slopless: Instructions file for Codex

Slopless V2

Goal

• Develop cheap, fast decision-model checks that generalize Slopless writing patterns. Scope tasks to named checks; measure missed slop and false positives by family and issue type.
• This document specifies the target workflow. Inspect the current schema and runners before implementing it; report which parts are operational.

Product Scope

• Target standardized sloppiness checks with fixed definitions and tested defaults; CLI flags select checks. Business-specific terminology belongs to the separate existing Python workflow.
• Judge the supplied passage or full text. External fact-checking, source verification, and decisions about a particular reader's prior knowledge belong to the separate editor.
• Keep readability measurement and terminology-policy enforcement outside this V2 iteration.
• Compare Jev with the old CLI before deciding which deterministic checks remain useful and whether textlint support should continue.
  • Base decisions on labeled evaluation, later corpus validation, highlighting quality, latency, and cost. Preserve production behavior while evaluating the replacement.
• Read the accepted V2 taxonomy when assigning families and issue types. It maps existing checks and records Humanizer additions and exclusions.
  • Use its consolidated semantic-thinness definitions for that family's datasets and checks. Treat the legacy identifiers as coverage sources rather than separate issue types.
  • For protected authoring and review, supply extracted issue definitions and an explicit permitted-input list. Keep the full taxonomy, development material and evaluation summaries outside those contexts.
• Exclude decorative-bold checks. Judge redundant label wording independently of its formatting.
• Declare input: passage or input: text in each finalized issue definition. The submitted writing is the complete input.
  • A passage can include several paragraphs. A text is the complete article, story, or reply; text-level checks use it intact.
  • Classify buried answers and repeated explanations within the submitted text as text-level checks. Keep density reporting separate from this input requirement.
• Use native paragraph ranges for current V2 highlighting. Paragraph checks report their input paragraph; whole-text checks receive the complete numbered text and ask independent paragraph-location questions in the same request as detection.
  • Batch all enabled paragraph checks over one shared paragraph. Each question carries its own rule definition, examples and judging instructions, and returns an independent probability. Split question groups only for model or request-size requirements.
  • Record detection/location disagreement and incorrect paragraph selections as known defects. On 2026-10-26, the user accepts the measured precision for proceeding with integration; preserve raw answers and existing labels.
  • Retain the measured whole-text detection gate for whole-text reporting. Exact offending-span selection remains outside this shipping step; paragraph ranges narrow where the writing agent should look.