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
fix(onnx): catch an in-place question rewrite in the scan-budget guard
sameedkhan17e0e655929368
diff --git a/laya/onnx_agent.py b/laya/onnx_agent.pyindex 5ca5fb93..27ab5c4d 100644--- a/laya/onnx_agent.py+++ b/laya/onnx_agent.py@@ -1,3 +1,4 @@+import copyimport jsonimport osimport threading@@ -476,6 +477,14 @@ class ONNXAgent(HookRegistry):internal = {qid: self._to_internal(questions[qid]) for qid in ids}budget, step_default, _ = window_budget(self.tok, [internal[qid] for qid in ids], max_len,head_max_len, window=window, stride=stride)+ # Snapshot the questions the scan was just sized against, BEFORE any start hook can+ # rewrite them, as the torch `predict_long` does (5f29170). `==` over the caller's mapping+ # cannot see an in-place rewrite: a hook that adds options to `ctx.questions[q]["criteria"]`+ # -- the `widen_for_high_cardinality` pattern in `docs/hooks/patterns.md` -- mutates the+ # same nested dict, so both sides change together, `_check_scan_budget` returns early, and+ # the windows are re-truncated while `usage["windows"]` reports them all read. A shallow+ # copy shares the nested dicts and misses it exactly the same way.+ asked = copy.deepcopy(questions)state_ids = encode_text(self.tok,@@ -492,7 +501,7 @@ class ONNXAgent(HookRegistry):# fit one window was silently truncated by a re-budgeting hook and still reported# `windows: 1`, i.e. "the model read all of it" -- measured, 138 of 240 state tokens# never reached the model, while a longer document on the identical input hard-failed.- _check_scan_budget(self, evidence, budget, max_len, head_max_len, questions)+ _check_scan_budget(self, evidence, budget, max_len, head_max_len, asked)single["usage"] = {**(single.get("usage") or {}), "windows": 0 if evidence["answered"] else 1}return single@@ -521,7 +530,7 @@ class ONNXAgent(HookRegistry):# `ctx.max_len`/`ctx.head_max_len` identically), so leaving it off here meant the bug was# fully live on this path while the other agent refused the identical input -- measured,# 34.6% of a document reaching no model.- _check_scan_budget(self, evidence, budget, max_len, head_max_len, questions)+ _check_scan_budget(self, evidence, budget, max_len, head_max_len, asked)if evidence["answered"]:# A hook answered the document before any window was scored: pass that answerdiff --git a/tests/test_onnx_long.py b/tests/test_onnx_long.pyindex 6c5b5de8..c5f55826 100644--- a/tests/test_onnx_long.py+++ b/tests/test_onnx_long.py@@ -333,6 +333,38 @@ check("questions/no-op preserves the unhooked result",_bare_onnx().predict_long(LONG_STATE, QUESTIONS))+# A hook that widens a question IN PLACE must be refused here exactly as on the torch agent+# (tests/test_predict_long.py, "an in-place question rewrite is refused"). Compared against the+# caller's own mapping it cannot be seen: the hook mutates the same nested dict, so+# `questions == asked` stays True and the scan proceeds with windows `build_sequence` re-truncates.+# Measured on this fixture, 8 added options cut the room from 28 to 12 state tokens: all 14 windows+# were truncated and 32 of 200 tokens reached no model, and a 24-token state that fit one window+# lost 12 while reporting `windows: 1`. The guard's own message is asserted, because too many+# options for max_len raise a different ValueError from `_encode_state` that would pass a bare+# `check_raises`. Fresh mappings per call, because the hook mutates what it is handed.+def _inplace_questions():+ return {"dept": {"type": "choice", "instructions": "?", "criteria": {"a": "x", "b": "y"}},+ "urgent": dict(QUESTIONS["urgent"])}+++def _widen_in_place(ctx):+ ctx.questions["dept"]["criteria"].update({"opt%02d" % i: "d" * 20 for i in range(8)})+++for name, inplace_state in (("a windowed state", LONG_STATE),+ ("a one-window state", LONG_STATE[:24])):+ try:+ with warnings.catch_warnings():+ warnings.simplefilter("ignore")+ _bare_onnx().predict_long(inplace_state, _inplace_questions(),+ on_predict_start=_widen_in_place)+ except ValueError as exc:+ check_true("questions/an in-place widening rewrite is refused on %s" % name,+ "after the start hooks ran" in str(exc), repr(exc))+ else:+ FAIL.append("questions/an in-place widening rewrite is refused on %s: did not raise" % name)++def _annotate_first_window(ctx):ctx.results[0]["answers"]["review"] = {"type": "noul", "noul": 0.9, "answer_confidence": 0.9}diff --git a/tests/test_predict_long.py b/tests/test_predict_long.pyindex 7199efc2..d2986620 100644--- a/tests/test_predict_long.py+++ b/tests/test_predict_long.py@@ -1050,7 +1050,7 @@ check_raises("questions/the torch scan validates before _to_internal", ValueErro# says the contract binds both agents._onnx_src = inspect.getsource(ONNXAgent.predict_long)check_true("hook budget/the ONNX scan checks its budget too",- "_check_scan_budget(self, evidence, budget, max_len, head_max_len, questions)" in _onnx_src,+ "_check_scan_budget(self, evidence, budget, max_len, head_max_len, asked)" in _onnx_src,"")check("hook budget/both agents check it on the single-window path too",(inspect.getsource(Agent.predict_long).count("_check_scan_budget("),