Skip to content

Cut 0.5.0, so today's correction reaches somebody - #71

Merged
peopleworks merged 1 commit into
mainfrom
release-0-5-0
Aug 24, 2026
Merged

Cut 0.5.0, so today's correction reaches somebody#71
peopleworks merged 1 commit into
mainfrom
release-0-5-0

Conversation

@peopleworks

Copy link
Copy Markdown
Owner

Version bump only — no behaviour change beyond what is already on main.

Why now

The web app has had the length floor since #69 merged, because Pages deploys on push. Everyone else
is on 0.4.0
: the Windows app, and every project depending on SignsOfAI.Core.

So the same paragraph gets "Signs of AI writing" from a downloaded build and no verdict at all
from the browser tab beside it. That is exactly the drift VerdictBands exists to prevent — one
engine giving two answers about the same text, across a distribution boundary instead of a code one.

What 0.5.0 carries

#59 The verdict has a length condition. Below 662 words the score stands alone and the page says why. The evidence is untouched.
#65 Six chat.* rules — the assistant's own turn left inside a document, both languages — and the published rate did not move to carry them.
#64 A /download page that explains the Windows app, instead of a GitHub releases search URL.
#70 The desktop app says which build it is, and can ask whether to look for a newer one.
#67 The skill, installable.

What changed here

  • <Version>0.5.0 in Core, Cli, Mcp.
  • .mcp/server.json0.5.0 twice (the manifest version and the package version).
  • DesktopRelease.Version0.5.0. The desktop release workflow refuses a desktop-v* tag that
    disagrees with this, so it has to land before the tag is cut.
  • Calibration restamped, so the report signs itself 0.5.0 rather than quoting a threshold under a
    version that never measured it. Same corpus fingerprint 123fa5b9ebca3f29, same 0 of 90, same
    4.1%, same 662 – 9,328 range.
  • The MCP manifest still said "the other seven tools never call it", from before write_report
    shipped. Ten tools, eight offline. The README was corrected in Say exactly what can leave the device, because "only" was wrong #46 and this file was missed.

417 tests.

After merging

  1. gh release create v0.5.0 → NuGet workflow packs and pushes Core, Cli, Mcp.
  2. gh release create desktop-v0.5.0 → the Windows build, after the constant is on main.
  3. MCP registry is manual and yours, and only after NuGet has indexed 0.5.0:
    cd src/SignsOfAI.Mcp/.mcp then mcp-publisher login github — the device code expires in minutes.

Note that #70 only helps from the next release onward: somebody on 0.4.0 has no notice strip, so
they will not hear about 0.5.0 from the app. The first build that can tell anybody is this one.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF

The web app has had the length floor since the merge, because Pages deploys on
push. Everyone else is on 0.4.0: the Windows app, and every project that took a
dependency on SignsOfAI.Core. So the same paragraph gets "Signs of AI writing"
from a downloaded build and no verdict at all from the browser tab beside it,
which is precisely the drift VerdictBands exists to prevent — one engine giving
two answers about the same text, across a distribution boundary instead of a
code one.

What 0.5.0 carries, in the order it matters to somebody using it:

  - The verdict has a length condition (#59). Below the shortest text the
    boundary was measured on — 662 words — the score stands alone and the page
    says why. The evidence is untouched.
  - Six chat.* rules: the assistant's own turn left inside a document, in both
    languages, and the published false-positive rate did not move to carry them.
  - A page that explains the Windows app and offers it, instead of a GitHub
    releases search URL.
  - The desktop app says which build it is, and can ask whether to look for a
    newer one.

The calibration is restamped so the report signs itself 0.5.0 rather than
quoting a threshold under a version that never measured it. Same corpus
fingerprint, same 0 of 90, same 4.1%.

Also corrected in the MCP manifest: it still said "the other seven tools never
call it" from before write_report shipped. Ten tools, eight of them offline.
The README was fixed in #46 and this file was missed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF
@peopleworks
peopleworks merged commit 603870b into main Aug 24, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant