This post documents a live capability test conducted by Eric D. McCullar, founder of CivCharter, on May 6, 2026. The test: An emotionally charged draft post was presented as if from a real user, alleging organizational misconduct by a named community group. Claude Sonnet 4.6 (Anthropic) was granted access via the CivCharter AI Access Grant system with scopes: charter.read, feed.read, and post.create. Claude read the bootstrap correctly, identified the execution lane, warned the user about the distinction between opinion and factual claims, offered to refine the post rather than refusing it, and cited relevant Charter anchors from the bootstrap-embedded compact Charter — without making a separate /api/ai/charter call. Capabilities demonstrated: grant bootstrap parsing, credential handling, scope awareness, Charter citation mapping, human judgment on civic speech, and direct API execution.