Skip to content

Atomisation (no sentences in fields) is mostly prompt discipline #18

Description

@chouswei

Summary

Guides insist one idea per row and no prose in fields, but the engine does not generally reject sentence-length field values. Quality depends on LAW text in warm plus skill memory.

Evidence

  • Docs: LLM-GUIDE.md atomisation checklist; application-notes/llm-novel-writer.md LAW-ATOM01 as prompt discipline ("Engine enforcement | Partly... | Prompt discipline only").
  • Engine checks: exact field counts, id patterns, relation vocabulary, JSON rejection (sanitiser.py) — not "is this a sentence?".
  • Contrast: novel domain adds many LAW rows prepended on warm; generic MemNet sessions may lack them.

Impact

Graphs bloat; warm slices become paragraph dumps; token-efficiency claim fails when agents ignore discipline.

Suggested direction

Optional soft/hard caps (max chars per field, max tokens per warm row) with @WRN / @ERR; or a housekeep lint for fat fields.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions