Using LLMs as peer-review devil’s advocates for research findings can’t hurt and can only help strengthen an argument. One of the common criticisms an LLM makes is that it will claim that a factor is “too weak” or insignificant to serve as a mechanism for some observed behavior. This has almost turned into a cliche since it caters to the consensus bias of “certainly if it was a real mechanism, someone would have noticed it by now“. Certainly, at least in climate science, the notion of weak factors that turn out to have a significant impact is well known. Consider CO2, which by density is a very weak constituent, yet it has an over-sized impact on radiative energy balance. Also, in the context of climate science, when one considers how often the butterfly effect is invoked, whereby a butterfly flapping its winds can initiate a hurricane down the road, puts to test the argument that any half-way potentially significant factor can be easily dismissed.

That brings one to tidal (lunar and solar) forcings as mechanisms for behaviors, beyond that of the well-known daily ocean tides. As I have invoked lunar forcing as a causal synchronization to QBO, the Chandler wobble, and ENSO and others here, an almost guaranteed response by an LLM is that tidal factors are too weak to be considered. That’s actually a challenging devil’s advocate proposition to address, since (a) controlled experiments aren’t possible to generate sensitivity numbers and (b) that there are many subtle ways that a forcing signal can be amplified without knowing which one is valid. For example, a weak yet incessantly periodic signal can build over time and overpower some stronger yet more erratic signal.

Another devil’s advocate argument that an LLM will bring up is the idea of fortuity and chance, in the sense that a numerical agreement can be merely a coincidence, or as a product of fiddling with the numbers until you find what you are looking for. As an antidote to this, an LLM will recommend that other reinforcing matches or spectral details be revealed to overcome the statistical odds of agreement by chance.

For the Chandler Wobble, an LLM may declare the 433-day cycle agreeing with an aliased lunar draconic period of 27.212/2 days to be a coincidence and dismiss it as such (since it is but a single value). Yet, if one looks at the detailed spectrum of the Earth’s orientation data (via X or Y polar position), one can see other values that – though much weaker – are also exact matches to what should be expected. So that, in the chart below, the spectral location for the 27.5545 lunar anomalistic is also shown to match — labeled Mm and Mm2 (for the weaker 1st harmonic). Other sub-bands of the draconic period set are shown as Drac2.

Importantly, the other well-known lunar tropical cycle of 27.326 days is not observed, because as I have shown elsewhere, it is not allowed via group theory for a wavenumber=0 behavior such as the Chandler Wobble (or QBO). In quantum physics, these are known as selection rules and are as important for excluding a match as they are for finding a match. The 27.554 day period is allowed so the fact that it matches to the spectra is strong substantiating evidence for a lunar forced mechanism.

For another class of criticism, an LLM may suggest that further matches in phase coherence of a waveform are required when matching to a model. This is rationalized as a means to avoid fortuitous matching of a simple sinusoidal wave.

For the QBO, detailed idiosyncratic phase details that arise from the lunar forcing model are straightforward to demonstrate via the time-series itself. A typical trace of the 30 hPA QBO time-series shows squared-off cycles that have characteristic shoulders or sub-plateaus that show up erratically dispersed within the approximately 28-month period. This is shown in the chart below, whereby though not perfectly matching, this characteristic is obvious in both the model and monthly data. The reason that this happens is the result of a stroboscopic-pulsed forcing creating a jagged sample-and-hole squared response. (A minimal lag of 1^st or 2^nd order will round the sharp edges.) Furthermore, the same draconic and anomalistic lunar periods contribute here as with the Chandler wobble model, substantiating the parsimonious aspects.

Importantly, this isn’t known to occur in a resonantly amplified system with a natural response, whereby the waves are invariably well-rounded sinusoidal cycles without this jagged erratic shape. This is actually an acid test for characterizing time-series, with features that anyone experienced with signal processing can appreciate.

---

This addresses some of the criticisms revealed when I prompted an LLM peer-review in a previous post, describing findings made in Mathematical Geoenergy (Wiley, 2019)