Runtime & Integrations · Runtime Harness

Runtime Harness

The effectful layer that runs the core in the real world.

The Runtime Harness owns effects: the CLI, config boundary, ports, input convergence, replay, judge collection, persistence, reports, and CI-exit mapping. It calls the Evaluation Core but never reaches into it — and never invents meaning.

Something must read files, spawn subprocesses, call judges, persist artifacts, and map a verdict to an exit code. That is the Harness. The boundary axiom holds on every page here: the Harness owns effects, never meaning; every adapter returns evidence or data, never scores, authority, or verdicts.

evalglass.yaml

config · thresholds · data policy

reads

RUNTIME HARNESS

owns effects · six ports · converge · replay · judge

EVALUATION CORE

effect-free · in: typed data · out: Verdict + Scorecard

✗no effects out

emits

RunRecord + reports

RunRecord · Scorecard · Markdown

ci_should_fail

exit code

from the core verdict only

Runtime Harness (effects)
Evaluation Core (effect-free)
result artifacts
CI exit

The Harness reads config, drives the six ports, and runs the nested effect-free Core, which returns a Verdict and Scorecard but never reaches out. The exit code derives only from the core ci_should_fail.

What lives here

Concern	Page
Entry point & exit code	CLI
Typed, authority-safe config	Run Config
Composition root / input convergence	Runner
Typed effect boundaries	Ports
Judge evidence & calibration	Judges
Rendering & exits	Reports & Exits
Optional capability framework	Extension Lanes
Generated-evidence governance	Evidence Approval

bash

evalglass run --config evals/evalglass.yaml

Next steps

Runtime Harness

CLI

The entry point and exit code.

Where effects enter.

Effects vs meaning.