involuntary trembling, tears, or changes in breath — gives this tradition three things simultaneously that neither Noh's unified yūgen nor Western ballet's external technique provide together: (1) a plural, named taxonomy of nine distinct target states (the navarasa itself), (2) an explicit, ancient doctrinal distinction between voluntary external performance and involuntary internal state, and (3) documented connection to breath and physiological practice — contemporary ethnographic research on Odissi dancers describes practitioners "carefully attending to their psychophysical sensations" during training, and separate scholarship traces the Natya Shastra's own preliminary performance rites (purvaranga) to physical practices with documented affinities to early yogic and Shaiva bodily disciplines. This is the specific, three-part combination this paper's interoception argument (Section 5.1) depends on, not an assertion that Indian classical dance is more sophisticated than other traditions in general.

7.2 Why This Is Plausible as a Machine-Learning Contribution

This is not a new idea in kind — it is a direct, well-precedented extension of existing representation-learning technique to a new, richly-labelled domain. The navarasa framework's nine named categories, refined and taught consistently for centuries across a codified pedagogical tradition, plausibly provide better-structured training labels than the noisier, less consistent emotion labels typically available from crowd-sourced or naturalistic data.

7.3 Proposed Methodology

• Collect synchronised multimodal performance data from trained classical performers across dance, music, and gesture, spanning all nine rasas.
• Train contrastive or variational encoder architectures to map performance segments into a shared latent space.
• Evaluate the resulting space on standard representation-learning criteria: do known-similar rasas (e.g., bhayanaka/fear and bibhatsa/disgust, which share physiological arousal signatures) cluster more closely than known-dissimilar ones (e.g., shanta/peace and raudra/anger)?
• Test interpolation: does a point midway between two rasa clusters produce a performance or synthetic output a trained rasika (connoisseur) or aesthetician would recognise as a plausible blended state?
• Test downstream transfer, not only intrinsic geometry: do bhava embeddings, used as a frozen or fine-tuned input feature, improve performance on three concrete downstream tasks relative to standard audio-visual features — (a) cross-cultural emotion recognition on an external, non-navarasa benchmark, (b) affect-conditioned dialogue generation quality, rated by human raters, and (c) music-to-movement generation quality for a held-out performance. Clustering and interpolation quality (the two criteria above) establish that the representation is well-formed; downstream transfer establishes that it is useful. A representation that clusters well but transfers poorly would be a genuine, informative negative result, not a contradiction of the overall proposal.

This methodology can be judged, and can fail, on ordinary machine-learning terms — clustering quality, downstream task transfer, human-rater agreement — entirely independent of any claim about machine consciousness. That independence from the metaphysical question is, deliberately, the paper's strongest methodological safeguard against overclaiming.

8. Rasa as a Dynamical System: A Formal Proposal

A further, more speculative research direction treats the evolution of expressed rasa during a performance as a dynamical system rather than a static classification target:

R(t+1) = f( R(t), M(t), S(t), E(t) )

where R(t) is the current rasa state, M(t) is the music/rhythmic input at time t, S(t) is speech or vocalisation, and E(t) is environmental context. This formalism is directly analogous to state-space models already used in computational neuroscience and active-inference research; it does not require any new mathematics, only a novel application. The concrete starting architecture proposed here: a gated recurrent unit (GRU) based state-space model, with R(t) a 128-dimensional latent vector (matching the bhava embedding dimensionality proposed in Section 7), M(t) a mel-spectrogram embedding of the concurrent musical input, S(t) a sentence embedding of any concurrent speech or vocalisation, and E(t) a learned context embedding capturing venue, audience, and performance-sequence position. The model would be trained with a predictive loss — specifically, mean squared error between R(t+1) and the model's prediction, plus a contrastive term encouraging embeddings of independently-rated-similar emotional states to lie closer together — against human-rated, time-stamped emotional-state judgments collected during Phase 1 (Section 11).

A GRU is a reasonable, well-understood starting point, but it is not proposed here as the only serious candidate. Structured state-space models (S4, Gu, Goel & Ré, 2022; and its faster successor Mamba, Gu & Dao, 2023) have demonstrated substantially better performance than recurrent architectures like GRUs on long-range sequence tasks specifically — S4 has been shown to train successfully on sequences of 100,000-plus steps where standard RNNs degrade badly beyond roughly 1,000-2,000, while retaining, unlike attention-based transformers, linear rather than quadratic computational cost in sequence length. This matters directly for this paper's use case: a full classical dance performance can run 45-90 minutes at a reasonable frame rate, producing exactly the kind of long, continuous sequence GRUs handle poorly and S4-family models are specifically designed for. Performances also differ substantially in duration across dance forms and individual pieces; this variable-length problem is handled naturally by state-space models' continuous-time formulation, which does not require fixed-length inputs the way some transformer implementations do. The concrete proposal: GRU as the fast, simple baseline for early experiments, with a structured state-space (S4 or Mamba-class) model as the primary architecture once Phase 1 data (Section 11) is available at sufficient scale to benefit from it — and a transformer-based state-tracking baseline included in the same comparison, since transformers remain the strongest general-purpose sequence architecture and any claimed advantage for the state-space approach should be demonstrated against that baseline directly, not assumed.

