Trust · Responsible AI

Responsible AI, as code

Every AI company has a responsible-AI page. Ours is different in one way: each rule below names where it is enforced in the product, because a rule that lives only in a document is a hope.

Under every chat box in Wavn, in plain sight: "Even the best are not perfect, and neither is AI. Wavn can make mistakes, always double check the responses. We challenge ourselves every day to make fewer of them."

The rules, and where they live

1. Wavn never says it did something it did not do

The single most important rule in the product, above being helpful and above sounding confident. Confirmations are generated from the record of what happened — the schedule row that came back, the file that landed — never from the intention to do it. The part of Wavn that can only talk tells you it can only talk.

2. An unknown is drawn as unknown, never as empty

If Wavn could not read something, the screen says it could not read it. There is no code path that renders "we do not know" as a reassuring blank. A blank that secretly means "unknown" is how people get caught out — and they would never learn Wavn had been guessing.

3. No invented facts on a screen

If Wavn does not know, it says so. Researched work carries real sources. When an answer used your material, the exact passages read are shown on request — and when nothing in your material covered the question, that is what the line says.

4. Consequences are named before you pay them

When switching something off has a cost — like busy-sharing, where turning it off makes you look free rather than hidden — Wavn states the cost in words you can picture, before you flip the switch.

5. Work is checked before it is delivered

Built files get a checking pass in the back room before the card lands. The check does not make Wavn perfect; it makes Wavn's failures honest — a job that fails its check arrives as a failure with its reason.

6. Wavn speaks about itself from a reviewed source

Asked what it is or how a feature works, Wavn answers from a single written, reviewed description of the product — and if it has not been taught the answer, it declines rather than inventing a screen or a feature. The tool that publishes that description refuses to publish it at all if internal machinery has leaked into it.

7. Nothing is sold before it exists

No "coming soon," no dates on unshipped work, no buttons on doors that do not open. This site describes the shipped product, and the release notes are its dated record.

Why this is the whole page

We could add sections about our values. But a user does not experience values — a user experiences whether the confirmation they just read was true. Every rule above is a way of making it true. That is what "responsible" means here.