Known limitations

Intelligence makes decision policy inspectable; it does not turn incomplete evidence into truth, a ranking into a calibrated probability, or an advisory artifact into permission to execute.

Decision limits

Limitation	Consequence	Responsible interpretation
every ranking is conditional on its candidate universe	omitted or filtered alternatives can change the apparent winner	publish candidates and exclusions with the result
component scores depend on orientation, scale, weights, and missing-data policy	a stable number can encode a changed decision rule	retain normalized policy and component explanations
confidence is only as broad as its calibration corpus	calibration may drift across workflow family, instrument, cohort, or consequence	name the corpus and avoid probability language outside it
contradiction and falsifier machinery exposes pressure; it does not resolve all disputes	a challenged claim can remain genuinely uncertain	preserve adverse evidence and use hold or refusal when needed
sensitivity covers declared perturbations	untested policy or evidence changes may still reverse the decision	report the explored envelope and observed reversals
regret depends on modeled alternatives and costs	unmodeled laboratory, time, or opportunity costs can dominate	pair decision review with Lab consequence evidence
learning uses retained outcomes and policy assumptions	biased or sparse outcomes can reinforce the wrong policy	version adaptations and preserve the prior decision history
Intelligence is advisory	downstream authorization, execution, and scientific acceptance remain separate	route authority to the responsible human, Lab, Runtime, or scientific owner

Inference boundary

flowchart LR
    E["bounded evidence"] --> P["declared policy"]
    P --> D["advisory decision"]
    D --> L["Lab feasibility and authority"]
    L --> R["Runtime execution"]
    R --> O["observed outcome"]
    O --> N["new evidence and policy review"]

Each arrow can narrow or reverse the prior posture. Intelligence cannot promise that the recommended assay is feasible, that execution will succeed, or that the observed outcome will support the original claim.

Report the uncertainty

State the evidence revision, candidate scope, policy, calibration corpus, challenge coverage, sensitivity range, modeled regret, and downstream authority. If one is absent, name the gap and use a weaker posture. “The model recommended” is never a sufficient explanation of what was known or why the action was justified.