This version keeps only the core evidence-backed harness evolution loop. Its three stacked rows are intended to stay readable in GitHub Markdown and translate cleanly into a roughly 4:3 graphic.
flowchart TB
subgraph OBSERVE["1. Observe"]
direction LR
G["Goal + environment"] --> A["Attempt real task"] --> I["Capture + inspect evidence"]
end
subgraph IMPROVE["2. Improve"]
direction LR
D["Trace likely failure"] --> C["Create candidate"] --> T["Retest + regressions"]
end
subgraph PROMOTE["3. Decide"]
direction LR
Q{"Improvement proven?"} -->|Yes| P["Promote version"] --> X["Continue, export, or package"]
Q -->|No| R["Revise candidate<br/>and repeat"]
end
I --> D
T --> Q
The promoted harness remains unchanged until the evidence gate confirms an improvement. A candidate may batch related changes, and independent candidates may coexist. Validation scales with impact; a rejected candidate returns to diagnosis and revision without replacing the current version.
See How Harneloop Works for the full architecture diagram.
The raw Mermaid source is available at docs/diagrams/framework-process-compact.mmd.