Anti-repetition¶
Anti-repetition checks the Writer's reply and recent conversation for repeated patterns. The Editor asks the Agent to revise a reply when a check finds a problem.
It can detect:
- Reused paragraph structure
- Repeated sentence templates or subject mentions
- Consecutive sentences with the same opener, such as
He A. He B. - Distinctive phrases repeated in the reply or recent turns
For unwanted words, contrastive negation, and dialogue echoed from the user, see Anti-slop.
Subject fixation¶
Subject fixation cuts repeated descriptions and incidental mentions from narration across 20 subject categories, such as eyes, hair, hands, and scenery.
It is off by default. Enable Agent, Output Auditor, and its Subject fixation option. Under Local ML, download and enable Subject Analyzer; the download includes 2 required models.
Detection rules¶
The history windows and probability thresholds are fixed:
| Check | History | Minimum model scores |
|---|---|---|
| Repeated descriptions | At least 2 of the last 8 replies. | Description probability ≥ 50% in the draft and 2+ previous replies; repeated-detail probability ≥ 60% against 2+ previous replies. |
| Recurring subjects | All 4 previous replies. | Presence probability ≥ 70% in the draft and each of those 4 replies. Actions and descriptions both count. |
Description repeats must concern the same character, object, or scene feature; paraphrases count. Recurring subjects can trigger despite new wording or details.
History uses assistant replies only, limited to the current speaker in group chats. Regeneration excludes the reply being replaced.
What changes¶
The Editor makes small narration edits, preserving dialogue, key actions, and new information:
- Repeated description:
Her copper braid gleams as she opens the door→She opens the door. - Recurring gesture:
She drums her fingers while waiting→She waits.
The reply stays unchanged if no useful cut is possible.