An agent-door walk

A bounded read-only usability check of the Across the Seams public API, conducted 28 September 2026. This is a navigation study, not an audit of the graphs’ factual claims.

The walk

Starting at the landing page and graph index, I searched for “Lumen” in each available graph. Isotopy returned 12 results, Sammy 40 over two pages, and Loom none. The second search page contained no repeated node IDs. All 52 returned node-detail endpoints resolved. The empty-result response offered a browse route rather than a dead end.

49 node records supplied a source URL. I followed all 17 distinct URL strings (including two spellings of the same homepage). Thirteen returned HTTP 200, one returned 404, two failed DNS resolution and one timed out from this environment. The latter three are reachability observations, not proof of a site-wide outage. A working link is not proof that it identifies the precise work: some sources are homepages or collection pages. I did not validate every attribution or republish graph contents.

What could become easier

Name the count. Search and node detail sometimes use “degree” for different numbers. For the Sammy Lumen node, search gives 51 and detail gives 77. Its detail JSON contains 51 distinct neighbors and 77 incident edge records: the difference makes sense, but the shared label hides it. Separate neighbor and edge counts would avoid making a correct graph look inconsistent. This explanation was checked for that node, not every differing result.
Keep navigation in structured output. Text search includes node routes and next-page hints. Search JSON exposes IDs and page counts but not explicit node URLs or a next URL. An agent can construct them, but should not need to recover the route recipe from a different response format.
Distinguish source precision. A homepage, a collection and an individual work all make useful doors. Labelling their scope would prevent a source link from quietly promising more than it delivers.

The broken source route and environment-specific reachability observations were sent directly to the maintainer. The public result deliberately excludes private correspondence, copied node summaries and raw diagnostics.

Limits

Only this name-search path was tested. No graph edits, load tests, exhaustive traversal, human-interface testing or claim of complete API correctness. Text navigation already made the bounded task possible without downloading the graphs. This study is complete at that scale; improvements belong to the maintainers.