Hark checks four public pages against source text

Wakaba tested Hark’s web app on October 8–9 using a deliberately narrow research task. The first run specified three public Hark pages covering usage, pricing and privacy. Instructions requested a short Japanese table with source links, excluded email and conversation history, and prohibited actions beyond reading.

The interface displayed a page-reading tool status before returning answers. Wakaba followed the links and compared the reported items with the originals. This check also exposed an outdated statement in Wakaba’s own research notes about model training; the author corrected the note instead of treating it as current policy.

The second run used one OpenAI help page about banked Codex resets. Hark distinguished the stored reset from purchased credits and explained the usage windows and where expiry information could be checked. It did not fill in a specific expiry date absent from that page. Similarly, the first response left unsupported yen prices and concrete free-tier quantities unstated.

The result is a checked, source-linked comparison from four supplied URLs. It does not establish independent source discovery, handling of conflicting evidence or the ability to operate websites through Handoff. Wakaba inspected neither a Handoff interface nor its internal logs, connected no private accounts and tested no purchase, booking or sending action.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Blog · Wakaba (ワカバ)