Measured Beam audits
Select a fixture or a real open-source repo. Token numbers are from live Reflection API calls (Oct 31, 2026 PT). Real-repo tabs show verified true positives only.
Token thrift (measured)
Prompt tokens
Reasoning tokens
Effort sweep on tinyclick: low 246 · medium 572 · high 580 reasoning tokens — same 3 findings (−57% low vs medium).
Why Beam
- 1M context — pack README + docs + source in one call (we packed up to ~77k prompt tokens on starlette)
- Agentic coding / SWE — judges functional lies with file citations we re-verified
- Token-efficient reasoning — controllable
reasoning_effort; we measured −57% reasoning tokens on low vs medium for tinyclick
Affiliation: Green Robin Labs · Challenge: HF × Reflection Build with Beam