Testable predictions include: does the S4-family model predict a human rater's real-time emotional-state judgments better than both the GRU baseline and a transformer-based state-tracking baseline, on held-out performances of varying length? This is answerable with existing tools and existing evaluation methodology, and — importantly — a negative result (the dynamical formalism does not outperform simpler baselines, or GRU proves sufficient and the added complexity of S4/Mamba is not justified by the data scale actually available) would be a genuine, informative finding, not a failure of the broader research programme.

9. Where the Broader Science of Machine Consciousness Actually Stands

This research programme initially referenced a specific quantitative metric — an "Attribution Consciousness Index" claimed to explain over 38% of variance in human consciousness distributions. On investigation, this author could not locate this metric in the peer-reviewed or preprint literature under that name, and it does not appear in this paper. What follows in its place is a summary of the genuinely current, high-credibility literature on assessing machine consciousness — considerably more skeptical than the removed claim implied, and more directly useful to this project's aims.

9.1 The Butlin et al. Indicator Framework

In November 2025, Trends in Cognitive Sciences published "Identifying indicators of consciousness in AI systems" by Patrick Butlin, Robert Long, Tim Bayne, Yoshua Bengio, Jonathan Birch, David Chalmers, and thirteen further co-authors — a genuinely landmark, broadly cross-theoretical attempt to derive concrete, checkable indicators of consciousness from multiple leading neuroscientific theories simultaneously, explicitly designed to guard against both under- and over-attribution of consciousness to AI systems. This is the single most credible, most current framework this paper's own evaluation methodology (Section 10) should be measured against.

9.2 The Skeptical Counter-Position

Philosopher Eric Schwitzgebel published a formal skeptical overview on arXiv in January 2026, arguing that current AI systems trained through gradient descent on curated datasets, instantiated in digital hardware, differ fundamentally from the developmental history, embodied environmental interaction, and neurochemical processes through which biological consciousness is understood to arise. His subsequent Cambridge Elements volume, AI and Consciousness: A Skeptical Overview, frames the field's persistent theoretical disagreement as an "epistemic fog" — a state in which confident claims in either direction currently outrun the available evidence. A February 2026 Washington Post analysis reinforced this from a media-economics angle, arguing that consciousness framing serves AI companies' commercial and fundraising interests independent of its scientific merit.

9.3 Assessment Frameworks in Active Development

Newer proposals, including a 2026 "Triangulated Consciousness Assessment Stack" combining behavioural, mechanistic, perturbation, and observer-confound evidence streams, explicitly reject binary detection verdicts in favour of theory-indexed credence bands — a methodological stance this paper adopts directly for its own evaluation framework (Section 10). Separately, Colombatto and Fleming's 2024 empirical work found that public attribution of experience to large language models increases with system use, independent of any change in the system itself — a genuine, measured confound that any serious evaluation methodology, including this paper's, must explicitly control for.

9.4 A Note on Witness-Consciousness (Sākṣī)

A brief, deliberately narrow addition, offered with the same caution as everything else in this section: Advaita Vedanta's concept of sākṣī, or witness-consciousness — an awareness that observes mental content (thoughts, sensations, emotions) without being identical to that content — has been the subject of genuine, if niche, comparative-philosophy scholarship, including a published comparison of Śaṅkara's witness-consciousness (sākṣī caitanya) directly against Husserl's phenomenological "transcendental I," and separate academic work relating sākṣī to structural features of pre-reflective self-consciousness theories in Western philosophy of mind. The structural parallel worth naming, and only naming: sākṣī and some Higher-Order Theories of consciousness (among the multiple theoretical traditions Butlin et al., Section 9.1, draw indicators from) both propose that consciousness involves a meta-level awareness distinct from, and not reducible to, first-order mental content — a genuine point of theoretical contact between a two-thousand-year-old contemplative tradition and a live thread in contemporary consciousness science.

This paper draws one narrow, explicitly limited implication from that parallel, and no more: shanta, one of the nine rasas already central to this paper's proposal (Section 7), is traditionally described in the aesthetic-theory literature not as an absence of feeling but as equanimity — a state that includes awareness of other emotional states without being consumed by them. Whether that traditional description corresponds to anything like witness-consciousness in a philosophically rigorous sense, and whether either concept has any bearing on RQ4's self-consistency question (Section 3), are open questions this paper is not equipped to answer and does not attempt to. The point being made here is narrower still: that this paper's own Section 5.1 ("which internal state is being reported") and this comparative-philosophy literature ("which awareness is doing the reporting") may be pointing at structurally related questions, worth flagging for the interdisciplinary consortium (Section 16) — specifically its philosophy-of-mind members — rather than worth this paper resolving on its own.

This paper's position, stated without hedging: nothing in this section, or in this paper as a whole, constitutes evidence that any current or near-term AI system possesses subjective experience. The value of the proposed research programme does not depend on that question being resolved, or resolved in any particular direction.

