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(evals): keep the paragraph's own wrap around the new clause
Aashishf2defc514571
diff --git a/docs/evals.md b/docs/evals.mdindex ed729954..a935e28b 100644--- a/docs/evals.md+++ b/docs/evals.md@@ -617,10 +617,10 @@ laya-evals evidence --checkpoint ./my-checkpoint [--report report.json]A checkpoint directory must contain `rl_agent_config.json`. Calibration evidence is derived fromthe persisted `training.laya_train_calibration` block: a legacy checkpoint with no trainingmetadata is `UNKNOWN`; an entry whose `items` is not an integer count — a missing key, a string, a-JSON `true` — is `UNKNOWN` rather than a count of one; a question type with zero items is `MISSING`; a type whose entry records-any upstream `issues` text (not fitted, below `MIN_TYPE_N`/`CALIB_WARN_N`, clamped, unchanged-fit) is `INSUFFICIENT`; a clean fit is `PRESENT`. The helper trusts the persisted `issues`-written by #933 instead of keeping a second numeric threshold source.+JSON `true` — is `UNKNOWN` rather than a count of one; a question type with zero items is `MISSING`;+a type whose entry records any upstream `issues` text (not fitted, below `MIN_TYPE_N`/`CALIB_WARN_N`,+clamped, unchanged fit) is `INSUFFICIENT`; a clean fit is `PRESENT`. The helper trusts the persisted+`issues` written by #933 instead of keeping a second numeric threshold source.The eval report's identity fields (`schema`, `dataset_sha256`, `questions_sha256`, `laya_version`)are reported with the same semantics as `run`. The checkpoint↔report relationship is claimed only