chunklaya: why Laya stops reading after ~200 tokens, and a chunking harness that fixes it

# chunklaya — unbounded-context System 1

**A decision model with no context limit. Tested at 1,000,000 tokens.**

Long-input harness for [Laya](https://github.com/NandhaKishorM/laya), the open-weights System 1
decision model: typed questions — `noul`, `choice`, `score` — answered over one piece of text in a
single forward pass, no autoregression. It's the open answer to TypeSafe's closed Jev.

Nothing in the harness scales with document length: chunk, locate the passage the question is about,
and spend exactly one forward pass on it. On the public Jev needle benchmark that holds from 24k
tokens to 1M — **AUC 0.993 at every size, unchanged to three decimals**, one passage scored, 0.9 s for
a million-token document and 15 ms per question after it's indexed. Jev itself stops at 32k.

The honest boundary, measured in the same runs: that's a *lookup* — a question naming something you
can search for. A *categorical* question ("is there any sports story in here?") has nothing to rank
on, so every passage gets scored, and the answer is then only as good as the detector's
false-positive rate times the number of passages — good to a few hundred passages, decaying after
(0.985 at 70, 0.785 at 5600). Both curves are below, with the things that didn't fix the second one.

Laya's checkpoints are configured for 512–1024 tokens, and Jev reads 32k. So what happens if you just
feed Laya more text? Not what you'd guess. The encoder handles long input fine — it's pretrained to
8192 tokens. The problem is the decision head sitting on top of it, and it's a weirder failure than
"runs out of context": past roughly the first 200 tokens, it stops paying attention to anything you
give it. A fact buried in the middle of a long document is effectively invisible to it.

This repo measures that fail