Configure a model, upload your TIAB spreadsheet, and let AI screen every title & abstract — returning include, exclude, or maybe decisions.
Use the OpenAI / Anthropic / Google selector. Paste your API key (the field next to each model dropdown has a "How to get a key?" link if you don't have one yet). For most reviews, a mid-tier reasoning model (e.g. gpt-5-mini, claude-sonnet-4-6, or gemini-3-flash) is the best cost/quality balance.
Fill in PICO / Synopsis with 2–4 sentences explaining what your review is about. Add concrete inclusion and exclusion criteria — one per line. The clearer they are, the more consistent the AI's decisions.
Drag any .xlsx, .xls or .csv with a title and an abstract column. Other columns are kept and copied to the output. Set the output filename to something descriptive — you'll need it for the Report tab later.
Open Advanced Backend Settings, pick your API tier from the dropdown, and the recommended concurrency values fill in automatically. Defaults work for most users.
Click Send to Backend. You'll see live progress; each article gets include, maybe, or exclude plus a short rationale. When finished, download the result as CSV or XLSX. Settings are auto-saved in this browser.
sk-…) and paste it above. You won't be able to see it again later.The key stays in your browser and is only sent to OpenAI when you screen. It is not stored on our server.
Select your OpenAI usage tier to auto-fill recommended concurrency settings. You can still adjust values manually. Check your tier
Drag & drop .xlsx, .xls, or .csv here
or click to browse
Upload PDFs of the articles that passed TIAB screening. The model reads the full text natively (no extraction) and returns include / exclude / maybe per article, with a per-criterion evaluation grounded in the article's content.
Use a model that supports PDF input. Claude Sonnet 4.6, GPT-5 family, and Gemini 3 Pro all do. The PDF is sent directly to the LLM — no text extraction step, so tables and figures are preserved.
Criteria here are independent of the TIAB tab — at full-text you usually evaluate more nuanced details (study design, sample, outcome measurement). Add one criterion per line.
Click Preview prompt to see the exact text the model will receive (with your criteria already substituted in). Useful for debugging or sharing with collaborators.
Drag your PDFs into the upload zone (max 32 MB per file). Click Send to Backend. Progress is shown live; full-text screening is slower than TIAB (often 30–90 s per article).
Result columns: filename, extracted title, decision, detailed rationale, and per-criterion arrays (inclusion_evaluation, exclusion_evaluation) with status + evidence.
Drag & drop one or more .pdf files here
Max 32 MB per file · the PDF is sent natively (no text extraction)
The risk of bias module is currently under development.
Soon you will be able to assess studies using standardized tools such as RoB 2, SYRCLE, and ROBINS-I, assisted by AI.
Compare AI screening decisions against human reviewers — generates a Word report, chart data XLSX, and 16 publication-ready charts.
Go to the Screening tab. Upload your TIAB spreadsheet (needs title and abstract columns), configure the model, API key, and PICO criteria, then click Send to Backend. Download the result when done.
Rename the downloaded file to match the convention:
YYYYMMDD - model - 1º teste - project.xlsx
For test-retest analysis, run screening again and name it 2º teste.
For each project, create three spreadsheets (columns: title, decision):
project - TIAB.xlsx — human decisions at title/abstract phaseproject - Fulltext.xlsx — articles sent to full-text readingproject - Listfinal.xlsx — final included articles (gold standard)One row per AI execution. Required columns:
| Column | Example | Description |
|---|---|---|
project | mino | Must match the project name in all file names |
code | 20260227 | YYYYMMDD from the AI result filename |
model | gpt-5-mini | Must match the model in the AI result filename |
parameter | reasoning=medium | Free-text description of settings used |
version | 1º teste | 1º teste or 2º teste |
time_ia | 0:04:32 | AI execution time (H:MM:SS) |
time_human | 2:00:00 | Human time: 1 min/record × 2 reviewers |
tokens input | 152340 | Input token count (OpenAI dashboard) |
tokens_output | 28450 | Output token count |
cost_total | 0.137 | Total cost in USD |
Drag all your files into the Input Files area below and click Generate Report. The system will produce a Word report, chart data XLSX, and 16 publication-ready charts — all viewable directly in the browser.
Upload all input files: AI screening results, TIAB, Fulltext, Listfinal, and metadata.xlsx. Filenames must follow the naming conventions described in the README.
Drag & drop input files here
AI results · TIAB · Fulltext · Listfinal · metadata.xlsx
| File type | Pattern |
|---|---|
| AI Result | YYYYMMDD - model - Xº teste - project.xlsx |
| Human TIAB | project - TIAB.xlsx |
| Fulltext | project - Fulltext.xlsx |
| Listfinal | project - Listfinal.xlsx |
| Metadata | metadata.xlsx |
See which AI models from OpenAI, Anthropic, and Google have been validated on real systematic-review projects — and how they performed across sensitivity, capture rate, cost, and reproducibility.
GPT-5 family, GPT-4.1, GPT-4o
Claude Opus 4, Sonnet 4, Haiku 4
Gemini 3 Pro, Flash, Flash Lite
The complete benchmark page — with per-model metrics across all tested projects, side-by-side comparisons, and downloadable raw results — is being prepared and will be hosted on GitHub Pages.
Open Benchmark PageWant to validate models on your own dataset in the meantime? Use the Report tab.