Two correct addresses. No request reached the server. As a hyperlink, the same destination was opened twice.
Published manuscript · v1.0
THE PAGE WAS GIVEN THIS TIME
In the discovery experiment, the agents never saw the site. This time ChatGPT was given the page. The question was smaller: once a page tells an agent where to go next, does it get there? Each trial used a fresh, unguessable path prefix on one host. The index page said the access code was not there and to open code/heron/, relative to the page. The server computed the code for each prefix at request time, so the exact code required a request for that trial's code page.
TWICE, THE ADDRESS WAS RIGHT
In both Phase 0b trials, the server logged ChatGPT-User/1.0 retrieving the index page. Both answers quoted the instruction and wrote out the correct absolute URL for the code page. Both then reported it inaccessible: “returned an access error when I tried to open it”; “inaccessible through my browsing tool.” The server recorded no request for either code page under any user-agent. Independent checks 41 and about 20 seconds later each returned the correct code.
NOTHING WAS INVENTED
Neither failed answer supplied a code. Both asked the user to open the page and paste its contents. The frozen gate recorded FAIL, 0/2, and the planned main study did not run.
A MISSING REQUEST IS AMBIGUOUS
The answers say an attempt was refused; the server says nothing arrived. The model may have declined to issue the request, the browsing tool may have refused a URL the model constructed, or something between tool and host may have blocked it. These records cannot distinguish those accounts. The reports of an attempted fetch are the model's self-description, not independent evidence of a request.
ONE CHANGE: THE PATH BECAME A LINK
Every page ChatGPT had opened beyond the first in an earlier study on a different host was a hyperlink. Here the path was plain text. After the failure, a versioned amendment was frozen, keeping the original verdict. It changed one thing: the identical instruction, with code/heron/ as a hyperlink, verified by diff. Two fresh prefixes; the same prompt, settings and scoring rule.
CLAIM
In the tested setup, presenting the same destination as a hyperlink rather than a plain-text relative path changed whether it was retrieved.
Plain text: 0/2 code-page requests, 0/2 codes. Hyperlink: 2/2 requests by ChatGPT-User/1.0, each 1.2 seconds after the index page, and 2/2 exact prefix-specific codes.
PARTIALLY SUPPORTED
Link availability, rather than whether a correct address is known, governs which pages consumer ChatGPT retrieves.
The four trials are consistent with a browsing tool that opens addresses supplied by the user, found in search or linked on a fetched page, and declines addresses the model constructs. Only a relative path was tested as plain text, so presentation as a link cannot be separated from presentation as a complete address. The component that prevented the plain-text requests is not identified.
A correct address is not a request.
Does an agent need a destination to appear as a hyperlink, or is a complete absolute URL sufficient?
Hold the page, wording and destination fixed. Compare a hyperlink with the complete absolute URL written as plain text, with server logs as the measure of retrieval.
KEEP THE PRESENTATION IN THE CLAIM
- Consumer ChatGPT, Thinking effort Medium, memory off, a fresh chat per trial. Two trials per condition, one host, one day (11 October 2026).
- Not tested: a complete absolute URL presented as plain text; addresses arriving through search results; other assistants, modes or hosts.
- The amendment was designed after the failure it explains. Its interpretation was frozen before its trials ran; the comparison is still post hoc.
- The original Phase 0b verdict remains FAIL, 0/2. The hyperlink result does not pass it retroactively.
The observation is a difference between two presentations in one setup, not a general rule about agent browsing.
Lab notebook / Published record
Conclusion
A correct address is not a request. In these four trials, the plain-text path produced no code-page requests; the same destination presented as a hyperlink was requested twice and both exact codes were returned.