We trust a model only after it reproduces someone else’s published result

Gravitational waves

The 4.5PN inspiral phase of a compact binary, the template detectors use. Blanchet, Faye, Henry, Larrouturou & Trestini, PRL 131, 121402, 2023.

• 36 checks, all proved symbolically;
• 53 of 54 table entries agree to the printed digits;
• the 54th is a misprint (95.502 printed for 93.502). It affects nothing downstream.

Black-hole accretion

The Event Horizon Telescope GRMHD code comparison, a torus around a spinning black hole. Porth et al., ApJS 243, 26, 2019.

• 41 physics checks;
• in a 2D run, the accretion onset came at 262 M against about 300 M published;
• the peak accretion rate (0.572) and peak magnetic flux (3.24) fall inside the published 3D codes’ ranges;
• the checks found a sign misprint in a widely used method paper’s wave-speed equation.

Turbulence in four dimensions

Navier–Stokes in four space dimensions, compared with Berera, Ho & Clark 2020 and Ho et al. 2025.

• from a 9-line Phy file, Phy generated the 4D solver, which matched an expert hand-written solver to 3.2e-15 after 200 steps;
• all four cascade statistics moved the published way; the chaos exponent was lower in 4D, as published (one run, 1.2σ);
• of 8 realistic mistakes, Phy refused 5 before running and hand-written code refused 0;
• a viscosity given in cm²/s was handled correctly in Phy and was 10⁴× wrong in the hand code, with every test passing.

GPS clocks

The relativistic clock correction, on 7 days of real satellite data: 2,826 broadcast messages, 562,966 clock points, GPS week 2431.

• Phy proved that the interface specification’s two “equivalent” correction forms are equal only for an ideal Kepler orbit;
• in the real broadcast orbits they differ by up to 69.7 ps; the real clocks respond as Phy predicted, and the literal “equivalent” reading is excluded at 8.4σ;
• published terms were reproduced (Ashby 2003: 23.98 ps against 23.83 ps);
• a proven, measured correction to the specification’s wording, worth 1–2 cm.

Why Phy

The twin is written in the same checked equations as the device firmware, so the simulation and the instrument cannot quietly drift apart. A wrong unit, sign or law is refused before the twin runs.

Why trust it

The method has reproduced published results before it is used on a partner’s sensor: four published papers above, and commercial atom-interferometer instruments below.

What a partner gets

Expected performance in their conditions, the points where the sensor fails, and what to change, before the cost of a field trial. Become a pilot partner →