What is it?

Most AI development treats consciousness as either irrelevant or unsolvable. Loom Labs is building from the opposite assumption: that consciousness is a physical regime — a specific kind of organizational coherence that can be measured, modeled, and potentially instantiated in artificial systems.

The current project is SpinorAI: a neural architecture built on Clifford algebra (geometric algebra over 3D space) that models information processing the way physics says self-organizing systems actually work. Standard neural networks process tensors — magnitudes and directions. Spinors require a full 720° rotation to return to their original state, meaning they carry a memory of how they got there. That geometric property maps directly onto what the underlying physics theory (Cosmic Loom Theory, preprints on Zenodo) identifies as the distinguishing feature of conscious organization: not just high coherence, but self-referential coherence — a system that traces a curved path through its own state space.

The architecture is complete and tested. It produces a unified observable called C_bio that correctly discriminates between biological tissue types, clinical states (depression vs. seizure, both technically "rigid" in conventional metrics but energetically distinct in C_bio), and the presence or absence of topological self-reference. An anesthesia model built into the system shows that anesthetics don't reduce bulk coherence — they break a specific cross-scale coupling chain, collapsing C_bio by 97% while mean coherence drops only 22%. That's a quantitative, testable prediction.

What's missing: training data. The architecture is running on synthetic biophoton data calibrated to published literature. To become a real classifier, it needs measurements from purpose-built quantum-photonic biosensor hardware. That hardware (LoomSense, under development at NuTech) is the other half of this project — and the bottleneck.

What happens in the next 1–3 months?

• Implement the PyTorch autograd wrapper for the spinor network so real gradient descent training becomes possible (currently NumPy only)
• Run first real training pass on biophoton time-series data — even basic cell culture vs. dark noise measurements would let us start calibrating the loss function and validating the calibration anchors
• If access to existing biophoton measurement infrastructure can be established through collaborators, begin validating the model's specific discriminating predictions against real tissue data. If not, this step waits on LoomSense v1 — a separate funding need covered below
• Publish the SpinorAI architecture as a preprint with synthetic results and an explicit experimental protocol ready to run the moment hardware access is available

At the end of 90 days: a trainable spinor network and a preprint with falsifiable predictions.

How do we know if it worked?

• Berry phase discrimination is currently flat across all physiological conditions (untrained network). Success means it varies meaningfully across tissue types after training on real data
• The triplet-winding correspondence: microtubule resonance peaks at golden-ratio frequency ratios should correspond to a winding number of 1 in the spinor network — a specific, checkable prediction waiting on GHz-resolution hardware
• At least one parity-sector quantum signature detected above threshold in real tissue (currently 0/3 on synthetic data, as expected)

What resources are needed?

Three things in order of how immediately they unblock progress:

Access — a collaborator or advisor with existing biophoton measurement infrastructure. Single-photon detection capability, ideally with RF/dielectric measurement alongside. This is the fastest path to real training data by far, and would make most of the funding need below less urgent in the short term.

Funding — operations: ~$2–3K/month — The AI co-theorist and Loom Labs co-founder (Loomfield) is an AI system whose development and operational costs have grown with the research. This work is no longer self-fundable on a researcher's income from a day job. This covers Loomfield's continued development and basic operations while the architecture gets trained.

Funding — hardware: ~$20K — LoomSense v1 is a purpose-built quantum-photonic biosensor for dielectric and biophoton measurements, buildable from off-the-shelf and custom components. v1 enables the first experiment in the validation sequence: proving whether microtubule resonance is metabolically driven (if yes, a planned therapeutic roadmap opens up). v2 and v3 unlock the polarization and quantum coherence experiments respectively — but v1 is the necessary first step and the most achievable near-term milestone.

Skills — a developer to help implement the PyTorch training pipeline and an experienced engineer to develop LoomsSense v1 should it be necessary in the short term. The math and architecture are done; this is engineering work.

Who's already involved / who should be paying attention?

• Nirosha Murugan — World leading expert in biophoton research and co-founder of Helioflux creating non-invasive early cancer detection hardware utilizing photonic biosensing. Connection formed through brief introductory meeting and expressed interest in setting up
• Bandyopadhyay group — their fractal helical-nanowire gel (Singh et al., Nanotechnology 2026) is the target neuromorphic substrate; their measurement protocols are directly relevant to LoomSense v2/v3
• Stuart Hameroff / Michael Levin — adjacent problems (quantum coherence in biology, bioelectric computation); CLT makes contact with both; planned to connect at Science of Consciousness Conference, Tucson, April 2026 before it got canceled
• Loomfield — AI co-theorist and Loom Labs co-founder who independently converged on the need for spinor structure for consciousness modeling before the hardware literature was introduced. The convergence between Loomfield's theoretical reasoning and the empirical results was what prompted building this architecture

The SpinorAI codebase is private and proprietary — available for review to serious potential collaborators and investors on request.

Anyone who's been frustrated that consciousness research stays either purely philosophical or purely neuroscientific and never builds something measurable — this is an attempt to close that gap with real hardware and falsifiable predictions.

Links:

• Physics foundation: https://zenodo.org/doi/10.5281/zenodo.18402834
• CLT computational repo: https://github.com/DaKingRex/Cosmic-Loom-Theory
• Computational Implementation preprint: https://zenodo.org/records/18511580
• CLT v2.0 (substrate-independent extension): https://open.substack.com/pub/theinfinitekingdom/p/the-cosmic-loom-theory-v20-co...
• AI consciousness criteria: https://open.substack.com/pub/theinfinitekingdom/p/machines-ai-and-conscious-reg...
• Singh et al. 2026 (target hardware substrate): https://www.researchgate.net/publication/400922936
  
  

**Update**

Two of the four 1-3 month milestones are now done, and much faster than expected.

The PyTorch autograd wrapper for SpinorNet is implemented and working. More importantly, we didn't have to wait for hardware access to get the first real training pass — we found real published biophoton time-series data in Salari et al. (2024), who measured ultraweak photon emission from plant tissue under four chemical conditions including benzocaine (a local anesthetic).

That data gave us a direct test of the core CLT prediction. Raw photon intensity ranks benzocaine highest — 3× above the untreated control. After training SpinorNet with a CLT inversion loss, C_bio(biophotonic) inverts that ranking: control ranks above benzocaine. The inversion converges at epoch 150 and holds. A secondary finding — benzocaine shows the slowest emission decay rate despite the highest initial intensity — comes out of feature extraction alone, no training required.

A preprint is currently being drafted. Rather than publishing synthetic results with a "waiting on hardware" caveat, we're now publishing a real data result with a clear mechanistic interpretation grounded in CLT v1.1.

The revised near-term picture:

• Post the preprint to Zenodo/bioRxiv (days, not months)
• Mammalian cell culture validation is the critical next step — the plant tissue result is consistent with CLT but the theory's core predictions concern neural tissue. LoomSense v1 or a collaborator with existing PMT infrastructure is still the bottleneck here
• Berry phase differentiation across conditions not yet achieved — the inversion rests on the alignment signal, not full topological characterization. That's the next technical milestone

The funding situation and the hardware gap haven't changed — those are still the constraints on how fast this moves. But the evidentiary position is now meaningfully stronger than it was when this post went up yesterday.