Runtime & Integrations · Evaluation Core

Evaluation Core

The effect-free centre that turns evidence into a verdict.

The Evaluation Core is the standard-library-only layer that holds all meaning: contracts, score semantics, aggregation, provenance, authority, and the Verdict Engine. It performs no effects — no file or network I/O, no clocks, randomness, subprocesses, or model calls.

Keeping meaning effect-free is what lets a green result never outrun its earned authority. The core receives typed data the Runtime Harness already collected, and returns a typed RunRecord. It cannot reach out to inflate a score, and nothing downstream can recompute its verdict.

What the core owns

StagePage
Input shapes & fail-closed parsingCore Contracts
Value vs meaning of a resultScore Semantics
Choosing, declaring, implementing metricsMetrics
Honest reduction of per-example scoresAggregation
Fingerprints & regression claimsProvenance & Baselines
May this metric gate?Authority Resolution
The single verdict pathThe Verdict Engine
Composing the whole pipelineRun Evaluation
The primary artifactsScorecard · RunRecord

Allowed tools

The core uses only dataclasses, enum, typing, json, hashlib, and (where exact threshold arithmetic needs it) decimal. No third-party validation, config, or reporting libraries; no vendor SDKs. Isolation is enforced in CI by tools/check_core_isolation.py.

Use it directly A vendored host or test calls from evalglass.core import run_evaluation. The Runtime Harness exists only to build that function’s inputs (plans, examples, evidence, dimensions) and to act on its output.

Next steps