Uncertainty in CSI

Canonical definition

Claim status: The multidimensional uncertainty taxonomy and propagation requirements are original CSI research-program propositions. They synthesize established construct-validity, multilevel, probabilistic-prediction, and human–AI interaction scholarship but have not yet completed peer review or empirical validation. See the Editorial and Citation Policy.

Why uncertainty is foundational

Collective states are latent, time-bound, and inferred across levels. Membership may change, constructs may overlap, evidence may be incomplete, and multiple composition models may remain plausible. Construct validity depends on the justification of an interpretation rather than a score alone (Messick, 1995), while multilevel emergence research emphasizes temporal and contextual dynamics (Kozlowski et al., 2013).

CSI therefore treats uncertainty as multidimensional. Statistical confidence cannot by itself resolve unclear boundaries, weak construct definition, missing evidence, or unsupported use.

Core requirements

Major forms of uncertainty

Variability and knowledge limitations

Some uncertainty reflects genuine variability in a collective; other uncertainty reflects limited knowledge. The distinction matters because variability may require distributions or trajectories, while knowledge limitations may require additional evidence, narrower claims, comparison, or abstention.

Representing uncertainty

A CSI output may combine calibrated probabilities, intervals, alternative-state rankings, sensitivity ranges, evidence-coverage indicators, model disagreement, boundary-confidence labels, and plain-language limitations. Proper scoring-rule scholarship establishes that probabilistic forecasts should be evaluated for both calibration and sharpness rather than point accuracy alone (Gneiting & Raftery, 2007).

No single representation suits every user. Technical users may need diagnostics; affected participants need clear explanations of missing evidence, plausible alternatives, and prohibited uses.

Uncertainty propagation

Uncertainty introduced early should not disappear downstream. Ambiguous membership affects evidence selection; evidence gaps affect composition; model disagreement affects estimates; and state uncertainty affects explanation, projection, and decision support. Human–AI guidance supports communicating system uncertainty and enabling correction rather than presenting categorical certainty (Amershi et al., 2019).

Illustrative example

A system estimates declining team cohesion, but one-third of interaction occurs in an unavailable channel, two contractors recently joined, and a competing dispersion model suggests subgroup polarization. A responsible result reports the leading estimate, competing interpretation, evidence gap, boundary change, and limited confidence in trend direction.

Abstention and decision thresholds

CSI should abstain or narrow its claim when uncertainty exceeds what the intended use can tolerate. Thresholds should reflect consequence. Abstention may still provide descriptive evidence, identify uncertainty sources, request review, or state which lower-risk conclusions remain supportable.

Communicating uncertainty without distortion

• Avoid false precision: do not imply more certainty than evidence, boundary, or validation warrants.
• Avoid vague disclaimers: identify the specific source and practical effect.
• Separate likelihood from consequence: low probability and high impact are different dimensions.
• Show alternatives: communicate credible competing states or explanations.
• Time-stamp the result: states and uncertainty change as collectives evolve.
• Support contestability: affected people should be able to challenge data, boundaries, assumptions, and use.

Common uncertainty errors

• Reporting model confidence as though it captured construct and boundary uncertainty
• Hiding disagreement among plausible composition models
• Treating missing communication as evidence of disengagement
• Presenting a point estimate when the collective is rapidly changing
• Converting an uncertain state into a confident causal explanation or forecast
• Using generic caveats that do not affect decision authority

Relationship to the CSI construct system

Uncertainty affects every CSI element: collective boundary, state definition, emergence, composition, evidence, validation, awareness, and system behavior. It is a cross-cutting property, not merely a final confidence score.

Scholarly foundations and provenance

Construct-interpretation uncertainty is supported by Messick (1995); temporal and emergent uncertainty by Kozlowski et al. (2013); probabilistic calibration by Gneiting and Raftery (2007); and communication and correction principles by Amershi et al. (2019). The integrated taxonomy and propagation rules are original CSI propositions.

Research status

RA-010 completes the initial architecture and remains a foundational draft. Version 0.2 adds scholarly provenance without changing the canonical definition. Future versions will add formal schemas, calibration and sensitivity protocols, abstention criteria, and empirical evaluation.