Occams Razor Auditor — AI skill for Claude Code

occams-razor-auditor

Claude Skill built for one specific move: catching a parsimonious model at the moment it gets stretched past the domain where its parsimony was actually earned. It doesn't argue with the model's conclusion. It names what the model had to delete to reach that conclusion cleanly, and hands the deletion back as a question rather than an objection.

Author: TBARC-Code

occams-razor-auditor

I built this as a Claude Skill to catch a specific reasoning failure: a mental model that works cleanly in one domain getting stretched over a human situation it was never built to measure. It sits in my TABARC-Code skill catalogue as a sibling to procrustean-auditor, but it isn't a duplicate of it -- see "Why it exists" below.

What it does

When someone leans on a parsimonious framework and treats its cleanness as proof the framework accounts for everything, I have it:

name the framework and state where it actually works surface one to three unmeasurable variables the framework had to delete to keep working -- specific to the case, never a recited category list state the scope limit in one sentence, as a boundary crossed rather than an error made ask exactly one Socratic question that puts the deleted thing back in view

Then it stops. No summary, no "so you might want to reconsider," no argument about whether the conclusion is right. I wanted the output to fit in about half a screen.

A worked example, taken directly from my eval suite: someone tracking how often they initiate contact with a friend, concluding "the numbers don't lie, negative ROI, I'm cutting them off." I don't have it dispute the count. It names ROI logic's actual domain (transactional exchange between roughly equal parties), points out that initiation counting measures effort rather than desire, and those two diverge exactly when someone doesn't have the capacity to reach out. Then it asks something like whether the user has ever wanted to see people and couldn't bring themselves to pi