SCIENCE UNDER AUTOMATION

How AI capabilities are compared with scientific work

Start with the work a field requires. Match each essential activity to what an AI system actually performed.

Define the work

Record representative workflows, their inputs, physical requirements, reasoning, originality, contextual judgement and validation. State which parts of the field the assessment covers.

Match the evidence

Record the system, version, date, task, denominator, comparator, tools and human contributions. Separate a directly tested capability from an extrapolation between fields.

Keep measurements distinct

Prediction accuracy, laboratory yield, expert preference and task completion measure different outcomes. They cannot be averaged into a field takeover percentage.

Score publication rule

The earlier prototype assigned untested requirements a 0–4 range. That could inflate an upper score. Those bounds are not approved for publication. A numerical assessment requires a justified mapping, explicit workflow weights, sensitivity checks and review. Unknown capability receives no speculative credit.

Corrections and versions

Evidence records retain their source dates. Reviewed releases record the change, reviewer and content fingerprint. Changing the content invalidates its previous approval.