Skip to main content
NiceEval separates post-run data into three layers. Record stores Run and published Attempt facts. Inspection operations close read semantics. Three consumers deliver the results.

​
Record stores facts

Every project has one canonical .niceeval/record.sqlite file. The Runner freezes expected slots when it creates a Run. Each Attempt becomes readable after its own atomic publication transaction completes, even if its origin Run remains active. Published Attempts and terminal Runs are immutable. After an Invocation finishes completely and passes the portable gate, the canonical file itself can be copied or archived. query and show validate an external copy selected by --record <file> completely before opening it read-only.

​
Inspection closes read semantics

An Inspection operation closes its source, selector, sealed cutoff, data gaps, issues, Evidence, and comparison before returning plain data. The fixed catalog covers Runs, Attempts, traces, diffs, sources, artifacts, and comparisons. It does not expose SQL, JSON paths, or formulas.

​
Three consumers deliver results

  • niceeval query is machine Query for AI, scripts, and CI. It outputs a niceeval.query/v1 JSON document.
  • niceeval show is terminal Show for people reviewing and drilling into results quickly.
  • niceeval view is browser View for people using fixed first-party Insight in a protected loopback browser.
All three consumers share the same fixed Inspection result. Each owns its JSON protocol, terminal text layout, or browser UI. They do not share formatters, routes, components, themes, or presentation schemas, and they do not reselect members or recompute business results. Read the CLI reference for commands and Record for external-file validation boundaries.