From market question to reproducible observation
LeMarché is the deterministic evidence layer. Use it to inspect data and calculations, not to infer a conversational AI answer from a proxy response.
Quick-test runbooks
Use a runbook when you need a result you can reproduce in under ten minutes. Each row names the minimum input, the correct surface, the verification step, and the expected handoff.
| Test | Surface | Minimum input | Verify with | Expected output | Time |
|---|
Curated stack inventory
Start with one job, not the whole file matrix. The names below are the intended routing map for a quick test.
Routes and boundaries
The stack is intentionally layered. Route work by the kind of truth it needs, then hand it to the next layer instead of blending every tool into one opaque answer.
Agent protocol
Every operator should make intent, allowed inputs, evidence, output, uncertainty, and the next action explicit. Use this contract as a prompt prefix or handoff template.
Show the workflow, not the file count
A credible demo proves the operating loop in a few minutes. Keep the first pass focused, visible, and honest about the boundary between generated interpretation and verified evidence.
Best practice checklist
Use these rules before a decision-relevant output is accepted, handed off, or shown in a demo.
Prompt library for the highest-value use cases
Choose a lane, replace the bracketed inputs, and copy the enhanced prompt into the appropriate Prime AI surface. Every prompt asks for bounded output and makes missing evidence visible.