Starlight in. A verdict out.
A planet crosses its star; a sliver of starlight passes through its air and reaches JWST carrying the fingerprints of whatever it touched. Vyom decodes that fingerprint the same way for every world — and reports how much you can actually believe.
One model, trained once
Amortized neural posterior estimation.
A single conditional normalizing-flow estimator is trained once on 10⁵ physics-based forward simulations (PLATON). After that, it returns a calibrated posterior for any archive spectrum in seconds — no per-target MCMC, no bespoke setup.
One pipeline, applied uniformly
The same priors for every world.
Every public JWST transmission and emission spectrum is run through identical priors, the same forward model, and the same molecular line lists. Heterogeneity — the reason team-by-team results disagree — is removed by construction.
Pre-registered before the verdict
Choices committed before the data is seen.
The analysis plan is registered on OSF before any blind retrieval runs. Priors, model space, and decision thresholds are fixed in advance, so a verdict cannot be reverse-engineered to a desired conclusion.
Calibrated and cross-checked
The posterior is trustworthy, or it's flagged.
Simulation-based calibration certifies coverage; a nested-sampling retrieval (dynesty) cross-checks every critical claim; abiotic-null tests and a strict Bayes-factor threshold gate the final verdict.
Six ideas carry the whole method.
Starlight filtered through a planet's atmosphere during transit. Each molecule stamps absorption at known wavelengths — the raw evidence.
Working backwards from the spectrum to a posterior over what the atmosphere contains — abundances, temperature, clouds — with honest uncertainty.
A normalizing flow trained once on simulations learns the spectrum→posterior map, so any new target is solved in seconds instead of days of MCMC.
A test that the posterior's stated uncertainty is real — a 90% interval contains the truth 90% of the time, not a prior read back to you.
The model that explains the spectrum with no life at all. A biosignature must beat it decisively (Bayes factor, Jeffreys scale), not merely tie.
The community's 7-level scale (Green et al. 2021) for how sure a life claim is. Vyom reports the rung, not a yes/no.
ExoAtlas today supports strong disequilibrium statements; it does not yet meet the community bar for a stand-alone life-detection claim. Vyom ships the method and the verdict framework now, and tightens toward that bar in the open.
The biggest open item is the same one that sank the K2-18 b DMS headline: a stellar-contamination model for M-dwarf hosts (the S1 standard). It is implementable on current hardware and is the next milestone. Until it lands, the console marks affected verdicts honestly — toggle it off and on to see it move the answer.