A Second Cosmology Experiment: Chasing the Cosmic Dipole
After the Hawking points experiment, I wanted to see how the same setup handles a bigger, messier question, one where the experts genuinely disagree. Same deal as before: Claude Code does the work, I direct it, everything is public: github.com/EdonZo/cosmic-dipole.
Same caveat too. I’m a software engineer, not a cosmologist. This is an exploration of what the tool can do, not a contribution to physics.
The question
Our galaxy is moving through space, and that motion should show up as a tiny lopsidedness in the sky: slightly more distant galaxies visible ahead of us than behind us. Astronomers can predict exactly how lopsided the sky should look. The puzzle is that when several teams measured it, the lopsidedness came out two to three times bigger than the prediction. If that’s real, one of cosmology’s founding assumptions, that the universe looks basically the same in every direction, has a crack in it. If it isn’t, something in the catalogs is fooling everyone. I had Claude Code try to find out which.
Not fooling yourself, again
Same playbook as last time: write down every rule before looking at the data, what counts as a find and what counts as nothing, so there’s no room to bend the answer afterward, and rehearse everything on thousands of simulated skies first. This time it ran much bigger: over twenty working sessions and three completely different sky surveys. One sees infrared light, one radio waves, and one is a brand-new map of quasars from the DESI telescope.
What the first two surveys said
Both showed the lopsidedness, bigger than our motion can explain. The convincing part is that they point the same way. These are different instruments catching different kinds of light, with completely different flaws. If the signal were a glitch, the two should disagree about its direction. They agree, within the margin we set in advance.
Still, by the rules we wrote down, this is not a discovery. One survey is too small to be trusted on its own, and the other lands just under the bar physics uses for discovery claims. So the verdict is “promising, but unproven,” and that’s the label it keeps.
The instrument that caught itself
The best story came last. The plan for the DESI survey was published openly, timestamped in the public record, before downloading a single byte of its data, so the test couldn’t be shaped after seeing the answer. And when the answer came, it pointed the wrong way. It looked like the whole idea was about to fall apart.
Then the tooling did what it did in the Hawking points run: it tested itself. It hid fake signals in the data to see whether they would survive the survey’s standard clean-up step. They didn’t. That clean-up was quietly erasing up to a third of any real signal, in exactly the direction that mattered. A measurement taken through that filter can’t prove anything, for or against. Two separate reviewer agents were then given opposite jobs, one to defend the bad news and one to tear it apart, and both reached the same conclusion: this round is a no-contest.
The result
The verdict stays “promising, but unproven.” What I took away is the pattern: the tooling refused to overclaim twice, in opposite directions. It wouldn’t call a discovery when things looked good, and it wouldn’t call a defeat when things looked bad, because both times its own checks found the reason to hold back.
Everything is in the public repo, and before writing this post I had one more independent agent double-check every claim above against the project’s records. I’ll keep running experiments like this.