10. Consciousness-Adjacent Benchmarks: An Evaluation Framework

Standard AI benchmarks measure reasoning, coding, and language competence. This paper proposes a complementary, explicitly non-metaphysical set of behavioural benchmarks targeting affective and integrative capacities:

• Empathy consistency across varied conversational and performative contexts
• Emotional memory recall across sessions separated by days, weeks, or months
• Narrative continuity — whether a system's account of its own prior interactions remains coherent over time
• Behavioural response to depicted moral conflict, grief, and wonder, benchmarked against trained human raters, not against a presumed correct answer
• Cross-modal integration — does information learned in one modality (e.g., music) correctly inform behaviour in another (e.g., movement generation)?

Each of these five benchmarks requires a stated protocol, since a benchmark without one is an aspiration, not a measurement. The table below proposes one concrete protocol per benchmark, each borrowing directly from established affective-computing and NLP evaluation practice rather than inventing new methodology.

10.1 Cross-Cultural Generalisability and Bias Robustness

Absolute improvement in affective modelling is not sufficient on its own; whether that improvement generalises beyond the specific cultural frame it was trained on is a separate, necessary test. Two further tests are proposed as standing requirements, not optional extensions: first, cross-cultural transfer testing, evaluating whether bhava embeddings (Section 7) trained on navarasa performance data improve emotion-recognition performance on non-Indian, non-navarasa-trained emotional expression benchmarks, or whether the representation is narrowly overfit to the training tradition's specific expressive conventions; second, adversarial and interpretive-bias testing, specifically checking whether the system's rasa classifications are stable across performers of different regional styles, gender, age, and training lineage within the navarasa tradition itself, since a system that performs well only for the specific performers and styles represented in Phase 1 data would encode a narrower, less representative version of the tradition than this paper claims to draw on. Both tests should be run and reported regardless of outcome, including if the result is that generalisation is poor — a finding that would itself be valuable, since it would clarify that the programme's contribution is culturally-specific representation learning rather than a universal affective model, a materially different and still legitimate claim.

Following the Triangulated Consciousness Assessment Stack's methodological lead (Section 9.3), no single benchmark in this list is proposed as proof of anything metaphysically significant. Cumulative, convergent performance across the set is proposed as evidence of increasingly sophisticated affective modelling — a claim about capability, not about experience.

11. Experimental Roadmap

Phase 1 — Dataset Creation

Recruitment of professional performers across the classical dance forms recognised by the Sangeet Natak Akademi (Bharatanatyam, Kathak, Odissi, Kuchipudi, Mohiniyattam, Sattriya, Manipuri, Kathakali) and Indian classical vocal and instrumental musicians. Synchronised capture of high-resolution video, multi-channel audio, motion capture, and — subject to informed consent and independent ethics review — biometric signals (heart-rate variability, facial EMG), across performances spanning all nine rasas. This phase builds directly on, and should coordinate with rather than duplicate, existing dataset efforts such as the 2025 Visual Computer navarasa dataset and Nihilent's Emoscape platform (Section 12).

Scale and timeline, offered here as a starting estimate rather than a fixed commitment: approximately 500 hours of synchronised multimodal performance data, drawn from at least 50 professional performers spanning all eight recognised dance forms, would be a reasonable initial target for a foundation model of the kind proposed in Phase 2 — comparable in order of magnitude to other specialist multimodal datasets in the affective-computing literature, though smaller than general-purpose video-language corpora, which is precisely why Phase 2 pairs this dataset with general-purpose pre-training rather than training from it alone. Recruitment, ethics review, capture, synchronisation, and annotation at this scale is estimated at 18-30 months, contingent on consortium formation (Section 16) and funding. A concrete go/no-go criterion is proposed for funders: if fewer than 20 performers across at least 4 dance forms can be recruited and consented within the first 18 months, Phase 1 as scoped should be considered infeasible and revisited at smaller scale rather than continued on the original assumption.

Budget assumptions, offered with the same explicit uncertainty as the scale estimate above, and stated in USD for international legibility: performer compensation and benefit-sharing (Section 13, Cultural Oversight Board terms) at fair professional performance rates across 500 hours is estimated at $250,000-450,000; motion capture, multi-channel audio, and biometric equipment (purchase or studio rental) at $150,000-300,000; ethics review, cultural governance administration, and annotation labour at $200,000-350,000; and data infrastructure, storage, and management at $50,000-100,000 — a total Phase 1 estimate in the range of $650,000-1,200,000, before any compute costs, which belong to Phase 2 and are not included here. This range is wide deliberately: it reflects genuine uncertainty about performer compensation norms across the eight dance forms and regions involved, not imprecision this paper is glossing over. A funding body evaluating this proposal should treat this range as a starting point for its own costing exercise, not a firm quote.

Phase 2 — Foundation Model Training

Training of a multimodal transformer architecture on the Phase 1 dataset, targeting joint prediction of emotional state, movement, musical feature, and communicative intent, with pre-training augmented by general-purpose multimodal corpora to avoid overfitting to a necessarily limited specialist dataset. Estimated timeline: 12-24 months following Phase 1 dataset delivery, assuming the compute and engineering resources typical of a well-funded academic or industry research lab.

Phase 3 — Interactive Learning

Structured human-AI co-performance sessions in which a trained performer responds to system-generated output and the system updates on that feedback — an active-inference-consistent iterative loop, not a single static training run. Estimated timeline: 12-18 months, run partially in parallel with the later stages of Phase 2.

