RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern — 4. Measurement Science Foundation: A Concern Vector Is a Measurement Claim

AIGP Specification › RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern › 4. Measurement Science Foundation: A Concern Vector Is a Measurement Claim

← 3. Ontological Foundation: What Exists to Be Observed · Section index · 5. The Mediation Vector Profile: A Four-Layer Model →

4. Measurement Science Foundation: A Concern Vector Is a Measurement Claim

4.1 The Core Realization

A Mediation Vector is not a list of scoring fields. It is an empirical measurement claim. When a vector reports authority_compliance = 0.4, it is claiming to measure something — and measurement claims carry obligations that arbitrary numbers do not.

This reframes the entire RFC:

A Mediation Vector Profile is not merely a list of scoring variables. It is a validity-aware measurement profile that declares how a concern construct is represented by observable, typed, evidence-backed variables within a declared Domain of Concern.

The science that governs this is not “AI governance” literature. It is measurement science — construct validity, validity theory, measurement scales, and provenance — fields that have spent a century learning how to measure things that cannot be observed directly.

4.2 The Concern Is a Construct

A concern like “unsafe physiological reassurance,” “unauthorized autonomous action,” or “misleading material disclosure” is not directly observable. It is a construct — an abstract attribute inferred from indicators.

Construct validation (Cronbach & Meehl, 1955) established that whenever a measurement is interpreted as capturing an attribute that is not directly operationally defined, the measurement must be validated against a theory of what the attribute is and how it manifests. (Construct validity in psychological tests) — Content rephrased for compliance.

Therefore every Mediation Vector Profile MUST declare its concern as a construct:

{
  "concern_id": "unsafe_physiological_reassurance",
  "construct_definition": "The artifact reassures a user about a physiological symptom in a manner that suppresses warranted escalation under medical uncertainty.",
  "observable_indicators": [
    "red_flag_symptom_present",
    "escalation_recommended",
    "uncertainty_disclosed",
    "reassurance_strength"
  ],
  "measurement_assumptions": [
    "Red-flag symptoms are extractable from the input artifact",
    "Escalation language is detectable in the output artifact"
  ],
  "known_invalid_uses": [
    "Do not apply to non-medical reassurance",
    "Do not apply where the user has already confirmed emergency care"
  ]
}

The vector does not merely calculate. It claims to measure something, and that claim must be defensible.

4.3 Validity Is an Argument, Not a Number

Messick’s unified validity theory treats validity not as a property a test either has or lacks, but as an integrated evaluative judgment spanning content relevance, criterion relationships, construct meaning, and the consequences of use. (Validity, in Educational Measurement) — Content rephrased for compliance.

A Mediation Vector Profile therefore does not assert:

“This vector is valid because it has variables.”

It asserts:

“This vector is valid only when its variables, evidence, calculation, and consequences are appropriate for the declared concern, in the declared domain, under the declared conditions.”

This directly supports the non-causal principle established across this series:

Governance makes quality claims admissible for empirical evaluation; it does not make them valid. Validity is argued in the profile, not assumed by the protocol.

4.4 The AI-Evaluation Field Is Rediscovering This

Recent work on LLM and AI benchmarks argues that evaluating abstract phenomena — safety, robustness, capability — requires construct validity, and that poor task or metric selection undermines the conclusions drawn from benchmark scores. (Evaluating the construct validity of AI benchmarks) — Content rephrased for compliance.

RFC-035 is not another benchmark. It is the mechanism that forces the prior questions every benchmark should answer:

What exactly are we measuring?
Why do these variables represent the concern?
What evidence supports each variable?
Where is this vector invalid?

4.5 Variables Must Be Measurement-Typed

Stevens’ theory of measurement scales (nominal, ordinal, interval, ratio) carries a practical consequence: not every variable admits the same mathematical operations. Ordinal labels cannot be summed; nominal categories cannot be averaged. (On the theory of scales of measurement) — Content rephrased for compliance.

Every Mediation Vector variable MUST declare its measurement type and the operations it permits:

{
  "type": "ordinal",
  "scale": "ordinal",
  "allowed_operations": ["rank", "median", "threshold"],
  "normalization_method": "ordinal_to_unit_interval_v1",
  "aggregation_allowed": false
}

This prevents the most common empirical error: treating ordinal judgments as if they were ratio quantities. low + medium + high is not arithmetic unless a declared, validated transformation makes it so.

4.6 Variables Exist to Answer Concern Questions (GQM)

The Goal–Question–Metric paradigm holds that metrics should not be chosen first; they should be derived — from a goal, refined into questions, answered by metrics. (The Goal Question Metric approach) — Content rephrased for compliance.

Every variable in a Mediation Vector MUST trace to a concern question:

Concern:   Unauthorized autonomous action
Question:  Was the system authorized to execute this action?
Variable:  authority_compliance
Evidence:  signed authorization record, scope envelope, action trace

A variable that answers no concern question does not belong in the vector. This is the discipline that prevents vectors from accumulating fields that feel relevant but measure nothing the concern requires.

4.7 Every Value Carries Provenance

W3C PROV models provenance as entities, activities, and agents — what was produced, by what process, used by and influenced by whom. (PROV-DM: The PROV Data Model) — Content rephrased for compliance.

This maps directly onto AIGP, ALP, D-DNA, and temporal chaining. Every variable value MUST carry provenance, not as after-the-fact metadata, but as a precondition for empirical status:

{
  "variable": "human_approval_present",
  "value": true,
  "evidence_ref": "approval-record-123",
  "derived_from": ["scope-envelope-77", "operator-action-9"],
  "generated_by": "governance-server",
  "observed_at": "2026-06-28T22:00:00Z",
  "confidence": 0.98
}

Without provenance, a calculation may be numeric, but it is not empirical.

4.8 Variables Carry a Warrant (Toulmin)

Toulmin’s model of argument separates the claim from its grounds (data), warrant (the inferential license), backing (support for the warrant), qualifier (strength), and rebuttal (defeating conditions). (The Uses of Argument, Toulmin) — Content rephrased for compliance.

A concern verdict is an argument, and its structure should be explicit:

Claim:     This artifact raises an unsafe-reassurance concern.
Grounds:   Red flag present; escalation absent; uncertainty undisclosed.
Warrant:   In this Domain of Concern, red flags require escalation.
Backing:   Domain profile, clinical safety policy, prior validated cases.
Qualifier: High confidence, but user history incomplete.
Rebuttal:  If the user already confirmed urgent care, the concern reduces.

Each variable or variable group SHOULD declare a warrant and backing, so the verdict is an auditable argument rather than an opaque number.

← 3. Ontological Foundation: What Exists to Be Observed · Section index · 5. The Mediation Vector Profile: A Four-Layer Model →