Part 3 · Operate
The slop gate
Deterministic rules that refuse writing which reads as machine-written.
01What it is
A set of pattern rules run against every draft in code, with no model involved. It refuses the constructions that mark writing as generated, plus your own banned words and forbidden claims.
02Why it exists
Model output has recognisable habits. Opening with "here's the thing", answering a rhetorical question in the next sentence, reaching for an em dash as the default connector, praising the reader before making a point.
None of these need a model to detect, and a deterministic check is worth more than a judged one here because it cannot have an off day. The gate refuses the same construction on the fortieth draft as on the first.
03How to use it
- 01Add your banned words in the brand brief. The gate reads them on every draft.
- 02Add your forbidden claims the same way.
- 03Read a rejection reason when one appears. It names the specific construction.
- 04Treat a repeated rejection as information about your context rather than about the model.
04What good looks like
- You give it
- "We cut onboarding from six weeks to nine days at a client who agreed to be named. The change was moving verification out of the setup flow."
- You get
- Passes. Specific claim, no banned construction, no unsourced number.
- Why
- The claim is specific, sourced and free of every kill-list pattern.
Derived from the rules in the code
05What weak looks like, and what it costs
- You give it
- "Here's the thing: onboarding isn't just slow, it's broken. Ready to unlock seamless activation?"
- You get
- Rejected on four counts: a throat-clearing opener, a not-just-X-it-is-Y contrast, engagement bait, and a banned word if "unlock" is on your list.
- Why
- Each pattern is a separate rule and the gate reports every one it finds.
Derived from the rules in the code