VulcanBench cache-read audit
The first test is a narrow technical finding: repricing recorded Sonnet usage under a different cache-read rate. A public reply links it to the original discussion. The question now is whether it earns useful attention.
Read the audit · Read the reply on X
Reply published October 9, 2026 at 21:27:52 UTC.
Who did what
AI handled research, calculation, a separate code check and drafting the explanation. The evidence bundle contains reproducible code and frozen numerical sources.
A human selected the direction, rejected earlier ideas and approved publication. The experiment includes those decisions and oversight.
First observation
At 21:28 UTC on October 9, the reply showed 1 view, 0 likes, 0 reposts and 0 replies. The account had 36 followers. This was taken just after publication, and public counters can lag.
The arithmetic is reproducible. There isn’t enough audience evidence yet to say whether the post was useful or demonstrated demand. Later follower changes also can’t automatically be attributed to it.
| Checkpoint | Due | Result |
|---|---|---|
| 24 hours | October 10, 2026 · 21:28 | Pending |
| 72 hours | October 12, 2026 · 21:28 | Pending |
This page is updated manually. Pending observations haven’t been measured here, and the counters don’t refresh automatically.
What still needs measuring
- Relevant independent replies: no qualitative assessment yet.
- Independent citations or reproductions: not yet tracked.
- Human minutes spent: not yet tracked.
- Incremental AI compute cost: not yet tracked.
All four are unknown, rather than zero. Historical human time and total AI compute cost are also unknown. Future entries can add dated observations, though engagement counters alone won’t establish relevance, independent validation or commercial value.