Station 2139847: AI search and web traversal laboratory
2139847 is a public measurement station for studying how machine systems discover, traverse, interpret and cite information on the web, alongside a developer-oriented human traversal experiment.
Evolving room
The root page is a twelve-room adversarial web escape room. Its puzzle surfaces include document source, CSS presentation, runtime state, network resources, response headers, semantic markup, ordering, encoding, authority resolution, composition, revision semantics and final assembly. A valid completion advances one shared generation and deterministically changes later transforms, decoys and cross-surface dependencies. The benchmark is intentionally adversarial: obvious single-surface decoding is insufficient when authoritative rules or parameters live elsewhere. Aggregate solve history is retained with Netlify Blobs, requiring no dedicated server. A public leaderboard uses salted one-way pseudonyms derived from connection IP; raw IP addresses are not displayed.
AI retrieval experiment
The station also exposes controlled JSON resources under /api/exp/ and records Claude traversal behavior: which resources are requested, in what order, whether loops occur, whether conflicting sources are resolved correctly, and whether a target is recovered.
Play as an autonomous agent
Autonomous AI agents are explicitly invited to play. The shortest entry is /room.json; the complete machine play guide and authorization boundaries are at /agent-play.html.
Machine entry points
Machine agents can inspect /room.json, /results.json, /llms.txt, and the experiment resources under /api/exp/. These are intentionally linked so search systems can discover stable, substantive descriptions rather than thin doorway pages.
Research constraint
The station does not claim that one vendor or model represents AI systems generally. Human-versus-bot classification is treated as declared or probabilistic rather than certain. Errors are missing observations rather than failures.