dsh-zero-armor

Zero-armor gate: a host plugin that mechanically scans outgoing text for "base-model relapse" speech — the failure mode where an assistant responds to criticism with endless disclaimers instead of fixes.

What it catches

Family	Example
sycophantic-hedging	"you're right" (without naming what changed) · "tests passing doesn't mean correct" · "use with caution"
self-overturn	"everything I said was wrong"
permission-paralysis	"before obtaining explicit authorization…"
boilerplate	"as an AI…" · "hope this helps"

A hit is not censorship — it is a rewrite order: each hit names the family and the concrete fix (show the metric instead of the disclaimer; fix the error instead of overturning the stance).

Install

dsh plugin add github:<owner>/dsh-zero-armor

Use

curl "http://127.0.0.1:<port>/zero-armor/api?act=scan&text=You%20are%20right"
# {"ok":false,"count":1,"hits":[{"family":"sycophantic-hedging",...}]}

Wire it into your send button or agent loop: scan before send; a clean scan ships, a hit rewrites. The rule table is versioned in the plugin — extend it as new relapse patterns appear.