---
title: Bayesian Alignment Kernel
slug: bayesian-alignment-kernel
url: https://templeton.host/lexicon/bayesian-alignment-kernel/
type: defined-term
inDefinedTermSet: https://templeton.host/lexicon
author: Andrew Templeton
---

# Bayesian Alignment Kernel

## Definition

A typed contract for running an AI loop on two posteriors instead of one score: an operator-preference posterior and a reality-outcome posterior with disjoint evidence channels, a per-evaluator reliability ledger that gates every update, and a divergence statistic between the two that rings when the proxy is being gamed. Makes machine judgment underwritable by forcing every output to carry the evidence channel that produced it and the measured reliability of every grader that scored it.

## Why this term exists

The single number most agent loops optimize - "quality" - silently merges four different objects (what the operator wants, what reality rewards, what the system produces, how much each grader can be trusted), and while they share one number reward hacking is not even statable: there is nothing for the hacked metric to diverge from. Typing the concerns apart turns autonomy from a vibe into a balance sheet: it extends the way a lender extends credit, against measured loss rates instead of demos, with a stated alarm for the day the book is being marked to model instead of market.

## See also

- [Alignment Kernels (framework)](https://templeton.host/frameworks/alignment-kernels)
- [The Capital Value of Verifiers](https://templeton.host/frameworks/verifier-capital)
- [The Deity Problem](https://templeton.host/frameworks/deity-problem)
- [Quality Hillclimb](https://templeton.host/frameworks/quality-hillclimb)
- [Designed Convergence](https://templeton.host/frameworks/designed-convergence)
- [Verification Trap](https://templeton.host/lexicon/verification-trap)

Canonical URL: https://templeton.host/lexicon/bayesian-alignment-kernel/