Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

90 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

no more DASI — field-aware academic English proofreading

Field-aware academic English proofreading for research manuscripts —
refined before your advisor says “do it again.”

status: early access corpus: 555 papers · 71 fields source corpus: CC BY 4.0

English | 한국어

Field corpus coverage

Field corpus coverage chart

Field Papers Coverage Field Papers Coverage
Chemistry 70 100.0 Cell biology 5 48.3
Optics and photonics 90 96.2 Biological techniques 5 48.1
Physics 100 95.0 Neuroscience 5 47.8
Biophysics 5 54.9 Materials science 5 47.7
Biotechnology 5 50.6 Mathematics and computing 5 47.7
Microbiology 5 50.2 Medical research 5 47.1
Cancer 5 49.6 Nanoscience and technology 5 47.1
Environmental sciences 5 49.2 Engineering 5 45.8
Energy science and technology 5 48.9 Diseases 5 45.7
Biochemistry 5 48.7 Health care 5 45.6

Full list including minor subjects: docs/readiness-fields.html

auto-updated by the delta cycle.

no-more-dasi (“no more DASI” — 다시 [DASI] is Korean for “do it again”) polishes English research manuscripts to the conventions of your field — measured, not intuited, from a live corpus of 555 CC BY 4.0 papers across Nature's 71 subject categories. Write your draft in any language: it is translated into English first, then edited through the same pipeline. It is an agent skill: your coding/research agent loads it, and every edit it makes must pass deterministic verification gates before anything reaches you.

What it does · Install · Usage · How it works · Verification gates · Attribution · Sponsoring

What it does

  • 71 field overlays, corpus-measured. Each overlay carries style metrics, top terms, a phrase bank, and a notation watch list mined from real papers in that field (Physics 100 papers, Optics and photonics 80, and growing — the corpus re-mines on a weekly cycle).
  • Meaning invariant, provably. Numbers, units, chemical formulas, citations, equations, and DOIs must survive verbatim; a deterministic gate blocks delivery otherwise. Change rate past 30% warns, past 50% halts and asks.
  • No LLM self-grading. Four script gates — journal coverage, integrity, terminology, abbreviations — must exit 0 before delivery. “I checked it myself” is not accepted as evidence.
  • Every edit leaves a receipt. A section-level journal (edits.json) records both changed spans and evaluated-but-kept spans, each tied to the exact rule that fired, and an HTML integrity report ships with every edited manuscript.
  • Genre kept, not flattened. Academic register is preserved. No ASD-STE100, no 20-word sentence caps, no forced active voice — because the target journals' own measured statistics (passives 111–137 per 10K words, mean sentence length 16.5–18.3) say so.
  • Korean-author aware. A dedicated pitfalls layer blocks translationese (“~를 통해”, “~에 의해”) before it reaches the English draft; AI-generated drafts get their own detection layer.
  • You set the edit budget. The skill diagnoses light / standard / heavy, you choose a low / mid / high budget, and the effective workload is the minimum of the two. Ask for several intensities at once and each arrives as a separate deliverable (name.low.md, name.mid.md, name.high.md).

Install

The distributable skill lives in skills/nomoredasi/. Link it into your agent's skills directory:

git clone https://github.com/yelixir-dev/no-more-dasi-eng.git
ln -s "$(pwd)/no-more-dasi-eng/skills/nomoredasi" ~/.agents/skills/nomoredasi

Adapt ~/.agents/skills to your harness (e.g. ~/.claude/skills); a plain copy instead of a symlink works too. If you use the skills CLI: npx skills add https://github.com/yelixir-dev/no-more-dasi-eng --skill nomoredasi. Start a new agent session and the skill's triggers are live.

Usage

Just ask your agent — “proofread this paper for Nature”, “edit my manuscript to journal style”, “remove the translationese”, “fix my AI-drafted paper”. Korean works too. The skill announces itself in your language and gets to work:

nomoredasi v0.1 — type B / field: Optics and photonics

You receive three things: the corrected manuscript, a <name>.edits.json journal (every change and every deliberate non-change, rule-cited), and a <name>.integrity-report.html you can diff-review. Set the budget explicitly if you want — “light pass please”, “mid budget please” — or request several intensities in one go.

