# ArcKit v6.21.0: A Second Look From Someone Who Wasn't in the Room **ArcKit v6.21.0 is out. Design reviews and vendor scores now get an independent second look before they stand: three fresh readers for every design review, and a second scorer and a neutral moderator for every bid.** *Mark Craddock ยท medium.com/arckit* --- ## Three readers who weren't in the room You can't proofread your own work, because you read what you expect to see. A review written in a long session has the same problem: by the time it's done, it has explained its own gaps to itself. So before `/arckit:hld-review` writes the review, three fresh reviewers now read the design, one each for comprehension, principles and requirements. They get the design, the principles and the requirements. They don't get the conversation, the draft findings, the scores or the verdict. - **The comprehension reader explains the design back in plain words.** Where its account differs from what the supplier meant, the design is unclear, and that's a finding. - **Every finding gets an outcome.** It's added to the review, matched to an issue the review already raised, or turned down with a reason. Nothing is dropped quietly. - **A new blocking finding means the verdict is checked again** before the review is written. In ArcKit's own tests on 8 November, on a fictional housing benefits portal, the three readers raised 34 findings. The review already had 23 of them. Ten were new, and one was turned down with its reason written down. Two of the new ones were blocking: the components had no interfaces between them, and nothing owned citizen accounts and sessions. The verdict was already Rejected and stayed that way. ## The same for detailed designs `/arckit:dld-review` gets three readers too. One explains the design back. One reads it as the engineer who'll build it and lists every decision it leaves to the builder. One checks it against the approved high-level design and its review's conditions, and reports any condition that isn't met, any departure that isn't explained, and any new component or store of personal data the high-level design doesn't have. If there's new architecture, the review asks whether the high-level design needs another look. ## Two scores, and a moderator Public-sector evaluators score a bid on their own, then meet to agree a final score. `/arckit:score` and `/arckit:evaluate` now work the same way. After a supplier is scored, an independent scorer scores the same claims against the same criteria. It never sees the first scores, the conversation or the vendor's name, and it can't open a file, so it can't go and look. Where the two scores differ, a neutral moderator decides. It sees both scores and their reasons as Evaluator A and Evaluator B, with nothing to say which is which, and gives a reason for the final score. "Averaged" doesn't count as one. Its decision is recorded as it stands: ArcKit can tell you it disagrees, but it can't change it. Both first scores are kept, so anyone checking later can see how the final score was reached. On a borderline sample bid, one evaluator gave a 1 and the other a 2. The moderator settled on 2, because the criterion as worded named neither of the unfinished features. ## What it needs The second look needs Claude Code, which can start a reader with no memory of the session. The other assistants can't, so they skip the step and the review says so. A review takes a little longer, because three more readers have to finish before it's written. ## Five short films I've made five 30-second films on how this works, one a day on the [ArcKit LinkedIn page](https://www.linkedin.com/company/arckit-org/) this week. They're silent, so they work with the sound off. ## Thank you The reader lenses are adapted from Google's [Elephant-Goldfish Model](https://egm.dev), which pairs a reviewer who knows everything with one who knows nothing. ## Get it ```bash # Claude Code /plugin marketplace add tractorjuice/arckit-claude claude plugin install arckit ``` Already using ArcKit? Run `claude plugin update arckit@arckit-claude`. --- *ArcKit is MIT licensed and maintained at [github.com/tractorjuice/arc-kit](https://github.com/tractorjuice/arc-kit). The full list of changes is in the [6.21.0 release notes](https://github.com/tractorjuice/arc-kit/releases/tag/v6.21.0).* ## Join the ArcKit Community - **Discord** - real-time conversation, help with commands, and what people are building: [discord.gg/HsA4Y3hQ4](https://discord.gg/HsA4Y3hQ4) - **LinkedIn Group** - announcements, case studies, and longer-form discussion: [linkedin.com/groups/17641034](https://www.linkedin.com/groups/17641034/) - **GitHub** - code, issues, and contributions: [github.com/tractorjuice/arc-kit](https://github.com/tractorjuice/arc-kit)