OctoView LabsQuestions
The questions we get, and the answers we give.
Is Topos a coding assistant?
No. It does not write code, and it does not make AI agents faster or cleverer. It is an operating system for agents, not a tool in the toolchain: it governs whatever agents you use and makes what they did provable. If you want more output, there are many vendors. If you need to answer for the output, there is Topos.
Which AI models or agent frameworks does it require?
None in particular. Topos sits under the agents as substrate: work enters as registered intents, results enter as evidence-bearing claims, and the rules apply regardless of which model produced the work. The industry's agents change monthly; obligations don't.
What do you actually mean by "evidence"?
Records that survive a skeptical reader: a registered intent stating what was attempted and why; a named human signature on every policy act; a completion claim anchored to a content hash; a verification verdict produced by something other than the author; and an append-only, hash-chained ledger holding all of it, from which every view is derived. Not screenshots, not chat logs, not assurances.
Why publish your refusal numbers? It looks bad.
It looks like the product working. When our system refuses machine claims that cannot prove what they assert, each refusal is a hallucination that did not enter the record. An audit trail with no failures in it has been edited. We would rather show you the losses than ask you to trust a record that only says yes. The same goes for the audit of our own system we published rather than buried: read it.
Why don't you name your customers?
Because institutions in regulated markets are entitled to decide for themselves what they disclose about their infrastructure, and that decision is not ours to pre-empt. What we do publish is checkable: the company registration, the founder's record, and the fact that our discipline runs EMIR Refit trade reporting live at a named European trade repository.
Can we run it inside our own perimeter?
Yes. Self-hosted is a first-class deployment, with data residency and retention configured per institution. A managed option exists for institutions that prefer it. In both cases the full audit ledger is yours, behind a typed API your own risk and audit functions can read directly. More on security and data.
What does a design-partner engagement mean?
It means we build against your real regulated workload rather than our assumptions about it, on a months-scale, with your obligations shaping the substrate. It is deliberately not a self-serve product today. That is slower and smaller than a launch; it is also how infrastructure for regulated industries has to be built.
Who is behind this?
A four-person team in Stockholm. The founder spent two decades in tier-1 financial infrastructure and architected the EMIR Refit reporting system live at REGIS-TR. The company is bootstrapped, no outside capital, which means the only people we answer to are our customers and the record. Details on the company page.
Isn't "the AI polices the AI" circular?
It would be, without two things: the enforcement machinery must prove it can catch a planted violation (a gate that cannot fire is theater), and the rules themselves are locked against the agents that work under them. Changing a rule takes a named human's signature, checked mechanically. Verification is machine-scale because it has to be; authority stays human because nothing else can carry accountability. The full doctrine is published here.
What is the status, plainly?
In production use on our own regulated engineering work, under active development, taking design-partner conversations in finance and, at an earlier stage, in law. We publish our progress, including the uncomfortable parts, as we go.