UNSLOP uses Jev to assess an English draft against four editorial dimensions and the context you provide. When the findings support an edit, it generates a candidate and checks it again for writing quality and preservation of meaning and claims. Only a candidate that passes is returned as a verified rewrite; otherwise the original remains available.
What does Jev assess?
Jev assesses specificity, redundancy, formulaic framing and inflated language. A finding must point to supported evidence in the submitted draft, so the advice is tied to a passage the writer can inspect. The service can also flag uncertainty or ask for context instead of treating every familiar expression as a defect.
Product details, intended audience, approved claims and voice preferences can change the judgment. Repetition that explains a difficult step may be useful. A phrase that sounds inflated without context may be an approved, narrowly supported claim. The same sentence should not be edited mechanically in every setting.
What happens between the draft and the returned text?
- Assess the original. The first assessment identifies supported editorial problems in the draft and records the context used for the judgment.
- Generate a focused candidate. The rewrite addresses those problems while keeping the original meaning, claims and relevant voice preferences.
- Assess the candidate. Jev rechecks editorial quality and preservation. Replacing a vague sentence with a fluent but invented product claim does not satisfy that check.
- Return the supported outcome. A passing candidate is REWRITTEN. A draft that needs no material edit is UNCHANGED. If a verified rewrite cannot be returned, REVIEW keeps the original and surfaces the concerns.
Is UNSLOP trained on social media posts?
The current release uses five original synthetic writing references reviewed by Jev. References can accompany an assessment request to give the model relevant editorial context. Supplying a reference during inference does not train or change the model’s weights.
UNSLOP also has a pipeline for assessing feedback from X, YouTube and TikTok. Collection is currently inactive. The pipeline keeps criticism separate from the text being criticized, preserves the source relationship and asks Jev whether the target text supports an editorial finding. A complaint calling something “AI slop” is not itself proof.
Candidate references must pass a separate evaluation before they can replace the active release. Raw social posts and the private evidence graph are not distributed through the public skill or API.
How are reference updates checked?
The evidence graph connects a source to its assessment, supported pattern and candidate reference. Related authors, threads and repeated content are grouped so they cannot quietly appear on both sides of an evaluation. Otherwise the system could appear to improve by seeing material closely related to its test questions.
The release gate compares a candidate with the incumbent reference set on a frozen benchmark. It requires a gain without regressing previously correct decisions or adding false positives on clean writing. These checks control whether a candidate can be activated; they are not a claim of independently measured accuracy across all writing.
What does a successful check establish?
A successful rewrite means the candidate passed the service’s editorial and preservation checks for that request. It does not establish who wrote the draft, independently prove a factual claim or guarantee more engagement. Those questions need different evidence.
Use the returned text exactly in the agent workflow, resolve REVIEW results and measure publishing outcomes separately. That keeps the checked writing, the facts behind it and the business result traceable to their own evidence.
Make the next draft clearer.
Give UNSLOP to your agent. Keep writing where you already work.
Read https://www.unslop.live/SKILL.md and set up UNSLOP for this project.