Phase 1 notes#

Done#

  • composite_step.recipes: the YAML format and its checks; the combine language, parsed with ast and evaluated over a whitelist (never eval) with cbs_power, hf_extrapolate_exp (G4 and Karton-Martin form), schwenke, scale, hf_part and corr_part; the library of shipped, user and job recipes, in which a local recipe may not shadow a shipped name and version; the capability check; fingerprints (the descriptive fields left out); and the thermal terms, H0 = E0 + s ZPE and H = E0 + s ZPE + [H - E - ZPE] at the frequency level.

  • The recipes r2SCAN-3c-HO (r2SCAN-3c) and revDSD-r2SCAN-3c-HO (r2SCAN-3c // revDSD-PBEP86-D4(2021)/def2-TZVPP), ZPE x 0.9899 pinned with its source; HO = the scaled harmonic thermal terms. They are what the thermochemistry step’s quick and default presets run today. Tests check that the factor is still seamm_thermochemistry’s, and that the recipes reproduce the thermochemistry step’s arithmetic bit for bit. Recipes are named by their content, not by a preset’s role (Paul, 2026-11-11), so a stored result’s label stays true when the presets move on.

  • CompositeStep.get_model_chemistry_options: SEAMM:CMP@<recipe> for the newest version of each recipe, with its geometry level, components and ZPE factor in the option. Whether the components’ codes are installed is not checked there (asking model-chemistry discovery from inside it would recurse); the step checks before it runs, with a matcher that lets the basis be any, as match_model_chemistry does.

The task route (next)#

seamm_exec’s Evaluator takes one task per structure from one provider, and a composite is several single points per structure, each on its component’s own code. For phase 1 the components run as one Evaluator per component, over all the structures at once, through the component’s own model chemistry; the composite then combines the energies per structure. This needs no change to seamm_exec, and it keeps every component an ordinary calculation with its ordinary settings fingerprint, so the fragment store finds and stores components, never composites (design §6), and the component of revDSD-r2SCAN-3c-HO is exactly the default preset’s energy level (science’s pass criterion (a)).

A later change can let the Evaluator take several tasks per structure from one provider (get_tasks / analyze_tasks), so that every consumer of a model chemistry gets composites without knowing about them, and all components run in one task set. That is a seamm_exec change for its own PR.

Pilot (science’s criteria, in order)#

  1. The recipe’s component settings fingerprints equal those the preset path produces, so the same store rows are hit.

  2. H0 and H(298.15 K) reproduce the step’s to round-off on the 20 demo molecules.

  3. Three to five molecules computed fresh through the task route, end to end.

Only today’s presets (from 2026-11-08: VERYTIGHTSCF, DEFGRID3, NoCOSX rules) are a valid reference.