Site review: · Dated evidence and limits
Claims checklist
Why
Help readers and reviewers distinguish verified engineering from an attractive narrative. Every workflow attribution needs evidence at the level claimed.
- Link every workflow claim to a demonstrating ticket, commit or study, and name what that evidence actually proves.
- Distinguish implemented capability, recorded usage, measured outcome and owner report. A workflow document is not proof of execution.
- Check cloud-agent authors and PR history before concluding that multi-agent evidence is absent; keep Jules attribution attached to source commits.
- State the snapshot date, counting window and limitations for numbers. Use pace and breadth; never infer quality or agent authorship from volume.
- Treat constrained-output validity, semantic accuracy and retrieval recall as separate results; preserve rejected routes and failed probes.
- Check synced index.html before offering a project website. Use the actual Codeberg repo name for stories; withhold links for unverified remotes.
- Scan authored summaries and linked local story sources for MACs, serials, private hosts/IPs, tokens and file URLs before sync. Public /home/uwe/ paths are allowed.
- Open generated output and run make check, uman website scan and uman website doctor before closing a site ticket. Do not claim remote publication from a local sync.
Evidence boundaries
Recorded cloud-agent PRs in Cati and Loom-games establish Jules contributions. They do not establish local advisor or reverse-sprint execution. A terminal list reported by the owner is not a versioned compatibility benchmark.
Stories are a fixed review snapshot. Upstream source links may become public only after the owner pushes those commits. No numeric quality scores are published.