SIRA

SIRA is an AI astronomer for new transients in the sky, like supernovae and gravitational lenses. It predicts how each object will change before the data come in, and every prediction can be checked later. Every night it takes new candidates from the ZTF survey. Muse Spark Open reads the first images and decides if the source is real. Muse Spark agents propose and argue over what it could be, and a physics library fits each idea to the light curve. SIRA then seals a forecast, hash-chained and timestamped by FreeTSA and on Bitcoin, so it can’t be changed after the fact. When the new data don’t match, Muse Wonder proposes the missing physics and SIRA tests it on measurements it hasn’t seen. We fixed the evaluation plan before running it. On 37 objects SIRA had never seen, it gave about 2.7 times more probability than Muse alone to what actually happened, and 1.35 times more than its own physics. Its surprise alarm went off on 10 of 112 unseen objects, all lenses or rare transients, and on none of 43 ordinary supernovae. With no lens model to work from, Muse Wonder wrote Einstein’s 1936 lens law on 2 of 6 lensing events, but overall its own equations won less often than building from known pieces. We published that too, along with a false discovery. The site, code, every sealed forecast and every number are public, and anyone can check and reproduce them without API keys. You can also watch SIRA investigate a candidate live, or play against it in Beat SIRA.