Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(sdk): say which usage keys a caller can read and why they exist
Aashishb32f0e599ada
diff --git a/sdk/typescript/README.md b/sdk/typescript/README.mdindex 4fa747c5..c28f6079 100644--- a/sdk/typescript/README.md+++ b/sdk/typescript/README.md@@ -119,6 +119,8 @@ result.answers.department.probabilities.billing;result.answers.department.confidence;result.answers.department.action?.act_probability; // Laya-only metadataresult.usage.input_tokens;+result.usage.truncated; // boolean: the state did not fit the room left for it+result.usage.truncated_questions; // string[]: the questions whose own window was cut```Choice criteria also accept an array of unique labels, serialized as an option-to-null@@ -132,6 +134,15 @@ Response fields retain Python's spelling (`input_tokens`, `act_probability`,etc.). Scores use zero-based rubric levels. Confidence and accuracy have thesame calibration limits as the Python model; see the root README.+`usage` is the whole record `/v1/systemone` reports: `input_tokens`,+`output_tokens` (always 0), `state_tokens`, `state_tokens_dropped`,+`truncated` and `truncated_questions`. How much of a state the model reads is a+token budget that moves with `max_len`, `head_max_len` and each question's own+option prompt, so these are the only fields that show a cut: the answer itself+is unchanged when evidence is dropped. `options` is added only when some+question's options no longer have a token span each. See the `usage` table in+`docs/http-api.md` for what each key measures.+## Local model selectionBy default, `laya-client` sends no `model` property. `laya-serve` then chooses