Rigorloomv0.17.0 · MIT

When an AI edit changes a number, a quotation or a negation, that paragraph goes back to the original.

Rigorloom creates Korean HWPX documents, fills in forms and edits text. AI polishes the sentences. Whether the facts are still the same is decided by a checker that follows fixed rules.

source on GitHubreleasesdesktop app in development

try it

the document is in Korean, because Rigorloom works on Korean text.

Below is a copy of a lab report with an AI’s proposed edit to paragraph 4. Pick a proposal, or rewrite it yourself. The check runs as you type.

실험보고서_사본.hwpxWaiting

3실험은 같은 과제를 두 조건에서 반복해 측정했다. 참가자는 각 조건을 무작위 순서로 수행했다.

4Needs rework

5다음 절에서는 조건별 오차를 비교한다.

Checking paragraph 4

Expressions that must not change are pulled from the original and the proposal and compared.

ItemOriginal → proposalResult

Applied
how it works

An AI finds awkward or off-register sentences and proposes rewrites. The AI that proposes and the AI that reviews run separately.

Then a checker compares each paragraph before and after. No AI is involved in this step. If anything differs, only that paragraph is restored and flagged for rework. Edits to the other paragraphs are kept.

the paragraph

Taken from a copy. The original file is never edited.

proposer

An AI writes an edit.

reviewer

A second AI, run separately, looks it over.

checker

No AI here. Fixed rules compare the paragraph before and after.

applied

Everything that had to stay the same did. The edit is kept.

restored

Something changed. Only this paragraph goes back to the original and is marked for rework.

what it checks

If any of these changes, the paragraph is reverted.

ItemReverted if it changes like this
Numbers18.6% → 19%
Units250 mL → 250 L
Dates2026년 3월 4일 → 3월 초
Quotations"일반화할 수 없다" → "일반화하기 어렵다"
Citations(Kim, 2021) → (Kim, 2021) removed
Formulas$v = d / t$ → $v = t / d$
Headings3. 실험 방법 → 3. 연구 설계
Negation차이가 없었다 → 차이가 있었다
Hedges약 30명 → 30명
Quantifiers모든 조건 → 대부분 조건
Causal words온도가 높았기 때문에 → 온도가 높았고

Examples are in Korean because Rigorloom works on Korean documents. URLs, links, source IDs and figure and table tags are compared too. The checker leans strict, so an edit that keeps every fact can still be reverted. We would rather look twice than miss a change.

document work

The original is never touched.

Fill in forms

Finds the blanks in official letters, applications and grant forms by address or field name, and fills them in.

Edit text

Replaces, inserts, deletes and reverts sentences in a finished document.

Keep a record

Every change is applied to a copy, with a record of what changed.

실험보고서.hwpx          original, unchanged
실험보고서_사본.hwpx     edits applied
실험보고서_변경기록      para 4 restored, para 7 applied …
CLIwith a manual written for agents
Desktop appin development, not yet released
MCPconnect it to AI tools
status
v0.17.0

Public release. Available on GitHub now.

next

In progress: the next public version.

Producing the final output for submission requires Hancom Office.

The verdicts above come from the same patterns and word lists as Rigorloom’s actual checker (prose_fidelity.py), ported to run in the browser. Proposals A and B were written by hand for illustration, not by an AI.