Evidential Weighting in Governance Chain

Mechanism

Weighting operates as a deterministic function applied at the moment an observation is admitted to the chain. Inputs to the function are five composite factors. First, the authority class of the observing source: a credentialed primary sensor under direct jurisdictional authority contributes differently from a derived inference produced by an automated process operating under a delegated credential. Second, credential continuity — the trust slope — measured as the recency, depth, and revocation status of the credential chain that links the observer to a recognized authority. A credential that has been continuously valid through many checkpoints contributes a higher continuity factor than a freshly issued or recently re-attested credential.

The third factor is corroboration: the count, diversity, and independence of additional observations, already present in the chain, that overlap the new observation in time, space, or subject. Corroboration is computed structurally — a single observation echoed verbatim across many subordinate sensors does not multiply weight, while genuinely independent observations from disjoint authority chains do. The fourth factor is governance policy: the operative jurisdictional policy at the moment of admission, which may declare class-specific weighting floors, ceilings, or multipliers. The fifth factor is operational context: the runtime mode (training, inference, audit replay), the declared mission scope, and the receiver's own governance scope.

The five factors combine into a structured weight record rather than a single scalar. Each factor's contribution is recorded individually, the combining rule is recorded by reference to the policy that governed the combination, and the resulting weight is bound to the observation's lineage entry. Downstream operations consume the structured weight; aggregations, decisions, and inferences integrate observations with their declared weights, and the integration step itself records which weights it consumed. Auditors replay both the weighting and the integration, verifying that the weight applied at decision time was the weight produced by the declared rule under the declared policy.

Weighting is not a confidence score. A confidence score is an internal property of an observation that estimates the observer's own uncertainty. Evidential weight is an external property assigned by the chain, reflecting the observation's structural standing under governance. An observer with high internal confidence whose credential continuity is poor receives low evidential weight; an observer with modest internal confidence whose corroboration and authority class are strong receives high evidential weight. The two properties are recorded separately and consumed separately.

Operating Parameters

The combining rule that converts the five factors into a final weight is itself a governance-declared parameter. In the default rule, factors are normalized to unit ranges and combined multiplicatively, with policy multipliers applied last. The choice of normalization, the choice of combining operator (multiplicative, additive with weights, lexicographic), and the placement of policy multipliers are all parameters that may be set per jurisdictional scope and per observation class. Changes to the rule are themselves observations that enter the chain, so any historical decision can be replayed under the rule that was operative at the time.

Trust-slope decay is parameterized. A credential that has not been re-attested within a policy-declared interval is treated as having a degraded continuity factor; the rate of degradation is a per-class parameter. Corroboration windows — the time and space radius within which observations are considered to overlap — are parameterized per observation class, recognizing that, for example, atmospheric observations may corroborate over hours and kilometers while transactional observations may corroborate only over seconds and identical subjects.

Floors and ceilings are operationally important. A floor specifies the minimum weight an observation may carry once admitted; a ceiling specifies the maximum. Floors prevent admitted observations from being effectively erased by adverse weighting; ceilings prevent any single observation, however well-credentialed, from dominating an aggregation. Both floors and ceilings are class-specific and policy-declared, and both enter lineage with the weight record so that downstream auditors can recover the bounded shape of the original computation.

Alternative Embodiments

In one embodiment the combining rule is fully explicit and stored as a structured expression that the runtime evaluates per observation. In a second embodiment, the rule is expressed as a credentialed reference to a published rulebook; the runtime fetches the rulebook through the same peer mechanism used for skill artifacts and validates its witness chain before applying. In a third embodiment, the rule is expressed as a learned function whose parameters are themselves credentialed and lineage-recorded, permitting machine-learning-derived weighting while preserving auditability of the function used.

Embodiments also vary in how the structured weight is propagated. In one embodiment, only the final scalar weight is propagated to downstream consumers, with the factor decomposition retained only in the lineage record. This minimizes runtime overhead but requires lineage access for re-explanation. In a second embodiment, the full factor decomposition travels with each observation, permitting downstream operations to apply alternative combining rules without re-fetching lineage. In a third embodiment, the decomposition is propagated lazily: scalar at first, with on-demand expansion to the full decomposition when an audit or dispute event requires it.

