ADR-040: Public Inspect Scorer for Claim Support¶
Status¶
Proposed (June 2026) — depends on the sandbox evidence slice (ADR-035).
Depends on ADR-051 (contract seam).
Context¶
Assay ingests structured outputs from promptfoo, mastra, pydantic, livekit, openfeature, and cyclonedx through a receipt-importer pattern. The one high-value eval host not yet wired is Inspect (UK AISI and Meridian). As of 2026 Inspect runs arbitrary external agents (Claude Code, Codex CLI, Gemini CLI) as agents-under-test, packages scorers as standard Python packages, and registers community evals through a /register/ folder .yaml submission pointing to external repositories.
Decision¶
Ship the claim-support scorer as a standard Python package and register it through the /register/ flow, not as a fork or a core-repo change. Be a scorer inside Inspect; do not build a competing eval harness. The scorer and any other consumer share the same claim-class contract (ADR-051); the vocabulary is not forked. Because Inspect can drive a coding agent as the agent-under-test, the scorer grades a coding-agent run observed by assay sandbox (ADR-035), giving one end-to-end demo.
Implementation slice¶
Lands as a Python package after the sandbox evidence record is consumable as the scorer's observed-evidence source. This ADR records the decision and the integration shape.
Consequences¶
- Inspect users score claim-support without leaving Inspect, and can score coding-agent runs end to end.
- Adds a Python package to publish and keep in step with the Inspect scorer API and the register flow.
Best-practice basis (2026)¶
- Inspect is the de-facto safety-eval framework; integrate as a first-class scorer via
/register/. Inspect natively drives Claude Code / Codex CLI / Gemini CLI.
References¶
- ADR-051 (contract seam), ADR-035 (sandbox evidence)