Braid programming and the organoid training problem

What the work claims

Jipdi and colleagues, a Cameroon-Gabon collaboration, build a complete theoretical stack on one physical object: a non-Hermitian Bogoliubov-de Gennes Hamiltonian, the kind of effective model that describes gain-loss or particle-hole balanced systems, with non-Hermiticity strength gamma and couplings J1, J2, and energy offset a. In such systems, exceptional points (EPs) are spectral degeneracies where eigenvalues and eigenvectors coalesce, and encircling one adiabatically swaps eigenstates while accumulating a geometric phase. The paper's first claim is a closed algebraic equation for the EP super-surface, together with a momentum quantisation rule k = sigma pi over N, which predicts the exact number and positions of all exceptional points in a finite system: N/2 of them for nonzero a, verified by exact diagonalisation at 2, 4, 6, and 8 unit cells.¹

The second claim is computational. Encirclements of the EPs act as quantum-like gates: a small loop yields Pauli-Y, a larger loop that couples in the auxiliary subspace yields Pauli-X, a half loop yields a Hadamard-like gate with a geometric phase of pi/15, the special case a = 0 adds Z-like and T phase gates, and two-qubit encirclements give SWAP and controlled-phase primitives. Because these holonomies depend only on the topology of the path, not its precise shape, the authors argue they form a noise-immune, non-Abelian gate library. The third claim is the paradigm: learning as braid programming, a discrete search over compositions of these gates, replacing gradient descent on continuous weights with combinatorial optimisation. A genetic algorithm with tournament selection, crossover, and mutation, searching braid words of up to length six over a ten-letter alphabet (the five gates plus inverses), recovered the standard Hadamard gate as a length-2 word and the H-dot-Z gate as a length-1 word at 0.989 Frobenius fidelity, reaching perfect fidelity under the more permissive state-overlap measure.¹

How it works

An exceptional point is a defective degeneracy: the Hamiltonian cannot be diagonalised there, and the eigenvectors merge. Berry's classic analysis showed that transporting states around such a point permutes the eigenvalues and imparts geometric phases with no Hermitian analogue; the swap fidelity takes values 0 or plus-minus 1, the normalised Berry phase takes values 0 or plus-minus pi/2, and a topological EP cannot have both invariants at zero. The paper derives exactly where these points live: a super-surface equation, cos-squared k equals gamma-squared times (J2-squared plus a-squared) over (J1-squared a-squared plus gamma-squared J2-squared), which in a finite system collapses onto quantised momentum layers. Once J1, J2, and a are fixed, the equation inventories every available exceptional point and the gamma values that reach them, no diagonalisation needed.¹

Braiding turns those points into operations. Encircling the odd-sector EP of a two-spin realisation yields a holonomy equal to Pauli-Y, exchanging the computational basis states |01> and |10> with a phase of plus-minus pi over 2. Widening the loop until the auxiliary even subspace couples rotates the holonomy into Pauli-X; a non-closing half loop produces the Hadamard-like gate. At a = 0 the same encirclements become phase gates, including a T-type gate, and combined loops give SWAP and controlled-T over the full four-dimensional space. A computation is then a braid word: a finite sequence of these generators, evaluated right to left, whose non-commutativity supplies universal computation. Learning becomes the search for a word whose unitary maps training input states to their target outputs within tolerance, a discrete combinatorial problem rather than continuous optimisation.¹

Where a skeptic should push

The most load-bearing assumption is adiabaticity, and the paper itself supplies the counterevidence. The claimed noise immunity is topological: holonomies depend only on the homotopy class of the encirclement path. But the authors' own encirclement at k = pi over 6 lands in a transitional regime where the Berry phase drops to fractional values (plus-minus pi over 4 in one block, plus-minus 3 pi over 4 in the other) because the path partially encloses the branch cut or non-adiabatic transitions leak. Their topological map quantifies the exposure: only 45.0% of Block 1 and 44.0% of Block 2 swapping points carry a quantised topological invariant, on a map where 49.6% and 49.3% of points swap at all. In other words, in nearly one swapping point in ten the gate works but the topological protection the paradigm is sold on does not. A physical realisation on a noisy, slow substrate would live near that boundary, and the paper offers no noise simulation, no fidelity-decay curve, and no decoherence analysis to say how immune the braids actually are.¹