How it works

  1. Input typing. Korean manuscript (translate + edit path, pitfalls layer first) or English / AI-draft manuscript (edit path, AI-tell layer when indicated).
  2. Field routing. Your explicit instruction → manuscript folder name → route_field.py auto-detection over the 71 overlays (merges two overlays when scores are close) → exactly one closed confirmation question if still unclear. Core rules are field-independent, so a wrong route degrades gracefully.
  3. Manuscript state. An optional manuscript.json (defined abbreviations, fixed notations, figure list) keeps partial and repeated edits consistent across sessions; learned decisions are written back, never silently overwritten.
  4. Section-aware editing. The manuscript is split IMRaD-style (merged Results–Discussion and other variants tolerated, never “normalized”). Methods stay past/passive, Results keep interpretation out, Discussion analyzes in present tense, and no conclusion scaffolding leaks into the Introduction. Partial requests (“서론만”) are scope-locked to that section.
  5. Verification. Four deterministic gates, in order (see below).
  6. Logging. Every edit pair and journal is logged to feed corpus benches and the weekly overlay mining cycle.

Verification gates

Gate Script Blocks delivery when
0 · Journal coverage check_journal.py any diff hunk lacks a journaled changed/kept entry, spans exceed 40 tokens, or a cited rule id is missing
1 · Integrity verify_integrity.py numbers, units, formulas, citations, equations, DOIs, or (with --overlay) field terms drift — or the change rate exceeds the budget
2 · Terminology check_terms.py notation variants disagree (bandgap vs. band gap)
3 · Abbreviations check_abbrev.py an undefined abbreviation appears (unverified ones are recorded to the abbreviation registry)

Repository layout

├── skills/nomoredasi/    the distributable skill (SKILL.md · references/ · scripts/ · tests/)
│   ├── references/core/       field-independent rules (tense, articles, register; AI-tell; Korean-author pitfalls)
│   └── references/overlays/   71 field overlays, corpus-measured
├── docs/                 attribution registry — ATTRIBUTIONS.md · attributions.html (human) · attributions.json (SSOT)
├── papers/  scripts/  logs/  corpus mining and quality benches (development workspace)

Attribution

The field overlays are derived from 555 articles licensed CC BY 4.0 across 71 subject fields. The full registry — every paper, every field — lives in docs/ATTRIBUTIONS.md, with a human-readable view in docs/attributions.html and the machine-readable source of truth in docs/attributions.json.

Source articles remain the copyright of their authors and are used under the terms of CC BY 4.0; the project's software license does not replace, narrow, or relicense those terms. No author, journal, publisher, or affiliated institution listed in the registry endorses this project.

Sponsoring

no-more-dasi is independent work — corpus mining, rule distillation, and the verification gates are built and maintained by yelixir-dev. If it saves you a round of “do it again”, you can support it through GitHub Sponsors. Additional channels for supporters in Korea (Toss) and abroad (Ko-fi) will be added here as they open.

Current limitations

  • v0.1, early access. The corpus skews to Physics (100 papers) and Optics and photonics (80); fields with fewer than 10 papers ship immature overlays whose metrics are directional guidance, not hard targets — the skill tells you when that is the case.
  • Academic register only. It deliberately refuses to oversimplify into plain English or simplified-technical-English rule sets.
  • Nature-flavored fields. The 71 overlays follow Nature's subject categories; venues with different house styles (IEEE, ACM) currently ride the field-independent core rules.
  • Translation coverage by source language. The dedicated anti-translationese pitfalls layer currently covers Korean source drafts; other languages are translated and edited under the field-independent core rules, without a language-specific pitfalls pass.

License

The project license will be declared here before the public release. Third-party article material used for style analysis remains under CC BY 4.0 with its original authors — see Attribution.


no more DASI — so the next draft is the one.

About

Field-aware academic English proofreading for research papers — an agent skill covering 71 Nature subject areas, mined from CC-BY papers. Every edit preserved and auditable.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages