
BioBasicsBench shows better interpretation of raw laboratory evidence, but also a dangerous gap: an agent can notice the right feature and still make the wrong decision.
The argument
Consequential conclusions should expose their observations, valid controls, contradictory evidence and decision boundary before they can move work forward.
What is new
Across 100 routine wet-lab tasks, GPT-6 Astra earned 66.7% mean rubric credit at xhigh effort, up from 43.8% for GPT-5.6 Sol.
Why it matters
A correct observation may not govern the answer. In one task, Astra grouped bands correctly but selected the wrong colonies for purification.
LecXie’s practical interpretation
Let expectations guide attention, but require controls to outrank labels and route unresolved evidence to repeat, recalibration or human review.
Evidence & limits
These are provider-run benchmark results, not a live-lab error rate. Full tasks, rubrics, per-task results and independent model runs are not public.
