Architecture

One request in, one report out. What happens between is a pipeline with clearly separated stages, and the separation is what keeps generated prose from contaminating observed fact.

The stages

1 · Routing

The request is classified. A 42-character hex string is an address; text is read to work out what kind of subject it names.

2 · Planning

A workflow is resolved against what is actually available. A step whose source is unreachable is skipped and recorded as skipped — not silently omitted.

3 · Execution

Independent steps run together, dependent ones after. A failing step does not fail the task; it produces an unknown.

4 · Evidence

Each observation is stored with its source URL and its state — confirmed, refuted, or unknown.

5 · Scoring

Pure functions over the collected facts. The same rules judge a token analysed alone and a token screened inside a portfolio.

6 · Summary

Written last, over findings that already exist. It can be absent — and says so — without invalidating anything above it.

Agents

An agent is a specialist: one gathers market data, another screens contract powers, another reads a project's public presence. They are independent by design, so losing one costs you that agent's findings and nothing else.

Confidence

A report's confidence is derived from what actually ran rather than asserted. When nothing could be established it is reported as unknown rather than low — low implies a weak answer, and there was no answer.