Second, the fidelity headline needs a footnote the abstract does not give. The genetic search reached 0.989 Frobenius (matrix) fidelity for both target gates, with the residual error attributed to the finite gate alphabet; perfect fidelity was reached only under state-overlap fidelity, which is insensitive to global phase. "Reproduce the standard Hadamard gate with perfect fidelity" is not what the stringent metric shows, and the discrepancy should be stated, not smoothed over. Third, everything is simulation of a two-level-plus-auxiliary toy system: the universal gate set and the learning demonstration operate on one qubit in a four-dimensional Hilbert space. Scaling to anything a machine learning practitioner would recognize as a task is an act of faith, with the braid search over longer words facing a combinatorially exploding space and no demonstrated heuristic that survives it. This is a framework paper, not an existence proof of a working learner.¹

Braid programming as an organoid training paradigm

The non-obvious implication for organoid intelligence is not the physics. Nothing in a dish of stem-cell derived neurons implements a non-Hermitian Bogoliubov-de Gennes Hamiltonian, and any article claiming otherwise should be discounted on sight. The transferable object is the training paradigm, and it maps onto the organoid problem with an uncanny precision the authors, writing about quantum hardware, never notice. The central fact of training biological computing systems is that backpropagation through living tissue does not exist: there is no differentiable path from your loss function to the synaptic weights of a culture. Every serious closed-loop organoid training system built so far does exactly what braid programming formalizes: apply physical perturbation primitives, observe the state change, and search, stochastically and compositionally, for sequences that improve task performance. Braid programming gives that folk practice a formal backbone: define your primitives, demand that they be discrete and robust, and replace gradient descent with combinatorial search over compositions.¹

The opportunity is three design constraints the paradigm imposes on any substrate, living or otherwise. First, primitives must be robust in the substrate's own physics: perturbations below some scale change an operation's parameters but not its identity, which for an organoid-electrode system translates into stimulation protocols whose effect on population state survives the culture's drift, an experimental question almost nobody asks. Second, programs must compose without interference; the paper's concatenation claim, that appending a new braid word does not corrupt stored words, is precisely the catastrophic-forgetting property continuous-weight systems lack and any lifelong-learning organoid would need. Third, training data requirements collapse: the genetic search learned its target from essentially nothing but the target unitary, suggesting a compositional primitive library could be searched with far less data than gradient methods demand, which matters when your training signal is hours of expensive electrophysiology per day.

The threat is the mirror image. If learning on noisy physical substrates is fundamentally discrete search over protected primitives, then much of the organoid training literature, which applies continuous optimization metaphors and reports single-task performance before the culture drifts, is measuring the wrong thing. A culture that cannot expose a stable primitive library cannot be braid-trained, and no readout engineering fixes that; the topological map's own transitional regime, where swaps work but protection fails, is a decent metaphor for an organoid that responds to stimulation differently tomorrow than today. The honest research program the paper implies is unglamorous: characterize the reproducibility envelope of stimulation primitives in long-lived cultures before claiming any learning at all.¹

The bottom line

Established in this preprint: a closed super-surface equation inventorying all exceptional points of a non-Hermitian BdG Hamiltonian at any system size, verified by exact diagonalisation up to eight cells; a universal braid gate library from adiabatic encirclements; and a proof of concept in which genetic search recovers single-qubit target gates as short braid words at 0.989 matrix fidelity, perfect under state-overlap fidelity. Asserted but not established: noise immunity (argued from topology, never tested against noise), guaranteed generalization, forgetting prevention at scale, and physical realizability of the Hamiltonian. For organoid intelligence the durable contribution is conceptual: a rigorous statement of what training without gradients looks like, discrete, compositional, built on robust primitives, which happens to describe the only kind of training living tissue can receive. What would confirm the transfer is evidence that organoid cultures expose stimulation primitives reproducible enough to compose; what would break it is the field's own data on culture drift, which currently suggests they do not.