Products
Intelligence for onchain work is not a lecture. It is traces, rubrics, and a place to act. Assay encodes that spectrum. Computer-use comes later.
A fluent paragraph about volume is not a trace. A trace is the steps: what the expert looked at, what they discarded, what they counted, and what they were willing to write down. We capture that on live Base cases — wash, inventory, launches, incidents, signing — and ship it as SFT.
An agent does not pass because the answer looks fluent. It passes a named expert rubric. Each criterion has a weight, a pass line, and a fail example. Rubrics are the reward surface. They are also how we refuse to grade vibes.
Training data without a place to act is a reading comprehension test. Assay environments give the agent a wallet it can inspect, an explorer, SQL over decoded Base schemas, and MCP tools. Default unit is USDC. Default canary network is Base.
Later. Human demonstrations of explorer, console, and incident-response work. We will film how the expert actually clicks when the feed is stale. Not in the first pack.