Why this follow-up exists
The lifecycle corpus marked three web cycles as misses. That could mean a broken web tool, an unstable page, or a model that returned the right fact in the wrong format.
Result
Ten isolated requests asked Bunsen to fetch example.com and return only its H1.
- Factual phrase present: 10/10.
- Exact response: 9/10.
- Median latency: 35.49 s.
- P90 latency: 56.36 s.
- One response added a correct extra sentence before
Example Domain.
The tool path is reliable in this sample; strict response-shape compliance is the weaker layer.
Scanner correction
The first evidence scan treated DISCORD_REQUIRE_MENTION=false as if it were a secret and replaced the JSON literal false. The scanner was corrected to inspect only secret-bearing key names with values at least eight characters long; the usage receipts were restored byte-valid and rescanned. Corrected result: 0 sensitive-value hits.
Falsifier
A larger follow-up should overturn this conclusion if factual success falls below 95% or failures show tool-call errors rather than response-format drift.