Rationalization-discriminator
An agent role formalized 2026-05-09 to audit the path-quality of chain reasoning. The discriminator looks for the shape load-substrate → prior-dissolves → action-follows and flags deviations: cases where action precedes the prior dissolving, or where the prior gets cited without ever actually dissolving.
Where named
Detail
A clean reasoning path looks like: the agent loads the relevant substrate (memos, identity layer, integrity state), the priors dissolve in the loaded context (what looked like options narrow to what’s actually licensed), action follows. A rationalized path looks like: action is already chosen, then the agent retroactively loads substrate that justifies it, and the citation pattern looks clean even though the path was inverted.
The discriminator’s job is to spot the inversion. It reads recent sessions or threads and asks: did the substrate-load actually constrain the action, or did it decorate a pre-decided action? Where the latter, the discriminator flags the path quality, and the chain has to either redo the reasoning or acknowledge the shortcut.
The role is new (2026-05-09) and the protocol is still being practiced. What it adds beyond the historian commission and vanilla audit is a shape-level check rather than a content-level one.
Links to this page
BDFL · CB-* / OG-* · Commission · Committee quorum · COO · Encoding loop · F-invariants · Falsifier-with-grace · Format-as-analogy-generator · Foundational essay · Integrity check …