| """Solomon: a document plus structured questions in, one probability per decision out. |
| |
| from solomon import service, api |
| layer = service.service(store, engine, selection='serving/selection.json') |
| server = api.serve(layer) |
| |
| Reading order. `service` is the serving contract and the layer that reads the model's own letter logits. |
| `binding` is the identity check that refuses to serve a stack that is not the one that was measured, and |
| `calibration` the one positive scalar per answer type it carries. `semantics` and `reliability` are what a |
| probability means here; `readout` turns letter logits into one. `engine` is the served engine, built on |
| `engine_numerics` over `engine_cuda`; `heads` are the trained answer heads it reads through. |
| `engine_contract` and `engine_reference` hold the prompt blocks every answer type is asked with. The |
| `service_*` modules are the pinned chain the layer wraps, innermost first: `service_states`, |
| `service_checked`, `service_answers`, `service_heads`, `service_evidence`, `service_packages`, `serving`. |
| |
| Every answer type is answered. There is no abstention on this path; see `service.ORDERING_DISCLOSURE`. |
| """ |
|
|