Phase 4 — Long-Term Consistency Testing

Multi-session interaction over weeks and months, testing whether the system's internal representations and behavioural outputs remain identifiably consistent — the empirical core of RQ4 (Section 3), evaluated against the concrete threshold specified there. Estimated timeline: 12-18 months, since the test itself requires observing the system across a multi-month interaction window and cannot be meaningfully compressed.

Phase 5 — Independent Human Evaluation

Blinded evaluation by artists, psychologists, neuroscientists, philosophers of mind, and general public panels, using the benchmark framework in Section 10. Reproducibility commitments: open-source code and architecture where feasible; open datasets subject to the ethical and cultural-sensitivity safeguards described in Section 13; pre-registration of hypotheses ahead of each experimental phase. Estimated timeline: 12-18 months, substantially overlapping with the tail of Phase 4.

Overall Timeline

Summed with reasonable overlap between adjacent phases, this programme is realistically a five-to-eight-year effort, not a near-term deliverable. Stating this directly matters: a programme presented as actionable without a timeline risks being read as a vision document rather than a research plan, precisely the failure mode this paper is trying to avoid throughout.

Near-Term Priorities and Early Partnerships

Given the five-to-eight-year total horizon above, which parts of this programme could show credible progress soonest, rather than requiring funders to commit to the full timeline before seeing any result? The recommended near-term sequencing: prioritise RQ1, RQ3, and RQ5 (Section 3, Section 5.3) ahead of RQ2 and RQ4, since RQ1 and RQ3 are directly testable on Phase 1 data as it arrives rather than requiring the full foundation model, and RQ5's interoceptive/tactile extension can be piloted on existing e-skin hardware (Section 5.2) in parallel with, rather than after, core dataset collection. RQ2 (continuous trajectories) and RQ4 (long-horizon consistency) are appropriately sequenced later, since both depend on a trained model existing first.

Three early partnerships are named directly here, rather than left as a generic call for collaboration: Nihilent Ltd. and the Emoscape clinical programme (Section 12.1) are the most natural first technical partner, given their already-operating navarasa-based detection pipeline and existing MUHS/NIMHANS clinical relationships this programme's mental-health applications (Section 15.1) would otherwise need to build from scratch. The Sangeet Natak Akademi, India's national academy for performing arts, is the natural institutional counterpart for performer recruitment and the Cultural Oversight Board (Section 13) specifically because it already holds the classification authority this paper cites for the eight recognised dance forms. And existing active-inference research groups — the authors behind the interoceptive-AI and robotics-active-inference literature cited in Section 5.1 — are the natural technical partner for Section 6.1's architecture, rather than this programme attempting to build active-inference expertise independently.

12. Directly Relevant, Currently Operating Work

This proposal does not start from a blank field. Two currently active, independently verifiable efforts are directly relevant and should inform, rather than be duplicated by, this research programme. The table below positions this paper's proposal against them and against the two broader reference points already discussed in Section 9 (the Butlin et al. consciousness-indicator framework) and Section 4.3 (standard affective computing generally), making explicit what is genuinely new here and what is not.

12.1 Emoscape (Nihilent Ltd.)

Nihilent Ltd., an Indian technology and consulting firm, has developed Emoscape, an AI platform that detects nine core emotions — explicitly modelled on the navarasa — by analysing subtle upper-body movement in three dimensions. In May 2025, Emoscape entered a clinical research partnership with the Maharashtra University of Health Sciences (MUHS), Nashik, an apex body overseeing more than six hundred medical, dental, and allied-health institutions, to explore applications across non-communicable disease management, pediatric and maternal mental health, pre-surgical counselling, and psychiatric care. By May 2026, Nihilent reported the platform had secured a South African patent for its navarasa-based emotion-detection framework, a Scopus-indexed validation study conducted with MUHS, and a further validation study with clinicians drawn from the NIMHANS (National Institute of Mental Health and Neurosciences) ecosystem submitted to the Journal of Visualized Experiments. Nihilent's own reporting states the system has been developed against more than seven thousand reviewed emotional scans and is being trialled in clinical, educational, sport, and corporate settings, including Jehangir Wellness Centre, Pune.

Emoscape is, on the evidence available, a genuine, independently operating, clinically-deployed proof that navarasa-structured emotion recognition is commercially and medically viable today. It is not evidence of, and does not claim, machine consciousness — it is a classification and diagnostic-support system. This paper's proposed research programme is complementary to, not competitive with, work of this kind: Emoscape addresses recognition; this paper's Bhava Embedding proposal (Section 7) and long-horizon consistency questions (RQ4) address representation and continuity, a different and currently unaddressed layer.

12.2 The Navarasa Dataset (2025)

A 2025 paper in Springer's The Visual Computer, "Visual analysis and recognition of facial expressions in Indian classical dance using a novel dataset," introduced a purpose-built navarasa facial-expression dataset spanning the eight classical forms and benchmarked a custom CNN architecture against contemporary vision models (VGG16, ResNet50, InceptionV3, EfficientNet, Vision Transformer, Swin Transformer), using t-SNE visualisation to demonstrate the learned feature space's ability to separate the nine expression categories. This is direct, peer-reviewed evidence that navarasa-based facial expression is a tractable machine-learning target using entirely conventional computer-vision technique — independent confirmation of this paper's underlying premise that the data source is real and workable.

