RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern — 2. Epistemological Foundation: What We Learn from Science

AIGP Specification › RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern › 2. Epistemological Foundation: What We Learn from Science

← 1. Abstract · Section index · 3. Ontological Foundation: What Exists to Be Observed →

2. Epistemological Foundation: What We Learn from Science

2.1 We Are Not Inventing a New Epistemology

The scientific method is a 400-year-old protocol for producing reliable knowledge under uncertainty. It took humanity from alchemy to chemistry, from astrology to astronomy, from anecdote to evidence-based medicine. It is the most successful knowledge-production system ever devised.

This RFC does not propose a novel framework for evaluating mediated intelligence. It applies the epistemology that already works — the one that science established — to a new substrate. The substrate is new (AI systems that act autonomously on behalf of humans). The observation method is not.

The position is deliberate: we draw from science to build a foundation of observation, and that foundation becomes the basis for internal and external mediated observation. Nothing bespoke. Nothing invented. The scientific method, made executable as protocol.

2.2 The Scientific Method as Protocol

Scientific Principle	What Science Learned	RFC-035 Implementation
Hypothesis before observation	You cannot validate what you did not predict. Post-hoc rationalization is not science.	ANTICIPATE is sealed before execution begins. The expected state is declared immutably before the outcome is known.
Independence of observer	The experimenter cannot also be the measurement instrument. Self-reporting is not evidence.	Observer lineage MUST be independent from the evaluated generation lineage. The producing entity, model, agent, or toolchain MUST NOT be the lineage that measures or grades the artifact. No self-evaluation.
Reproducibility	A result that cannot be independently replicated is not a finding — it is an anecdote.	D-DNA recomputation witnesses allow any qualified third party to re-derive the verdict from retained evidence alone.
Declared methodology	The method must be stated before results are reported so others can evaluate the observation process, not just the outcome.	Extraction methods are declared per variable. The process by which evidence becomes a value is transparent and auditable.
Falsifiability	A hypothesis that cannot be disproven is not scientific — it is unfalsifiable assertion.	VERIFY can produce MISMATCH and VIOLATION. The anticipation can fail. The system is designed to say “no.”
Controlled variables	Confounding factors must be isolated. Not everything that changes is causally related.	The four-state model separates anticipated, actual, evidenced, and perceived — preventing conflation of expectation with outcome, or belief with measurement.
Peer review	No single observer is authoritative. Knowledge requires independent qualified judgment.	Domains of Concern declare observer competence requirements. Verdicts require qualified observers — human, machine, or hybrid.
Null result is a result	The absence of a finding is informative. Negative results must be published.	INDETERMINATE is a verdict class. The system surfacing what it cannot evaluate is as important as what it can.
Progressive refinement	Theories evolve from accumulated evidence. Yesterday’s edge case becomes tomorrow’s understood phenomenon.	Second-order criterion evolution: the system generates new observation criteria from its own empirical history.
Observer bias acknowledgment	All observers have biases. Science doesn’t eliminate bias — it declares it, measures it, and accounts for it.	Perceived state is captured explicitly and separated from evidenced state. The gap between them (perception gap) is itself a measurable signal.

2.3 Why This Matters for AI Governance

The strongest counter-argument to governed AI evaluation is: “the model is good enough — governance is overhead.” Science’s answer to that claim has been consistent for four centuries:

You do not get to claim something works until you have observed it working, independently, reproducibly, under declared conditions, with the possibility of falsification.

That is not a governance argument. It is an epistemic argument. It does not depend on regulation, compliance frameworks, or organizational policy. It depends on the claim that reliable knowledge requires reliable observation — and reliable observation has structural requirements that AI systems do not satisfy by default.

Those structural requirements are what RFC-035 implements.

2.4 The Internal/External Observer Problem

Science encountered the observer problem centuries ago and solved it structurally:

Problem	Scientific Solution	RFC-035 Solution
The experimenter sees what they expect to see	Double-blind studies	Four-state separation: anticipated ≠ perceived ≠ evidenced
Single observations are unreliable	Replication by independent labs	Recomputation witnesses; qualified external observers
Self-reported data is unreliable	Independent measurement instruments	Evidence admissibility levels; D-DNA signing
Expert consensus can be wrong	Falsification; anomaly tracking	INDETERMINATE verdicts; second-order criterion evolution
The observer changes the system being observed (Heisenberg)	Declare the measurement’s influence; account for it	Governance overhead is measured (EVALUATION-GAPS.md §11); its cost is a known variable

This maps directly to the Fourth Law (RFC-034 series, DECLARATION.md): the internal observer (first-order, inside the loop) is structurally insufficient. Science’s answer is peer review — the external observer. RFC-035’s answer is the same: qualified observers, declared competence, independent verdict, reproducible evidence.

2.5 What This Position Gives Us

Legitimacy — We are not asking the market to trust a novel framework. We are asking them to trust the scientific method, applied to a new substrate.
Durability — Bespoke AI evaluation frameworks date with each model generation. The scientific method does not. This pattern will not need to be rewritten when the next architecture arrives.
Cross-domain applicability — The same observation pattern works for autonomous drones, clinical AI, financial disclosure, and software supply chains. Because it is how observation works, not how AI evaluation works.
Defense against dismissal — “Just trust the model” is not a scientific claim. It is a faith claim. RFC-035 doesn’t argue with faith — it simply provides the apparatus for anyone who requires evidence.
Unification — Internal and external observation, human and machine observers, software verdicts and AI mediation verdicts — all follow the same epistemological pattern. One structure, many substrates.

2.6 The Starting Point

Everything that follows in this RFC — the four-state model, the gap calculus, the evidence admissibility system, the extraction methods, the observer modes — is derived from what science already knows about producing reliable knowledge under uncertainty.

We start here because this is where reliable knowledge starts. Not with a framework. Not with a standard. Not with a regulation. With the question: how do we know what we claim to know, and what structural apparatus makes that knowledge defensible?

← 1. Abstract · Section index · 3. Ontological Foundation: What Exists to Be Observed →