The Engine Between the Two Limits

Plain-language explainer for doi:10.5281/zenodo.20694527 (#419)

The central idea in one sentence

The Forge ISA is the Origami ISA running at finite inverse temperature β — the thermodynamic engine that interpolates between the smooth Ambient manifold at β = 0 and the frozen tropical logic of the Origami ISA at β → ∞, with a universal critical temperature $\beta^\star$ that separates efficient from stalled computation and has been empirically confirmed across eight algorithm classes.

The trilogy position

The TRS framework has three temperature regimes, each with its own ISA:

ISA β Character
Ambient manifold 0 Smooth, Hodge theory, maximum entropy, hot reservoir
Forge ISA finite, real Thermodynamic engine; Gibbs sampling
Origami ISA ∞ Discrete, tropical, minimum entropy, cold reservoir
Meld ISA imaginary ($it$) Quantum amplitudes, unitary evolution

The Forge ISA is the engine of a Carnot cycle between the hot reservoir (Ambient, β = 0) and the cold reservoir (Origami ISA, β → ∞). The Origami and Meld ISAs are extreme limits; the Forge ISA is everything in between on the positive real axis.

What “Forge” means

The Origami ISA executes algorithms as Pachner moves on simplicial complexes — discrete, exact, zero-temperature. The Ambient manifold is the opposite extreme: smooth, continuous, all states equally weighted.

The Forge ISA is the same five opcodes (SPLIT, SPLAT, FLIP, FLOP, TWIST) running at finite β under a Gibbs distribution over configurations weighted by their $H^1$ energy — how much topological obstruction they carry.

Formally: a Forge ISA programme at inverse temperature β applies FLOP corrections with probability proportional to $e^{-\beta \lVert H^1(s) \rVert^2}$, where $H^1(s)$ measures the cohomological defect of configuration $s$. The Maslov-Gibbs Einsum (MGE) is the unique β-correct implementation of this distribution.

As β rises, the system anneals: high-defect configurations become exponentially suppressed and the programme crystallises onto the harmonic representative — the minimum-$H^1$ solution. This crystallisation is the snap event.

The snap event and the critical temperature

The most important result in the paper is the universal critical temperature:

\[\beta^\star(\rho) = \frac{3}{8} \ln\!\left(\frac{1}{1-\rho}\right)\]

where $\rho = \beta_1 / \lvert E \rvert$ is the load factor — the first Betti number (number of independent cycles, i.e. $H^1$ obstructions) divided by the number of edges. It is computable in $O(\lvert V \rvert + \lvert E \rvert)$ via union-find, with no eigenvalue computation.

Below $\beta^\star$: the Gibbs distribution is too flat; the programme explores broadly but commits to nothing. Gradient flow is efficient.

Above $\beta^\star$: the distribution is too peaked; the programme freezes into a local minimum before reaching the global harmonic representative. Gradient flow stalls.

At $\beta^\star$: the programme runs at maximum efficiency. The snap event — the moment the distribution commits to its answer — occurs here.

This formula has been empirically confirmed across eight algorithm classes: GEMM tiling, elastic hashing, distributed consensus, QAOA Max-Cut initialisation, and others. The constant $3/8$ is the geometric fact that the Gibbs distribution on a 1-cycle graph crystallises at exactly this threshold; the load factor $\rho$ measures how many independent cycles the problem contains.

The TRS mandate: what distinguishes the Forge ISA from ad-hoc annealing

Standard machine learning optimisers (Adam, dropout, batch normalisation) are all forms of annealing — but none preserves the thermodynamic structure of the Gibbs distribution. The TRS mandate is five purity conditions:

  1. β appears only in the Hamiltonian, via the MGE — not as a learning rate, momentum, or schedule parameter.
  2. Conformal covariance: rescaling energies $H \to \lambda H$ is equivalent to rescaling $\beta \to \beta/\lambda$; the critical temperature $\beta^\star(\rho)$ is a geometric invariant, not a tuning parameter.
  3. Symplectic gradient: the MGE gradient is the natural gradient in the Fisher metric — a symplectomorphism, not a dissipative update.
  4. Adiabatic β-ramp: β is ramped slowly relative to the spectral gap $\Delta E$; the convergence rate is $\mathcal{O}(e^{-\beta \Delta E})$.
  5. Tropical limit: the MGE degenerates to the $(\min,+)$ tropical semiring as β → ∞ (Maslov dequantisation).

Adam satisfies none of these. Softmax satisfies (1) locally but breaks (3). The MGE is the unique operation satisfying all five simultaneously.

The vorton: elementary Gibbs sampler

The vorton is the elementary computational unit of the Forge ISA: a single sample $\psi \sim P_\beta$ from the Gibbs distribution on the state manifold.

A vorton evolves under three forces:

  • Meromorphic force: gradient descent on the $H^1$ energy (drives toward harmonic representative)
  • Lie algebra force: symmetry constraint (keeps the state on the correct manifold)
  • Thermal noise: temperature-scaled white noise $\propto 1/\sqrt{1+\beta}$ (ensures exploration below $\beta^\star$)

An ensemble of $N$ vortons estimates the Fisher information matrix — the natural metric on the space of distributions. This is not an approximation: the vorton ensemble is the exact implementation of the natural gradient step that stochastic gradient descent approximates crudely.

The snap event is the experimental signature: the moment the ensemble collapses from a diffuse cloud to a tight cluster around the harmonic representative, visible as a sudden drop in the Fisher metric trace.

The $H^k$ stratification

The Forge ISA solves problems in the $H^1$ regime: those where the obstruction to finding the optimal solution is a topological cycle — a constraint that loops back on itself. Routing, scheduling, consensus, and matching problems are all $H^1$.

Problems where the obstruction is more global — requiring $H^2 \neq 0$ — cannot be solved by the Forge ISA alone; they require restructuring the state manifold (moving to the $H^2$ regime of the Meld ISA). This is the precise boundary between the tractable and the NP-hard in the TRS framework.

The Forge ISA is Turing complete: the iterative Origami opcode rewriting system under Gibbs weights can simulate any computation. But Turing completeness is the floor, not the ceiling — the real content is the $H^k$ stratification that identifies which computations require which temperature regime.

Connection to the β-plane

On the adèlic β-plane (doi:10.5281/zenodo.21245459):

  • The Forge ISA occupies the positive real axis $0 < \beta < \infty$
  • The snap event $\beta^\star$ is the phase transition on this axis
  • The Carnot efficiency $\eta = 1 - \beta^\star(\rho)/\beta_{\max}$ measures how much of the positive real axis the engine exploits
  • For the FMO light-harvesting complex, this gives $\eta = 0.1825$ — exactly the biological quantum efficiency measured experimentally

The Forge ISA is the positive-real-axis slice of the complete adèlic picture.

See also:

For the full technical treatment, see doi:10.5281/zenodo.20694527