{
  "cases": [
    {
      "name": "white noise counted twice",
      "model": "published-spectrum",
      "hellingsDownsSignalToNoise": 0.4111278148553253,
      "phaseShiftNullStandardDeviations": 2.4445413329141155,
      "lesson": "the NG15 spectra already include white noise; adding the TOA uncertainties again inflates the noise model and weakens the statistic"
    },
    {
      "name": "intrinsic red noise ignored",
      "model": "common-process-only",
      "hellingsDownsSignalToNoise": 8.310532066823592,
      "phaseShiftNullStandardDeviation": 7.446047446843209,
      "phaseShiftNullStandardDeviations": 1.1179739153552404,
      "lesson": "the formal S/N looks large, but the phase-shift null is just as wide: unmodelled red noise in a few pulsars makes every pair correlation noisy"
    },
    {
      "name": "timing model not marginalised",
      "model": "spectrum-minus-white-floor without the quadratic projection",
      "hellingsDownsAmplitudeSquared": 3.3096975635695677e-31,
      "hellingsDownsSignalToNoise": 0.718580497813314,
      "lesson": "without the projection the normalisation still counts quadratic-shaped power that the published spin fit already removed, so A^2 drops by about a third while the S/N barely moves"
    },
    {
      "name": "amplitude not reproduced",
      "model": "spectrum-minus-white-floor",
      "recoveredOverPublishedAmplitudeSquared": 0.0850524783919247,
      "phaseShiftNullStandardDeviation": 0.18864094652898744,
      "lesson": "the recovered A^2 is far below the published value and the null spread is far below 1. Both point to a noise model that is too large for the post-fit residuals; the likely cause (power absorbed by the full timing and DMX fit that this tutorial does not model) is not verified here. Quote significance against the null, not the amplitude."
    }
  ]
}
