Every how-to page teaches one task. This area is where the facts and judgment calls behind those tasks live — written once, so a how-to page can link here instead of re-explaining them each time.
Foundations covers what AQVEN is actually built on, what surviving a process crash really means, and
how a project’s files sit on disk. How a step runs goes one level deeper into the mechanics of a
model call: the three real files behind one llm node, the ten node kinds as a map, when to reach for
which of the three prompt levels, the five cases of a dynamic input or output shape, and what happens —
outcomes, retries, redaction — every time a model is called. Debugging and changing a project covers
the engineering loop from an unexpected result to a verified fix, how to address the exact step that went
wrong inside a parallel branch or a loop, the two ways to change a project’s files, and how to decide
whether a step should be split, diverge, or get a critic loop in the first place. Experiments covers
how a flow proves it works: experiments, series and findings, working and held-out cases, what a series
costs, and how a series turns intervals and margins into a verdict.
- What this is built on
- A run survives a process crash
- Files as source of truth
- Agent, Inference, and the llm node
- Ten kinds of nodes
- Three prompt levels
- Five cases of dynamic input and output shape
- What happens when a model is called
- The engineering loop: from incident to verified fix
- Finding the node where a workflow went wrong
- Two ways to change a project
- Designing reliable workflows
- Experiments, series and findings
- How a series decides