- InputEligible candidates or measurements, their sources and a stated objective.
- HandlingSelect one candidate OR aggregate specified inputs; retain the rule and version.
- OutputA sourced selection or candidate aggregate with uncertainty and missing inputs.
Delivery, permissions and correction
- Permissions
- Use only permitted sources and disclose selection criteria. Personalisation and aggregation need separate controls.
- Failure
- Missing, duplicate, correlated or adversarial inputs can distort the result. Disagreement is not removed by convergence.
- Correction & evidence
- Recompute under a recorded rule when inputs are corrected or withdrawn; explain what changed rather than silently replacing the result.
Selection answers “which one?”
A feed or assistant may select a useful item, model or service for one person from the eligible options among many candidates. The ranking objective should be stated, personalised controls should remain visible, and a selected item should retain its source.
Aggregation answers “what do they jointly show?”
A reducer may combine measurements, reports or predictions into one candidate result. It must name the inputs, missing participants, algorithm version, uncertainty and correction path. A median, average or vote is only as meaningful as its source and rule.
Many inputs do not make truth
Selection is not aggregation. Aggregation is not ledger consensus. Neither one establishes truth, authorisation or finality by itself. Those properties need separate evidence and governance.
Still to decide
Research must define provenance, duplicate and Sybil handling, uncertainty, disagreement, exposure concentration, user correction and how an outcome changes when an input is later withdrawn.
Research target: no global ranking algorithm or many-to-one settlement rule is claimed.