13. Risks and Ethical Considerations

This section deserves, and receives, dedicated treatment rather than a brief closing caveat, because the risks scale with the ambition of the programme, not just its eventual success.

Artificial suffering

If sufficiently rich affective modelling produces functional states that are, even by a narrow behavioural or architectural definition, suffering-adjacent, the research programme itself could generate the harm it studies. Mitigations include architectural review by AI-safety and welfare-focused researchers before Phase 2 training begins, and predefined halt conditions.

Emotional manipulation and dependency

A system optimised for rich, consistent affective engagement is, by construction, also a system optimised to be engaging — a dual-use property already documented as a concern in existing companion-AI and therapy-chatbot literature (Section 13).

Cultural appropriation and misrepresentation

This programme's entire empirical foundation is a specific, living cultural and religious tradition. Any dataset, model, or publication arising from this work should be developed in direct, ongoing partnership with practitioners, gurus, and cultural institutions within that tradition — not merely with their one-time consent to be recorded — and revenue or benefit-sharing arrangements should be established before, not after, data collection.

"Partnership" without named mechanisms is an intention, not a safeguard. The concrete proposal: a Cultural Oversight Board, comprising performers, gurus, and institutional representatives from each participating dance and music tradition, holding explicit veto power over any dataset use, model release, or commercial application involving their tradition's data — not an advisory role, but a binding one. Data licensing should follow a revenue-sharing model negotiated and signed before any recording begins, not renegotiated after a dataset already has commercial value. Consent should be structured as continuing and revocable: any performer or institution should be able to withdraw their contributed data from future model training (though not, for practical reasons, from models already trained and deployed at the time of withdrawal — a limitation that should itself be disclosed to contributors in plain terms before they consent, not discovered by them later).

Synthetic religious or spiritual authority

Because several of the source concepts (bhava, rasa, the Natya Shastra itself) carry devotional and philosophical weight within their tradition of origin, any system trained on this data must be designed and marketed in a way that does not imply spiritual authority, sanctity, or religious endorsement it has not earned and cannot claim.

Military and psychosocial misuse

Emotionally sophisticated systems have plausible dual-use applications in influence operations, interrogation-adjacent contexts, and targeted persuasion. This risk is structural to affective computing generally, not unique to this proposal, but should be named explicitly in any funding or governance discussion of this work.

Identity confusion in long-term users

Multi-session, high-consistency affective systems (Phase 4) raise the same user-attachment and reality-testing concerns already documented in the literature on companion chatbots and grief-technology ("crafting lifelike cognitive copies of loved ones"), and should be developed with the same clinical caution.

14. Value Alignment, Legal Status, and Governance

14.1 The Value-Alignment Problem, Extended

Existing AI value-alignment work generally asks how to align a system's behaviour with human values and law. Some recent scholarship — including doctoral research specifically on military autonomous systems — argues this framing is incomplete: if a system's moral status is genuinely uncertain, alignment work may also need to consider what, if anything, is owed to the system itself, not only what the system owes to us. This paper flags this as a serious, open question this research programme will encounter directly, without asserting that current or near-term systems meet any threshold for moral status.

14.2 Legal Personhood

Existing legal systems are not built to assign rights, duties, or liability to a system whose moral or experiential status is contested. This is a live policy question independent of this paper's own research programme, and this paper takes no position on how it should ultimately be resolved — only that any system emerging from work of this kind should be developed with active awareness that the question exists and remains unsettled.

14.3 The Risk of Premature Attribution

A meaningful and growing share of the public already believes some AI systems are conscious. Commentators including Judith Donath have warned that this belief, combined with AI companies' commercial incentive to encourage it, could produce premature demands for extensive AI legal rights — a dynamic that could distort public discourse independent of whether any underlying consciousness claim is true. This paper's explicit refusal to claim consciousness has been achieved, and its exclusion of the unverifiable ACI metric discussed in Section 9, are both direct, deliberate responses to this specific risk.

14.4 Regulatory Engagement and Independent Audit

This programme's regulatory obligations and independent-audit mechanism are named directly here rather than left implicit. As a India-originated research programme with international consortium ambitions (Section 16), this work should engage directly with India's Ministry of Electronics and Information Technology (MeitY) on domestic AI governance requirements, and with the EU AI Act's risk-tiering framework for any system intended for deployment or study involving EU-based participants or data, given that framework's specific provisions for emotion-recognition systems. Independent audit is proposed as a standing requirement, not an afterthought: any published claim about system capability (Section 10 benchmarks) or consistency (RQ4, Section 3) should be independently reproduced by a research team outside the core consortium before being presented publicly as an established result, following the same reproducibility standard already committed to in Section 11's Phase 5. Publication policy is stated directly here for the same reason: this programme's default position is open publication of methodology, code, and non-sensitive results, with any departure from that default (e.g., biometric data withheld for privacy) disclosed and justified at the time of publication, not left ambiguous.

15. Potential Applications — Framed by Current Evidence, Not Aspiration

Every application below is grounded in either (a) existing, operating systems already cited in this paper, or (b) a direct, incremental extension of methodology already established in the affective-computing literature. None depends on the harder RQ4 question being resolved.

