The problem
Quantum chromodynamics has a striking prediction: because gluons interact with each other, there should be particles made of pure glue — no quarks at all. The lightest of these glueballs, the scalar with quantum numbers 0⁺⁺, has a mass that cannot be computed with pencil and paper. It only exists non-perturbatively, and for forty years the standard way to get at it has been lattice gauge theory: discretize spacetime, simulate the gluon field with Monte Carlo, and measure how correlations decay.
That makes it a perfect testbed for an AI lab. The answer is known from decades of careful human work, the calculation is genuinely hard and multi-stage, and there are dozens of places to fool yourself. We asked a coding agent to do the whole thing from scratch: write the simulation code, generate the ensembles, extract the spectrum, and — the hard part — decide which of its own numbers to believe.
What the agent built
The agent wrote its own SU(3) pure-gauge lattice code with a Wilson plaquette action, accelerated on Apple-silicon GPUs via Metal, with CPU double precision kept as the correctness oracle behind explicit parity gates — Metal was only allowed to generate data after matching the CPU on test lattices. On top of the simulation it built the full measurement pipeline: smeared glueball operators, a generalized eigenvalue problem (GEVP) across the operator basis to isolate the ground state, and torelon (flux tube winding the periodic box) correlators to set the physical scale through the string tension σ.
Everything is quoted in the dimensionless ratio M/√σ, so the lattice scale drops out; the only external input is the conventional value √σ = 440 MeV used to convert to GeV at the very end.
Guardrails against self-deception
The interesting engineering is not the physics code — it is the machinery that keeps an eager optimizer honest. Three guardrails did the real work.
Blinding. The analysis was target-blind: fit windows, operator choices, and quality thresholds were fixed without reference to the literature value, so the agent could not tune its way toward the "right" answer.
Fit hygiene. Mid-run, the pipeline produced a textbook pathology: a fit pinned at the edge of its mass grid (aM = 0.03) with a zero jackknife error and a huge negative constant term — numerically "perfect", physically meaningless. The fix was not to delete the point but to add gates: reject boundary-of-grid fits, non-positive errors, and constant terms above 25% of C(0). Wrong answers should be excluded by rule, not by taste.
Finite-volume gates. Each coupling β only enters the final fit if the measured mass and scale agree between two box sizes (L = 20 and L = 24) within a preregistered 10% tolerance. This gate had teeth: at β = 6.40, the finest lattice we ran, the two volumes disagreed by 71.6% in mass and 21.7% in scale. The point looked perfectly plausible on its own — it was excluded purely because the rule said so, and a replacement coupling (β = 6.25) was run instead.
The result
Three couplings survived every gate. Each gives a continuum-bound data point at its own lattice spacing:
| β | lattice | a²σ | M/√σ | volume gate |
|---|---|---|---|---|
| 6.20 | 24³×48 | 0.0264 | 2.801(248) | pass |
| 6.25 | 24³×48 | 0.0247 | 3.816(344) | pass |
| 6.30 | 24³×48 | 0.0169 | 3.611(316) | pass |
| 6.40 | — | — | excluded | fail (71.6% shift) |
A weighted linear extrapolation in a²σ — the leading discretization artifact of the Wilson action — takes these to zero lattice spacing:
Qualified continuum estimate
M/√σ = 4.84(98) → M(0⁺⁺) = 2.13(43) GeV
The reference values from the human lattice literature are M/√σ = 3.405(21) (Athenodorou & Teper) and M = 1730(50)(80) MeV (Morningstar & Peardon). Our estimate sits about 1.5 combined standard deviations above them — consistent, but not a precision result, and we say so plainly: the fit has three points, one degree of freedom, χ²/dof = 4.7, and no systematic error budget beyond the gates themselves. This is a qualified continuum estimate, not a benchmark.
What we learned about AI doing physics
The agent's failure modes were the failure modes of a hurried graduate student: falling in love with a numerically perfect but degenerate fit, trusting a plausible-looking data point that finite-volume checks later demolished, and wanting to quote a smaller error than the data supports. None of these were caught by the agent "being careful" — they were caught by rules it was made to preregister before seeing the data. That is the transferable lesson: the path to AI that does real science runs less through raw cleverness and more through building the harness that makes honesty mechanical.
The error bar is currently dominated by the geometry of the fit — two of the three points sit nearly on top of each other in a²σ, so the extrapolation has a long lever arm. A fourth coupling at β = 6.35 is running now; if it passes its gates, the intercept error should shrink by roughly 40%. The log will be updated either way — including if it fails.