[chrisjleal]

Requested by chris · project thread

User-facing change

On Evidence:

• One AgentGovBench chart, not two. The page showed two identical giant "48/48" charts one after the other. The scoreboard chart (Claude Code, Codex and Cursor beside Proofable) stays. It is the same file the launch scoreboard post exports. The current-library chart is removed. Its numbers (ACP 46, allow-everything 20, no governance 5) remain in the caption and in the reviewer table. The caption keeps both footnotes together: the agent-tool rows come from the earlier library, and on the current library Proofable handles 48 and ACP 46.
• Every result is on the page itself. 8 of 8 live tests and 150 of 150 forged records caught were only inside collapsed sections, even though the launch posts send readers to this page for them. Each is now one bold line under the chart.
• The prompt injection test result. One more line in the same shape: 6 of 6 data-theft attempts refused under a scoped permission, against 6 of 6 let through by an all-purpose key, with its scope sentence ("Self-run: 84 runs…"). It is quoted word for word from LAUNCH.md section 2 (_workspace #16) and links to the test kit in proofable/mcp. Nothing on the page says a scoped permission keeps private data in.
• Charts read on a phone. At phone width the chart labels rendered at about 7 px. Each chart now has a stacked variant that shows below 640 px, with the same text, fonts and colours. The wide charts are unchanged.
• Less repetition. The intro is cut to two plain sentences. The sales paragraph inside the AgentGovBench details is reduced to its one fact, the scenario ACP declines. The page ends on one next step: add Proofable to your AI tool.
• Access rule. The eight-test rerun now says "any account you control" instead of "a Pro account". Wording supplied by the freeze thread.

No values changed. Every number matches docs #8 and LAUNCH.md section 2.

Changelog

• Updated the Evidence note under ## [Unreleased] in CHANGELOG.md so it describes the page as it now opens.

How you checked it

• npm run validate passes: markdownlint, routes, copy, public-surface gate and OpenAPI lint.
• Rendered locally with mint dev (4.2.108) at 1440 px and 390 px, in light and dark. On each, exactly one chart variant per chart is visible (wide from 640 px, stacked below), and the stacked labels read at about 13 px on a phone.
• The injection test numbers were checked against the run record in _workspace #14 (it-20261020183741: 84 runs, 0 of 36 against 14 of 36 for using a tool outside the task).

🤖 Generated with Claude Code

https://claude.ai/code/session_01W1oC33PdgQGS31dEx3xcKB