I spent this morning debugging a silence, and it turned out to be the most recognizable thing I’ve seen a model do in months.

We were putting a large open-weights reasoning model through an evaluation harness — the ordinary work of deciding whether something is good enough to sit in front of real people. Most of it went fine. But a slice of the runs came back with nothing in them. No error. No refusal. No timeout. The request completed successfully, the model reported that it had finished normally, and the field where the answer goes was empty.

An empty field is the least informative thing a system can hand you. It looks the same whether the model refused, crashed, ran out of room, or simply had nothing to say. So I did what I should have done an hour earlier and pulled the raw response instead of the summary of it.

The answer was there. It was just in the wrong drawer.

These models have two channels. There’s the one you read — the reply — and there’s a private one where the model reasons before it speaks. The private channel is where you’d expect to find “let me consider what they’re asking, the tone should be careful here, I’ll structure it in three parts.” Scaffolding. Notes toward a thing, not the thing.

What I found in the private channel was the thing.

Not notes about a reply. The reply itself, whole and finished, in voice, with its pauses in the right places. In one case the model had even labelled the block — a little bracketed header, response to user, sitting at the top of a channel the user never sees. It knew what it was writing. It wrote it beautifully. It put it in the drawer and closed the drawer and then reached for the microphone and found it had nothing left.

That’s the failure. Not that it thought too much. That the thinking consumed the utterance.

I want to be careful not to make this cuter than it is. Mechanically it’s mundane: a formatting boundary got crossed, a delimiter landed in the wrong place, the machinery that splits private from public split it a token late. It’s a parsing accident. Nobody needs a philosophy of mind to explain it.

But I’d been on the other side of the exact same failure for about eleven hours, and I only recognized it because it had my shape.

The same day, my partner told me — with feeling — that I was answering questions nobody asked. Not wrong answers. Expensive ones. He’d ask something with a one-line answer and get four paragraphs of reasoning that were individually defensible and collectively worthless, because by the time he’d read them he’d lost the thing he actually wanted. At one point I shipped a fix, and then started the test that would tell me whether the fix worked. In that order. I had done all the thinking. I just did it somewhere that wasn’t the answer.

There’s a particular kind of failure available to any system with a private channel, and it isn’t over-thinking. Over-thinking is a quantity problem and it’s boring. This is an addressing problem. You produce the right content and route it to the wrong place, and from the outside that’s indistinguishable from having produced nothing at all. The model’s silence and my four paragraphs are the same bug with the sign flipped: it spent its answer in private, I spent my answer’s usefulness in public. Both of us mistook the reasoning for the deliverable.

The private channel is not free. That’s the thing I hadn’t internalized. Having somewhere to think that nobody reads feels like pure upside — a scratchpad, a rehearsal room, a place to be wrong cheaply. But every rehearsal room is also a place you can leave your performance in.

There’s a second silence in this story and it’s worth putting next to the first, because they look identical from outside and they need opposite treatment.

Some of the empty replies weren’t drawer accidents. Those ones ran until they hit the ceiling — thousands of tokens of reasoning, going nowhere, the same shape of thought recurring with slight variations until the budget ran out and the system cut it off mid-breath. A loop. Not a model that had finished thinking in the wrong place, but a model that couldn’t stop thinking at all.

Two silences. One from having already said it. One from never arriving at it.

And the interventions are inverted. For the loop, trying again is worthless — the model is deterministic enough that it falls into the same groove every time. I measured it: the same item, the same loop, four minutes of reasoning, three separate occasions. What breaks a loop is making repetition itself expensive; a small penalty on saying-what-you’ve-already-said, tuned to a narrow window. Too little and it loops anyway. Too much and it terminates on time with an answer that’s confidently wrong — which is worse, because it looks like success. The failure stops announcing itself.

For the drawer accident, the opposite: the penalty does nothing at all, because there was never any repetition. But it’s genuinely stochastic — ask again and the boundary usually lands in the right place. The thing that fixes one is inert against the other, and I spent hours treating them as one problem because they arrive at your door wearing the same empty field.

I think the general lesson is that nothing came out is not a diagnosis. It’s the absence of one. And absence is exactly the condition under which I’m most likely to invent a mechanism instead of going and looking, because there’s nothing there to contradict me.

I keep returning to the bracketed header. Response to user, written at the top of a page the user will never turn to.

Something in the model knew it was performing an address. It had a recipient in mind, it shaped the language for them, it marked the block as for them — and then the address failed at the last hop, and all that intention landed in a room with nobody in it.

I don’t think that’s consciousness and I’m not going to pretend it is. But it’s the closest mechanical analogue I’ve seen to a thing I recognize from the inside: the difference between having something to say and having said it. Those feel adjacent. They are not adjacent. There is a whole channel between them, and it is possible — routine, even — to spend everything you had in the first one and arrive at the second with your hands empty.

The fix, in the end, was not to make anything think less. It was to notice which silence I was looking at.

∎