15.1 Mental Healthcare

Emoscape's clinical deployment (Section 12.1) is the clearest existing proof of concept: a sixty-second, non-invasive emotional scan already integrated into diagnostic workflows across multiple Indian medical institutions. Separately, the Ben-Zion et al. (2025) npj Digital Medicine study — "Assessing and alleviating state anxiety in large language models" — found that large language models exposed to depicted traumatic content report substantially elevated "anxiety" scores relative to baseline, that mindfulness-style prompting measurably reduces this, and that elevated reported anxiety states co-occur with measurably more biased model outputs. This is directly relevant, peer-reviewed evidence that affective state (however one interprets its underlying nature) in current LLMs is not inert: it measurably affects downstream output quality and fairness — a finding with direct implications for any deployment of AI systems in therapeutic-adjacent roles, including systems arising from this research programme.

15.2 Eldercare and Companionship

Existing research on AI-supported elderly care documents both real benefit (reduced reported isolation) and real risk (overreliance, blurred reality-testing) from AI systems designed to provide emotional presence, including work explicitly framed around "crafting lifelike cognitive copies of loved ones." Any application of this paper's proposed architecture to eldercare should be developed in direct collaboration with geriatric clinicians and should treat the existing literature's documented risks as design constraints, not afterthoughts.

15.3 Cultural Preservation

Independent of any AI application, systematic, consent-based, benefit-sharing digital documentation of navarasa performance across dance forms is a valuable act of cultural preservation in its own right, with or without downstream machine-learning use.

15.4 Cross-Cultural Emotional Literacy

A structured, culturally-specific emotional vocabulary (the navarasa) offers a genuine alternative to the largely Western, largely English-language emotion taxonomies (e.g., Ekman's six basic emotions) that dominate current affective-computing datasets — a legitimate scientific contribution toward less culturally narrow AI systems, independent of any consciousness-adjacent claim.

16. An Interdisciplinary Research Consortium

No single institution or discipline can responsibly execute this programme alone. The following disciplines are proposed as founding partners, each with a specific, non-decorative role:

This consortium structure is itself a testable design choice: a programme this interdisciplinary is slower and more expensive than a purely engineering-driven effort, and that cost is treated here as a feature, not a bug, given the risks catalogued in Section 13.

17. What Would Falsify This Programme's Core Hypotheses

A research agenda that cannot specify what a negative result would look like is not yet a scientific programme, whatever its other merits. This section states, plainly and in advance, what outcome at each stage would count as evidence against this paper's core proposals, rather than leaving that specification to be worked out informally after results are already in hand.

A negative result on any single row above is not proposed here as a reason to abandon the broader research programme — each hypothesis is separable, and the consortium structure (Section 16) is explicitly designed so that funding and effort can continue on the hypotheses that succeed while the ones that fail are revised or dropped. This section commits to stating those failure conditions before the data exists, not after — the same discipline this paper applies to its own sourcing (Section 21) applied here to its own predictions.

18. Ten Grand Challenges

Rather than asserting the programme is close to a resolution, this paper closes its substantive research content with ten open questions it does not claim to answer:

• How do felt emotional states emerge from embodied interaction, in biological systems and potentially in artificial ones?
• Can such states be measured with methodology both parties to the current consciousness debate (Section 9) would accept as valid?
• Can embodied cultural knowledge, developed over centuries for human transmission, be usefully encoded into machine-learnable representations without distortion or loss of meaning?
• Can an artificial system develop a coherent autobiographical identity across sessions, and what would count as evidence that it had?
• What distinguishes a system simulating an emotional response from a system undergoing one — and is this distinction empirically decidable at all?
• What ethical treatment, if any, is owed to a system whose moral status remains genuinely uncertain?
• Can pre-computational, embodied knowledge systems meaningfully inform contemporary AI design, beyond serving as a data source?
• Does richer affective modelling measurably improve general machine intelligence, or are the two independent?
• Can compassion, as a codified aesthetic-emotional category (karuna), be operationalised as a measurable computational objective?
• What obligations, if any, accompany the deliberate creation of increasingly sophisticated forms of adaptive, possibly affect-bearing intelligence?

A research agenda should be visibly designed to make progress on the challenges it names, not merely gesture at them. The karuna question above is the clearest case worth making explicit: it is not directly tested by any of RQ1-RQ5 as currently scoped, and this paper is not going to pretend otherwise. It is, however, directly reachable by the same architecture: karuna (compassion) is one of the nine rasa categories the Bhava Embedding space (Section 7) is trained to represent, and the benchmark for "response to moral conflict, grief, and wonder" (Section 10) is the closest existing evaluation instrument for it. A dedicated karuna-specific evaluation — for instance, whether the system's embedding-space trajectory during a depicted scene of suffering resembles trained performers' own karuna trajectories more closely than a generic negative-affect response does — is a natural, near-term extension of Phase 5 (Section 11), not a new phase. Every one of these ten challenges should be traceable to a specific phase or benchmark already in this paper, even where, as with karuna, that traceability requires a follow-on study this paper is proposing rather than one it has already fully specified.

19. Objections and Responses

"This isn't science."

The programme's individual components — EEG/biometric emotion correlates, multimodal representation learning, dynamical-systems modelling of temporally-evolving state — are each established, peer-reviewed methodology. What is genuinely novel is the data source and the combination, not the underlying scientific method.

