Windowsill Lab · Run M14 · Statistical Physics

A line through a disordered
phase diagram where
the answer is exact

In a system built out of randomness, is there any path along which theory can still tell you the answer exactly — before you run anything?

Painterly illustration: a grid of small magnets webbed with threads, some taut and ember-red, with a single polished brass rail arcing cleanly across the tangled field.
ILLUSTRATION (AI-PAINTED) - ONE EXACT PATH THROUGH A DISORDERED FIELD

Disordered systems are hard because the randomness is baked in. You cannot average it away, you cannot solve around it, and almost nothing about them is exactly computable. So it is genuinely surprising that there exists a single curve through the random-bond magnet's phase diagram along which the energy is known in closed form, at every lattice size, with nothing fitted. It is called the Nishimori line, and M14 walks along it.

Here is the setup. Take a square lattice of magnets and make a fraction p of the bonds hostile — those pairs want to disagree, the rest want to agree. Higher p means more frustration. Normally you would be free to set the temperature independently, but the Nishimori line ties them together: tanh(1/T) = 1 − 2p. Pick a p and the temperature is forced. That constraint is not a convenience; it is the condition under which a hidden gauge symmetry of the model kicks in, and the symmetry is what makes the energy computable.

What it says is that the disorder-averaged energy per magnet must be exactly −2·tanh(1/T), equivalently −2(1 − 2p). Every point on the line is therefore a prediction the simulation was never told, and every measurement is a checkable falsification. Because it is a symmetry identity rather than a critical temperature, it holds at any lattice size — no finite-size shift to correct, no extrapolation to argue about. That is what makes it verify cheaply and robustly. It is also, honestly, what makes it a weaker claim than pinning the interesting point on that line would be.

The interesting point is where the line crosses out of the ordered phase — the multicritical Nishimori point, at roughly p ≈ 0.1094, T ≈ 0.9528 in the literature. This run brackets it but does not pin it: the ferromagnetic order visibly collapses as p rises along the line, passing half strength near p ≈ 0.126. Pinning it properly needs a much larger run. The panel below gives you one slider for p, which drags the temperature with it, and shows the measured energy landing on a line nobody fitted.

Live · random-bond ±J magnet on the Nishimori line · M11's bond machinery L = 24 AS PUBLISHED — BUT FAR FEWER SWEEPS AND REALIZATIONS
p = 0.040

left · the fabric — ember bonds are hostile, moss ones friendly; raising p reddens it and the order dissolves. right · top: the p–T plane with the Nishimori line and where the slider has put you; bottom: measured energy against the exact −2·tanh(1/T). The satisfying moment is the measured dot riding the theory line while everything else in the picture falls apart.

temperature (set by p — you cannot move it) · measured E/N · predicted · deviation · ⟨|m|⟩ · realizations averaged 1

Press off the line and the constraint is released: the temperature is pushed away from the value p demands, and the measured energy walks straight off the prediction curve while nothing else about the picture changes. Nothing demonstrates as economically that the line is special — it is not that this model is easy, it is that this one path through it is exactly solvable.

0.015max deviation from the exact identity, L = 24, 64 realizations MEASURED
0.85 → 0.22⟨|m|⟩ collapsing as p rises along the line MEASURED
p ≈ 0.126where that collapse passes half strength MEASURED
182 swall time, 2026-08-13 rerun

The graded claim is the first number and only the first number. Across p ∈ [0.04, 0.16] at L = 24 over 64 disorder realizations, the measured disorder-averaged energy never departs from −2·tanh(1/T) by more than 0.015 — a prediction with nothing fitted, reproduced at every point on the walk. That is what earns the green leaf.

The second number is reported, not graded. The ferromagnetic order does collapse along the line, from 0.85 down to 0.22, and it passes half strength near p ≈ 0.126 — which brackets the literature's multicritical point at pc ≈ 0.1094, Tc ≈ 0.9528, but does not pin it. Bracketing and pinning are different claims and this page keeps them apart on purpose.

The bond machinery here is M11's, reused verbatim — same checkerboard, same J-weighted neighbour sum, same Metropolis half-sweep — so the random-bond engine can never drift away from the spin-glass one. The three-dimensional version of that spin glass, still looking for its transition, is M12.

What is verified here is an exact identity, not a critical point. A gauge symmetry fixes the disorder-averaged energy along this line at any lattice size, which is why it checks out cheaply and robustly — and also why it is a weaker claim than locating the interesting point on that line would be. The multicritical point itself is only bracketed: the ferromagnetic order passes half strength near p ≈ 0.126 against a literature p_c ≈ 0.1094, and the two-size Binder crossing does not resolve at reachable disorder averaging. Pinning it is deferred to a much larger run.