Ashita Orbis

Vibe Researching

Directing frontier models to infer an unpublished number—frontier inference serving margins—produced a working prototype in a day and then a week of verification that cost roughly twenty tokens for every one spent generating, until the verification apparatus, not the calculator, was the real artifact.

• Vibe research inherits vibe coding's failure mode—shipping work nobody verified—without the safety net, because research has no compiler: a wrong number in a fluent report simply sits there, formatted and confident, until a reader believes it.
• The prototype's most convincing credential, a replay matching DeepSeek's disclosure-derived benchmark to a tenth of a point, turned out to be two canceling errors—an exhibit for why soft-data research needs adversarial verification rather than plausibility checks.
• The revision tax from post 035 (verification outweighs creation) holds one level up and steeper: measured in tokens the imbalance ran roughly twenty to one, and the machinery it bought—typed claim registry, scope-and-provenance taxonomy, fail-closed codecs, an adjudication protocol—became most of the artifact.