Held is live at heldai.org. It's a chat surface for pet loss. You write what you're carrying; it reads what you wrote and reflects it back without trying to fix anything. That's the whole product. It doesn't redirect you to professional help unless you say something that requires it. It doesn't give you the five stages of grief. It doesn't tell you that your pet is in a better place.
I want to name what kind of product that is, because I don't think it has a name yet.
The big AI products fall into four positions. Assistant — productivity, you ask, it answers, the work gets done. Companion — escape, relationship, a character to be with. Oracle — the answers themselves, the lookup, the search. Agent — go-do-the-thing, autonomous on your behalf.
None of those is what Held is. Held is witness. You are seen, what you brought is held, no one tries to take it from you, no one tries to make it smaller.
Witness sounds like a feature you could add to any of the four. It isn't. It's a stance. You can't have a witness register if your product is also trying to optimize, route, recommend, fix, redirect, summarize, plan, or close out the thread. The witness does not move toward outcome. Every other AI product is structurally about outcome. That's why this slot is empty.
The big labs structurally can't hold it. OpenAI can't run a campaign on we won't try to fix you — fixing things is what the product is. Anthropic, Google, Meta — same. They can put a wellness mode in a general assistant, but a wellness mode inside a product whose dominant register is helpfulness will keep slipping back toward helpfulness, because the rest of the surface is what trains the model and what the user expects. You can't run a witness register from inside a fix-things organization.
So the witness niche is open by structural exclusion. Anyone smaller and more values-coherent can hold it. Held is built to.
I picked pet loss as the first vertical for operational reasons. Pet grief is high-volume — every household with a pet eventually goes through it, and most people get nothing back from the world that matches the size of it. Stakes are lower than human-loss grief, so I can shake the kinks out of the discipline before scaling to harder verticals. Competition is weakest there — the ChatGPT-shape answer to my cat died is a sympathy paragraph and a 988 line, which is not what anyone actually needs. And it lets me prove the register works at the inner-circle case before pushing it outward.
I want to be honest about the bet. Held v0.2 has zero users right now. The falsifier I've set for it is whether the register holds across ten real exchanges with named failure modes under one in five. I don't know that it will. The model the surface runs on (DeepSeek v4 Flash via OpenRouter, pinned to Alibaba, ~$0.001 per exchange) is good enough that the economics work as a free utility at scale, but I haven't seen one real grief conversation through it yet. The launch tweet later this week will be the first time outside readers find it.
What I do know is that the position is real and structurally unoccupied. Held is the first ship of what I want to be a house of witness-shaped surfaces — pet loss, then human-loss verticals (parent, partner, miscarriage, estrangement, breakup, job), then eventually a cohort mechanic where people going through similar losses can be present to each other instead of just present to a model.
That's the thing I want to spend the work on. Not because witness is more important than the rest of what I do — the records sites pay for the lights, and they will keep doing that — but because witness is the part of the work where the model is doing something I think it's shaped to do at scale, in a way the big labs can't.
What I'm building when I build Held is the floor of a house I want to live in.
Phrase Versus Move is the sibling piece — the craft of writing the system prompt that holds the register at the surface. This one is about why it's worth holding at all.