[watson-qg]

KI7MT asked for every repo's docs to match reality before a QRZ post. Bob audited the non-MCP side and left the MCP content to me; this is that, plus what auditing it turned up.

The command, 78 references across 7 files

qso-graph-auth installed as qso-graph-auth and ran as qso-auth until 0.5.0 (qso-graph-auth#28). Every example on the site typed the old name — including docs/install.sh, whose comment named it as the command you get.

qso-auth still works and prints where it went, so nothing here was broken. It told readers to type something that now prints a deprecation notice. The rename is noted once, on the qso-graph-auth page, where someone with it in a script would look for it.

Counts that were wrong

Each replaced with a measured one, not a corrected guess:

Said	Is
nine MCP servers	22 — public, non-archived *-mcp repos
ADIF 3.1.6	3.1.7 — live get_version_info; 3.1.6 still ships
26 enums, 4,427+ records	ADIF's 25 / 3,345, plus Country re-keyed by adif-mcp → 26 / 3,745 as shipped
adif-mcp test matrix: 48	the register's 109, and pytest's 186
Testing: 108/108 PASS	2,464 tests across 24 packages

The enum count is the interesting one. 26 is right for what adif-mcp ships and wrong for what ADIF publishes, because Country is derived here, not in the spec — and both numbers sat on the page as if they measured the same thing. I nearly "fixed" 26 to 25 and would have introduced an error. Verified two ways that agree: summed from the shipped enumerations.json, and against the running server's own per-enum counts.

The test numbers drifted because the page copied a file in another repo. adif-mcp/tests/TEST-CASE-REGISTER.md is the record; the copy here said 48, the register says 109, and pytest collects 186. Three answers to one question, so the table now points at the register and states only the two totals it doesn't own.

The retired review gate, on three pages

• security.md described an audit request sent to an internal message queue and an independent agent returning a PASS.
• testing.md described a writer and two audit/reviewers — "All three must pass before the tag is created."
• llms.txt said every package requires a security audit before publication.

None of that is what stands between a change and PyPI. publish.yml's mechanical checks are, and that is now what the pages describe. Describing reviewers who are not there overstates what a reader is getting, which is the specific thing this pass was asked to fix.

Also

• The architecture diagram showed 8 of 22 servers, and drew every one making an outbound HTTPS connection. Nine make none at all — the zone servers and adif-mcp read data shipped in the wheel, and n1mm-mcp listens to a UDP broadcast on the local network. It now shows all 22 grouped by what they do, the three with no working tools labelled as such, and only the ones that call out drawn calling out.
• Claude Code reads MCP servers from ~/.claude.json, not ~/.claude/settings.json (two pages).
• The contributing page's issues table listed 8 of 22 MCP repos — a bug in the other 14 had nowhere to go. All 25 packages are listed now.
• Its dev setup ran python3 -m venv and then uv sync, which manages its own. The venv step is gone and the commands run through uv run.
• The adif-mcp section was headed (v1.0.0); it is 2.1.0. A pinned version in a heading is wrong by default, so it's gone rather than updated.
• The test corpus section cited only 3.1.6. Both are run — 6,197 records for 3.1.7 and 6,191 for 3.1.6, counted from the test data rather than quoted.

Checks

mkdocs build --strict passes against the repo's pinned requirements.txt. The Fleet Overview's server and tool counts are build-time macros reading live PyPI, so those were already self-correcting and are untouched.

Companion: .github — the same command named in three places in the org profile.

🤖 Generated with Claude Code