This contract applies to every candidate, experiment, report, and automated component.
- Every consequential factual claim must identify its source and precise source location.
- AI-generated summaries are unverified until checked against the source.
- Secondary sources may orient retrieval but do not replace primary evidence for scientific claims.
- Missing, inaccessible, retracted, corrected, or disputed evidence must remain visible.
- The corpus must have documented inclusion, exclusion, coverage, and stopping rules.
- Records must distinguish observation, statistical relationship, model inference, causal hypothesis, and speculation.
- Every promoted question must include competing explanations and evidence against novelty.
- Terminology similarity is a lead, not proof that phenomena or mechanisms are equivalent.
- Importance, novelty, tractability, and available-data quality are separate judgments.
- Uncertainty and reviewer disagreement must not be collapsed into a single confidence score.
- The question, hypotheses, data rules, primary outcomes, holdouts, tests, and falsification criteria must be preregistered before final evaluation.
- Training, exploration, validation, and final holdout data must be separated where the design permits.
- Baseline reproduction precedes novel analysis.
- Leakage, confounding, robustness, measurement error, and alternative explanations must be tested.
- Deviations from preregistration must be timestamped and labeled exploratory.
- Negative results and failed experiments remain in the record.
- Reproducibility includes code, configuration, environment, data provenance, and permitted acquisition steps.
- Correlation cannot be described as causation.
- “Novel” means no prior work was found after a documented adversarial search; it is never an absolute guarantee.
- Computational evidence cannot be presented as physical or clinical validation.
- Expert scrutiny is required when interpretation depends materially on domain knowledge unavailable within the evidence record.
- Publication, public claims, human-subject work, restricted data, laboratory work, or consequential external action requires a separate explicit decision.
- Automation may retrieve, structure, compare, generate candidates, and run declared computations.
- Automation may not silently expand scope, suppress contrary evidence, select a final question, or promote a scientific claim.
- The system must preserve an auditable path from source to claim to candidate to experiment to conclusion.