Changelog

For users and operators. Reference.

There is no tagged release yet, so changes are grouped by the date they reached the main branch. Technical detail is in the repository’s own CHANGELOG.md and its commit history. The tool never writes to OpenStreetMap in any of these versions.

20 October 2026

• The demo is now fifteen real changesets, recorded on 20 October 2026. Each has a reading from the Strands Decider 2B. The lanes came out as 4 needs local knowledge, 4 review suggested and 7 insufficient evidence. None is in looks consistent.
• Queue rows show the mapper’s own changeset comment.
• A fresh workspace can be claimed only by an OpenStreetMap account named in the server settings. See first admin.
• The queue, case, draft, history, config and settings screens now share one look, with a banner for the person claiming a workspace.
• The public home page gains a section on how decisions are made, with readings from a real run, and scroll motion with a switch for less motion.
• The map now loads in the development server, and the dropdown focus ring is fixed.
• This documentation site is added.

19 October 2026

• Sign in with OpenStreetMap, with read-only access. Invite codes still work. See sign in.
• Shareable decision receipts, which anyone can open from a link. See read a receipt.
• Community packs, and calibration from reviewers’ recorded decisions. See tune it to your community and calibration.
• A settings page for admins: workspace name, tagline, accent colour, logo, invites and the sign-in switch. See brand the workspace.
• Public home, About and Privacy pages.
• A redesigned evidence desk: a side rail, a command menu, a queue grid and a timeline.
• The queue lists cases where the mature-deletion rule fired ahead of every lane. They keep their lane.
• A command for comparing model sizes, cso eval ablate-model. The comparison itself has not been run.
• For operators: continuous database backup, secrets read at boot, alarms and a safer deploy for the single AWS host. See deploy on AWS.

18 October 2026

• Safety and honesty fixes from a review: parsed facts now set a floor under the model’s readings, drafts are metered and checked more strictly, and imports have a quota.
• The evaluation write-up now says plainly that a constant guess matches the model on the same readings. See what it cannot do.

17 October 2026

• First working version.
  • Collects the public evidence for a changeset, with a call budget and a coverage record.
  • Asks the Strands Decider 2B yes or no questions about it, applies fixed rules, and puts the case in a lane.
  • Drafts a comment with Claude Haiku 4.5 on Amazon Bedrock, with a plain template as a fallback.
  • A keyboard-first queue, case view, draft editor, outcome recording and history.
  • Invite-code sign-in, an HTTP API and a command line.
  • Evaluation tools and the first measured results.