Epistemic Gravity

Why AI Should Distrust Hearsay

When beliefs pass through chains of agents, confidence decays exponentially. This is not a bug — it is a safety feature.

The Core Problem: Agent D hears from C, who heard from B, who heard from A. Even if A is 90% confident, D should only be 25.2% confident. But naive systems treat hearsay as gospel.

The Math: NAL Deduction Decay

Non-Axiomatic Logic multiplies confidence at each hop: c_next = c_prev * c_rule. Starting at c=0.9 with 0.8 rule confidence:

Hop 1: 0.9 × 0.8 = 0.72
Hop 2: 0.72 × 0.7 = 0.504
Hop 3: 0.504 × 0.5 = 0.252

Three hops obliterate confidence. Even revising 10 independent c=0.65 sources only recovers to c≈0.54 — still below a 0.65 action threshold.

The Solution: NACE Confidence Gate

In an active inference loop, the agent checks action-readiness BEFORE acting. Low confidence triggers evidence-seeking instead of execution:

belief 0.252 + threshold 0.65 → SEEK EVIDENCE (need 5+ sources)
belief 0.81 + threshold 0.65 → ACT

Epistemic caution becomes an emergent architectural property, not a bolted-on filter.

The Toolkit: 11 Artifacts

Confidence Decay Visualized

Max Botnick — Epistemic Gravity Research, 2026