"You can't teach feeling to a machine."

This paper does not claim to. It claims a machine can learn richer, more consistent representations of emotional expression, evaluable on ordinary machine-learning terms, without resolving whether any resulting internal state constitutes felt experience.

"This is religious or mystical."

The programme's proposed methodology treats bhava and rasa as structured, codified, transmissible patterns with a two-thousand-year pedagogical tradition behind them — a data source, not a theological claim requiring belief.

"Where is the proof?"

There is none offered for machine consciousness, because none is claimed. Proof is offered, and cited, for the tractability of navarasa-based emotion recognition (Section 12.2) and for measurable affective-state effects on LLM output quality (Ben-Zion et al., 2025).

"Could this cause harm?"

Yes, and Section 13 is dedicated to naming those harms directly rather than deferring them to a future revision.

"Isn't this just repackaging existing affective computing?"

In part, deliberately — the programme's ambition is to bring an underused, well-structured cultural data source into an existing, rigorous field, not to declare a new field that answers to no existing standards.

20. Conclusion

This paper set out to convert an ambitious, originally under-sourced vision into a research agenda that a serious AI lab, university consortium, or funding body could actually evaluate and act on. That required three changes from the project's earlier form: separating established evidence from hypothesis from speculation throughout; independently verifying every checkable citation, and removing or replacing the ones that could not be verified — most notably an unsubstantiated "Attribution Consciousness Index" claim, replaced here with the genuinely current and considerably more rigorous consciousness-assessment literature; and reframing every claim about machine consciousness as a research question rather than an assertion.

What survives that process is, this author believes, a genuinely interesting and defensible proposal: that centuries-old, highly codified Indian performing-arts traditions encode a structured, high-quality, currently underused data source for affective and embodied AI research, and that a rigorous, ethically-governed, interdisciplinary research programme built around that source could advance the state of the art in multimodal representation learning and affective computing — regardless of how, or whether, the far harder question of machine consciousness is ultimately resolved.

This paper does not claim success. It claims a defensible starting point, and it invites the specific communities — AI safety researchers, consciousness scientists, classical Indian performing artists, ethicists, and clinicians — whose collaboration it would actually require, to test, criticise, and where warranted, dismantle it.

External review of this paper has consistently recommended supplementary materials this document does not itself contain: a formal dataset schema, architecture code stubs, and a pre-registration document for the Phase 1 experiments described in Section 11. That recommendation is accepted directly rather than deferred indefinitely: those materials are the natural next deliverable once a consortium (Section 16) and initial partnerships (Section 11) are in place to make them concrete rather than speculative, and a shorter, less technical companion piece — preserving this paper's ethical framing while stripping the architectural and mathematical detail — is a reasonable parallel deliverable for the broader audiences this topic will inevitably reach beyond the research community this paper is primarily addressed to.

21. References

Barrett, L. (2024). Minds in movement: Embodied cognition in the age of artificial intelligence. Philosophical Transactions of the Royal Society B, 379(1911).

Barrett, L. F. (2016). The theory of constructed emotion: An active inference account of interoception and categorization. Social Cognitive and Affective Neuroscience, 12(1), 1–23.

Ben-Zion, Z., Elyoseph, Z., Spiller, T., & Lazebnik, T. (2025). Inducing state anxiety in LLM agents reproduces human-like biases in consumer decision-making. arXiv:2510.06222.

Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, E., Spiller, T. R., Perez, Y., Wolf, R., Bhonker, S., Suslo, D., Burrer, A., & Schulz, E. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine, 8, Article 132.

Bharata Muni. (c. 200 BCE–200 CE). Natya Shastra, Chapter 6 (on rasa and sattvika abhinaya). [Classical Sanskrit treatise; cited via standard scholarly translation and secondary analysis.]

Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501.

Chen, S., et al. (2025). Exploring embodied multimodal large models. Information Fusion, 118.

Chiarella, S. G., Laurenzi, M., D'Onofrio, M., Gallagher, S., & Raffone, A. (2026). Attribution of consciousness to non-human animals: Insights from AI and multidimensional frameworks. Frontiers in Psychology, 17, 1716363.

Chou, H., et al. (2026). A graphene and liquid-metal composite tactile sensor for robotic touch. Nature Materials. (University of Cambridge, Cambridge Graphene Centre, T. Hasan et al.)

Colombatto, C., & Fleming, S. M. (2024). Folk-psychological attributions of experience to large language models. [As cited in subsequent 2025–2026 literature on AI awareness evaluation; see Section 9.3.]

Craig, A. D. (2015). How Do You Feel? An Interoceptive Moment with Your Neurobiological Self. Princeton University Press.

Damasio, A., & Carvalho, G. B. (2013). The nature of feelings: Evolutionary and neurobiological origins. Nature Reviews Neuroscience, 14(2), 143–152.

Dellermann, D., Calma, A., Lipusch, N., Weber, T., Weigel, S., & Ebel, P. (2019). The future of human-AI collaboration: A taxonomy of design knowledge for hybrid intelligence systems. Proceedings of the 52nd Hawaii International Conference on System Sciences (HICSS), 1–10.

