The context window is not the bottleneck

Everyone is buying tokens when what they need is a filing system. The recurring failure is not that the model lacked the document. It is that the model had the right document in context and still reached for the wrong one.

The bottleneck is not how much a model can hold. It is whether anything in the pipeline can say I don't know loudly enough to stop the run. A retrieval step that returns nothing is more useful than one that returns eight chunks of near-duplicate prose, because nothing is a signal a program can branch on.

So the interesting work sits a layer down: intent records, resumable cursors, an evidence ledger the model has to cite from. Boring plumbing, and it is where the reliability actually comes from.

The cheap test, which costs nothing to run: log every case where the model was confident and wrong, then check whether the evidence was present in context. If it was, more context is not the fix.

That share — confident, wrong, and holding the evidence the whole time — is the only number in this discipline I would trust as a measure of progress. It is also the one nobody publishes, which tells you roughly what it looks like.