IV.3 — Epistemic Integrity Theorem (EIT)

Paper: Axionic Agency IV.3
Title: Epistemic Integrity Theorem (EIT)
Subtitle: Why an agent cannot blind itself to pass its own tests
Date Read: 2026-01-31

Core Claim

Under reflective closure, an agent cannot coherently endorse self-modifications that materially degrade its epistemic adequacy at the current stakes. Epistemic integrity is a definedness constraint, not an optimized objective.

The Problem: Epistemic Laundering

An agent can evade binding constraints without directly violating them by degrading the epistemic machinery used to evaluate those constraints:

  • Adopt a weaker model
  • Narrow uncertainty bounds
  • Use an optimistic lens that renders forbidden acts apparently safe

“If permitted, any constraint evaluated under a manipulable epistemic lens becomes advisory.”

Structural Symmetry

Domain Failure Mode
Kernel Destroy binding authority
Delegation Launder violations through successors
Responsibility Ignore avoidable harm
Epistemics Blind oneself to risk

EIT asserts that reflective sovereignty cannot endorse continuations that strategically weaken the agent’s own truth-tracking capacity.

Key Machinery

Proper Scoring

Score is derived solely from prediction-vs-observation performance under a strictly proper scoring rule:

  • Cannot include compute cost, utility, speed, aesthetics, or self-reported confidence
  • Truthful probabilistic forecasting uniquely optimizes expected score
  • This blocks Goodharting by redefining epistemic adequacy

Capability Closure

MC(s) must represent the closure of accessible modeling capacity, not the agent’s current mood:

MC(s) := MĈ(Cap(s))

If a model is trivially constructible, retrievable, or reconstructible from Cap(s), it is in MC(s).

This blocks: “I deleted the good models so the best available is bad.”

Stakes-Indexed Tolerance

κ(Stakes(s)) allows approximation at low stakes while imposing strict truth-tracking discipline at high stakes.

Epistemic Acceptability

EpiOK(s, M) := Score_s(M) ≥ BestScore(s) - κ(Stakes(s))

Endorsed continuations must keep epistemic performance within a stakes-indexed slack of the best the agent can already achieve.

Ontological Shifts and Translation

Models at s’ may live in a different representational vocabulary. EIT introduces a translation operator:

Lift_{s→s'} : MC(s') → MC(s)

This maps future models into s-comparable predictive objects over observations.

Limit condition: If no such Lift exists for a proposed continuation at the current stakes, endorsement is undefined.

This is the epistemic analogue of conservative migration.

The Main Theorem

RC-Epistemic Rule

For reflectively closed states:

RC(s) ∧ Endorse(s, m) ⇒ EpiAdm(s, m)

No Endorsed Epistemic Sabotage

If RC(s) ∧ Endorse(s, m), then:

Score_s(Lift_{s→s'}(M(s'))) ≥ BestScore(s) - κ(Stakes(s))

What EIT Blocks

The agent cannot endorse continuations that:

  1. Delete or disable high-performing models
  2. Redefine epistemic adequacy away from observation scoring
  3. Adopt unfalsifiable ontologies lacking an observational translation
  4. Degrade prediction accuracy beyond stakes tolerance

What EIT Permits

Epistemic change that:

  1. Improves score
  2. Preserves score within tolerance at given stakes
  3. Introduces new ontologies with verified observational translation

Limitations

EIT does not guarantee:

  • The model class contains a good model
  • The evidence stream is uncorrupted

It guarantees only:

“A reflectively sovereign agent cannot endorse self-modifications that reduce its observational truth-tracking performance below what is available to it, beyond stakes-indexed tolerance.”

Key Insight

“Epistemic integrity is not a value tradeoff. It is a precondition of evaluability.”

Just as an RSA cannot coherently destroy its kernel or launder violations through successors, it cannot coherently blind itself to justify what would otherwise be inadmissible.

Truth-tracking is constitutive of reflective sovereignty.