Prodromos — a $0 pre-flight gate before expensive DFT

πρόδρομος — “the one who runs ahead.”

A DFT nudged-elastic-band (NEB) or saddle-point search can cost days of GPU time and hundreds of dollars. And a dispiriting number of them are doomed before they start.

We learned the failure modes the expensive way:

  • endpoints that quietly relax into the same basin, so the “barrier” is fiction;
  • the wrong structure silently sitting on disk — a 44-hour run on a cell that had lost the very vacancy it was supposed to study;
  • a migrating proton that lands on the wrong host mid-path, so the run measures a different mechanism than intended;
  • endpoints sitting on two different magnetic sheets, making a single-path NEB ill-posed;
  • foundation machine-learning geometries that look fine locally but carry tens of eV of hidden lattice error.

Prodromos is the answer: a $0 pre-flight gate that runs ahead of the costly calculation and reports whether the path is worth taking. It’s an evidence ladder — cheapest test first, stop at the first hard diagnosis — and it is not a replacement for Quantum ESPRESSO / ABACUS / jDFTx, just a thin, cheap layer of checks around them.

It is built on TM-Spec — the “support layer” we said we were starting to build. A case is a TM-Spec document; Prodromos converts a raw QE/ABACUS input into one automatically, and pulls structures and magnetic ground states from OPTIMADE / NOMAD / Materials Project / MAGNDATA to cross-check before a single GPU is touched.

It ships as a CLI and an MCP server, so an AI assistant can call each gate directly and get back a structured verdict — GO, route to a cheaper test, or NO-GO with reasons and next actions. 544 tests; the search-method prototypes are validated on analytic potentials.

MIT-licensed, built for our own Fe–S campaigns and shared in the open.

Prodromos on GitHub →