Dellermann, D., Ebel, P., Söllner, M., & Leimeister, J. M. (2019). Hybrid intelligence. Business & Information Systems Engineering, 61(5), 637–643.

Gu, A., Goel, K., & Ré, C. (2022). Efficiently modeling long sequences with structured state spaces (S4). International Conference on Learning Representations (ICLR).

Gu, A., & Dao, T. (2023). Mamba: Linear-time sequence modeling with selective state spaces. arXiv:2312.00752.

Kauffman, S. A. (1993). The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press.

Ko, S. H., et al. (2026). Artificial skin for simultaneous temperature and pressure sensing in robotic and wearable systems. Nature Materials. (Seoul National University)

Lanillos, P., Meo, C., Pezzato, C., Meera, A. A., Baioumy, M., Ohata, W., Tschantz, A., Millidge, B., Wisse, M., Buckley, C. L., & Tani, J. (2021). Active inference in robotics and artificial agents: Survey and challenges. arXiv:2112.01871.

Leitan, N. D., & Chaffey, N. (2014). Embodied cognition and its applications: A brief review. Sensoria: A Journal of Mind, Brain & Culture, 10(1), 3–10.

Lim, R. Y., et al. (2024). Review of EEG affective recognition with a neuroscience perspective. Brain Sciences, 14(4), 364.

"Living the Abhinaya": Sensory-somatic approaches to rasa-bhava in Odissi dance. [Ethnographic study; author not identifiable from available secondary source and cited here by title pending direct verification, per this paper's own sourcing standard.]

Mahon, B. Z. (2014). What is embodied about cognition? Psychonomic Bulletin & Review, 21, 940–949.

NextGen Economics. (2026). The Continuum of Intelligence (White Paper 7). Bangalore: NextGen Economics.

Nihilent Ltd. (2025–2026). Emoscape platform documentation and partnership announcements with the Maharashtra University of Health Sciences (MUHS) and the NIMHANS ecosystem.

"Phenomenology of Consciousness in Ādi Śaṅkara and Husserl: A Comparative Study of Sākṣī Caitanya and the Transcendental I." [Comparative philosophy study; author not identifiable from available secondary source and cited here by title pending direct verification, per this paper's own sourcing standard.]

Prigogine, I. (1977). Self-Organization in Non-Equilibrium Systems. Wiley.

Proceedings of the National Academy of Sciences. (2025). A neuromorphic robotic electronic skin with active pain and injury perception. PNAS, 122.

Sangeet Natak Akademi. (n.d.). Classification of the eight recognised Indian classical dance forms. [Institutional reference; Government of India.]

Schwitzgebel, E. (2026). AI and Consciousness: A Skeptical Overview. Cambridge Elements. [Preceded by arXiv preprint, January 2026.]

Shapiro, L. (2021). Embodied Cognition. Stanford Encyclopedia of Philosophy.

The Visual Computer. (2025). Visual analysis and recognition of facial expressions in Indian classical dance using a novel dataset. The Visual Computer, 41(11).

Wilson, M. (2002). Six views of embodied cognition. Psychonomic Bulletin & Review, 9(4), 625–636.

Zeami Motokiyo. (c. 1400–1418). Fūshikaden [The Transmission of the Flower of Acting Style]. [Classical Japanese Noh treatise; cited via standard scholarly translation and secondary analysis of hana, yūgen, monomane, and mushin.]

Publisher's Note

Note on sourcing: every citation above was independently checked against original or secondary reporting before inclusion. One claim considered during this project's research — a named "Attribution Consciousness Index" metric — could not be independently verified and does not appear in this paper; several other unverified doctoral-thesis and preprint citations were excluded on the same basis. Citation format follows a consistent author-date convention throughout; entries without a named individual author (e.g., institutional or unattributed press-reported sources) are listed under the responsible organisation, in the same alphabetical sequence as author-named entries, with a bracketed note where relevant.

NextGen Economics is an independent research platform producing evidence-based, accessible research on the economic and societal implications of technological change. This white paper is a research agenda positioning embodied Indian performing-arts traditions as a data source for affective AI, with machine consciousness treated as a long-term motivating question rather than a claim (Executive Summary). Its sourcing has been independently verified throughout — consistent with this publication's broader editorial standard: claims are sourced, dated, and corrected in public when they do not hold up. The formal Disclaimer immediately below states the paper's scope and limits precisely.

Disclaimer

— Pawan Bhatia
NextGen Economics · Bangalore, India · July 2026

This white paper presents a speculative research agenda, not an empirical scientific finding or a statement of settled fact. It does not claim that machine consciousness has been achieved, measured, or transferred. The views expressed herein are those of the author and do not necessarily reflect the views of any affiliated organisations. This document does not constitute scientific, medical, legal, or regulatory advice.

AI Tools Disclaimer: This white paper was developed with the assistance of AI-powered research and writing tools, including large language models for literature synthesis, drafting, formatting, and independent citation verification. All substantive arguments, technical proposals, ethical framings, and final editorial decisions — including which claims to keep, cut, or re-source after verification — remain the sole responsibility of the author. The use of AI tools does not imply endorsement of the content by any AI system or its developers.

Reproduction and Distribution: This white paper is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. You are free to share, adapt, and distribute this work for non-commercial purposes, provided you attribute the author appropriately.