The autopilot: a whole review with two AI reviewers
One click takes a review from imported records to the PDF with no human step. What it does, how it keeps going for hours, and what people should still check.
Overview > Run the whole review with two AI reviewers takes a review from the imported records to the built PDF with no human step.
What it does, in order
- Merges duplicates.
- Has two AI reviewers screen titles and abstracts, blind, two votes per record.
- Lets them discuss their conflicts.
- Settles what remains by the review's rules (Config > Autopilot). The methods report them as rules, not as human decisions.
- Retrieves open-access PDFs.
- Screens full texts the same way.
- Designs the extraction form and extracts the data.
- Rates the prior surveys.
- Writes the paper and builds the PDF, checking every page. Problems in a section's own text are sent back to be fixed, and the PDF is built again.
Built to keep going
A run takes hours, so:
- Every stage only does what is left. Stopping and resuming never repeats work.
- A stage that does not finish is tried again after a pause (1, 5, 15, then 30 minutes), up to the number of attempts set in Config; then the run stops and waits for you.
- Each stage attempt runs in a fresh process, so an update to the app applies from the next stage, and a crash costs one attempt, not the run.
- A watchdog relaunches the run when it stops showing signs of life, for example after a reboot.
- If the AI reaches its usage limit, the run waits until the limit lifts.
- Stop cancels the current AI job, and no further job starts.
The pair of AI reviewers
Each reviewer's model is recorded with its work and named in the paper. On one model, two reviewers differ only in their instructions and agree on almost everything; two different models make their agreement meaningful. A change to the pair applies to the records not screened yet.
Before you submit
The paper reports the autopilot's decisions as AI work that no person has verified yet. People should still check the AI's decisions before submission: screen a sample, look at the conflicts the rules settled, verify the extracted values against their quotes, and read every section.