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
"""Deterministic, model-free Python float-repr oracle for Rust's JSON formatter.
Store IEEE-754 bits, not JSON numbers: parsing a decimal must not change the
input float before the differential test reaches PythonJson::repr.
"""
import json
import math
import os
import random
import struct
def build_repr_probe():
bits = {0x43179085685D83C9}
def add(value):
bits.add(struct.unpack(">Q", struct.pack(">d", value))[0])
for exponent in range(50, 54):
for quarter in range(-16, 33):
value = float(2 ** exponent) + quarter / 4
add(value)
add(-value)
for raw in [0, 1, 2, 3, 15, 0x000FFFFFFFFFFFFE, 0x000FFFFFFFFFFFFF,
0x0010000000000000, 0x0010000000000001, 0x7FEFFFFFFFFFFFFF]:
bits.add(raw)
bits.add(raw | (1 << 63))
for value in [1e-5, 1e-4, 1e16, 1e22]:
for neighbour in [math.nextafter(value, -math.inf), value,
math.nextafter(value, math.inf)]:
add(neighbour)
add(-neighbour)
rng = random.Random("laya-repr-v1")
accepted = 0
while accepted < 2000:
raw = rng.getrandbits(64)
if (raw >> 52) & 0x7FF == 0x7FF:
continue
bits.add(raw)
accepted += 1
return [
{"bits": "%016x" % raw,
"repr": repr(struct.unpack(">d", struct.pack(">Q", raw))[0])}
for raw in sorted(bits)
]
def main():
root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
probe = build_repr_probe()
for checkpoint in ["english", "multilingual", "typed-decisions"]:
out = os.path.join(root, "tests", "golden", checkpoint)
os.makedirs(out, exist_ok=True)
with open(os.path.join(out, "repr_probe.json"), "w", encoding="utf-8", newline="\n") as f:
json.dump(probe, f, indent=2)
f.write("\n")
print("recorded %d finite float repr probes per checkpoint" % len(probe))
if __name__ == "__main__":
main()