Can an evidence checklist reduce unsupported bid claims?
Protocol drafted. Compare reviewed drafts using the same quality rubric. Measure: Unsupported-claim rate per reviewed section.
Research as an operating method
Test the work. Publish the limits.
“Lab” means controlled experiments with explicit hypotheses, risks, permissions, measures, outcomes and decisions. It is not an exemption from law, privacy or commercial responsibility.
Experiment register
Questions before conclusions.
These are proposed protocols, not studies in progress and not evidence of a performance improvement.
Does explaining signal connections improve relevance judgements?
Not recruiting. Compare source-first and summary-first review flows. Measure: Accurate relevance decisions and time to source inspection.
Does calm mode improve completion without reducing comprehension?
Protocol drafted. Run an accessibility-led task-completion study with optional motion. Measure: Completion, comprehension and participant comfort.
Research permission is optional.
Essential service processing and optional research consent stay separate. Private tender documents are not used to train models or publish case studies without specific authorisation.
Claims keep their sources.
Supported drafting links each factual statement to evidence and records human review. Tender-specific AI disclosure requirements are checked for the actual procurement.
Small samples stay small.
Any outcome rate needs a denominator, time window, eligibility definition and treatment of withdrawn or unknown cases. Association is not quietly rewritten as causation.
Every experiment can stop.
A registered method records risk, result, limitations and the decision to adopt, revise or reject the intervention. Participant welfare and service quality take priority over a dramatic result.