Report structure
Checks required fields such as sample number, lot number, test date, test method, analyst and reviewer.
Chemical test report quality review
Keep values intact.
Catch what matters.
Kurage Proofread reviews Japanese chemical test reports for missing information, inconsistent sample and lot identifiers, specification/result relationships, units, pass/fail logic and missing OOS context. Local Gemma 4 adds semantic review while a human reviewer keeps final control.
Sign in with the shared Kurage X login. Files and review history are separated by user.
Direct answer
What is Kurage Proofread?
It is an AI-assisted proofreading and verification workspace designed for Japanese chemical test result reports. It detects review points and produces a reviewed copy plus a separate review report. It does not decide or rewrite scientific results.
Review coverage
One workflow, three layers of review
Deterministic checks handle facts and relationships. Japanese writing checks improve readability. Gemma 4 reviews meaning and context.
Numeric consistency
Compares specification ranges, measured results, units and pass/fail statements without changing them.
Identity consistency
Finds conflicting sample IDs, lot IDs and dates across paragraphs, tables, sheets and cells.
OOS context
Flags out-of-specification results that lack discussion, deviation records or an investigation trail.
Japanese proofreading
Uses textlint-based technical writing checks for unclear, unfinished or inconsistent Japanese expressions.
Semantic AI review
Local Ollama Gemma 4 reviews contradictions between prose and tables, missing discussion and logical gaps.
System specifications
Current production specification
The following reflects the running system, not a future roadmap.
- Input formats
- DOCX, XLSX, PDF, TXT and CSV. Maximum file size: 25 MB.
- AI engine
- Ollama with gemma4:12b-it-qat, running on an EXBRIDGE-operated machine.
- Rule engine
- Python-based deterministic chemical report checks, Pandera data validation, Pint unit handling and Japanese textlint rules.
- Output
- A reviewed DOCX/XLSX or generated DOCX, plus a separate DOCX review report with severity, location, reason and suggestions.
- Access
- Kurage shared X login, CSRF protection, internal API authentication and user-separated review history.
- Human approval
- Required. The system presents review findings; the responsible reviewer confirms scientific and quality decisions.
Safety boundary
AI assists the reviewer. It does not become the reviewer.
Kurage Proofread intentionally separates writing corrections from scientific judgement.
- Measured values are never auto-changed.
- Specification values and unit conversions are never auto-confirmed.
- Pass/fail decisions remain with the responsible reviewer.
- Every finding includes its location and reason for review.
FAQ
Frequently asked questions
What is Kurage Proofread?
A web-based review workspace for Japanese chemical test reports. It combines deterministic validation, Japanese proofreading and local Gemma 4 semantic review.
Does it automatically correct measured values?
No. Measured values, specifications, unit conversions and pass/fail decisions are only flagged. A qualified human reviewer decides what to change.
Which formats can I upload?
DOCX, XLSX, PDF, TXT and CSV files up to 25 MB. DOCX and XLSX retain their native output format; PDF, TXT and CSV reviews are exported to DOCX.
Is the report sent to an external LLM?
The production AI review uses Ollama and Gemma 4 12B on an EXBRIDGE-operated machine. The source file itself is not sent to an external LLM provider.
Review the report without surrendering the decision.
Upload a report, inspect evidence-backed findings, and download both the reviewed file and the review report.
Open Kurage Proofread