Treatment of corroboration admits multiple embodiments. A naive embodiment counts corroborating observations within a window. A more refined embodiment computes structural independence by examining the credential chains of the corroborators and discounting overlap. A still more refined embodiment treats corroboration as a graph problem, computing the effective rank of the corroborating set under a declared similarity metric and using that rank as the corroboration factor.

Composition With the Five-Property Chain

Weighting is property two of five. Property one — admissibility — gates whether an observation enters the chain at all. Weighting then assigns the structured contribution. The remaining three properties (lineage retention, dispute, and replay) operate on observations carrying their weights: lineage retains the weight alongside the observation; dispute may target the weight independently of the observation, allowing a party to challenge how an observation was weighted without challenging the observation's admission; replay recomputes downstream operations under the recorded weights, or under counterfactual weights, to produce comparative explanations.

Composition with the marketplace primitive is structural. Bilateral pair-settlement under the chain consumes weighted observations as inputs to settlement computations; the weighting record enters the settlement's audit trail. Composition with spatial adaptation is similarly structural: adaptation parameters derived from training observations carry forward the weighted contributions of those observations, so that a deployed adaptation can be traced to the weighted evidence base from which it was derived.

Distinction From Prior Art

Access-control list approaches treat observation admission as binary: an observer is on the list and contributes equally with all other admitted observers, or is not on the list and does not contribute. Weighted-sensor-fusion approaches in classical signal processing assign scalar weights but typically derive those weights from internal-confidence estimates, not from credential structure or governance policy. Bayesian belief networks combine evidence under priors but operate over modeled probabilities rather than over governance-credentialed structured weights, and do not retain factor decomposition for audit.

The mechanism described here is structurally distinct on three axes. First, weight is derived from the chain's own properties (authority, continuity, corroboration, policy, context), not from observer self-report. Second, the weight is a structured record retaining factor decomposition, not a scalar. Third, the combining rule is itself a governance-declared, lineage-bound object subject to dispute and replay. No prior approach combines these three properties.

Worked Illustration

Consider a multi-source environmental observation fed into a downstream decision aggregator. Three observers contribute reports of an event at approximately the same time and location. Observer A is a credentialed primary sensor whose credential has been continuously valid for eighteen months and whose authority class places it in the highest tier; corroboration from Observers B and C is therefore additive rather than redundant. Observer B is a derived inference produced by an automated process under a delegated credential issued only days before the observation; its credential continuity factor is low and its authority class is mid-tier. Observer C is another primary sensor but its credential, while currently valid, was revoked and re-issued during the observation window; the trust-slope decay parameter applies and its continuity contribution is reduced relative to A.

The combining rule for this observation class, as declared by the operative jurisdictional policy and recorded in the chain, normalizes each factor to the unit interval and combines them multiplicatively, with policy multipliers applied last. Floors prevent any of the three contributions from being driven below a class-declared minimum; ceilings prevent A from dominating the aggregation despite its strong factors. The aggregator consumes all three weighted contributions and produces a result whose lineage records every factor decomposition, every policy multiplier, and the rule version under which the combination was performed. A subsequent dispute targeting B's weighting — perhaps on the ground that the delegated credential was issued outside policy — can be evaluated against the recorded decomposition without re-running the entire aggregation.

Disclosure Scope

This disclosure covers the five-factor weighting function, the structured weight record, the lineage binding, the parameter set governing factor normalization and combination, the floor/ceiling discipline, the worked-illustration discipline for replay and dispute, and the composition of weighting with the remaining four governance-chain properties. Alternative embodiments — explicit rule, credentialed-reference rule, learned-function rule; eager, lazy, and on-demand decomposition propagation; naive, structural, and graph-rank corroboration — are within scope. Implementations that admit observations binarily, or that derive weight solely from observer self-confidence without governance binding, fall outside the disclosed mechanism.