Systematic review software that keeps the PRISMA flow live, with two blind votes on every record and every number tied to its page
Two blind votes on every record, a person on each conflict, every number tied to its page.
See one case, screen by screen ↓Systematic Review & Evidence Synthesis is systematic review software for library evidence synthesis services and review teams. Agents capture the criteria from the registered protocol, re-run the saved searches monthly, vote blind on every title and abstract, extract data pinned to page and table, re-derive every effect size and draft the PRISMA 2020 flow and checklist. People decide every conflict, confirm every exclusion and value, and the principal investigator signs the flow before it exports.
AI in education: a buyer's guide to agentic AI for colleges, universities and school districts →
Updated
Systematic review software is meant to carry a review from a registered protocol to a published PRISMA 2020 flow diagram, yet a systematic review still takes 67.3 weeks on average from registration to publication. A university library's evidence synthesis service turns the protocol into searches across bibliographic databases, removes duplicates, organizes dual title and abstract screening of thousands of records, chases full texts, extracts data, judges risk of bias and reports the whole journey in the flow and the 27-item checklist.
Today that work is spread over five or six tools, spreadsheets and email. Searches live in database histories and go stale before submission; counts live in spreadsheets, and the flow is redrawn by hand at the end, so its numbers often do not add up. A sample size copied from the abstract instead of the results table quietly changes a study's weight in the meta-analysis. And since the 2025 RAISE recommendations, any automated step that makes or suggests a judgment has to be reported, with how people oversaw it. Systematic Review & Evidence Synthesis puts the review on one living PRISMA flow, where agents propose, quote and cite, and people decide.
Estimated split of the weeks a library team spends on the stages this solution supports, by hand and with the agents; the 'after' column keeps both human decisions on every record and is in line with the estimated 50–70% cut in screening and extraction time. Our arithmetic, kept modest; journal peer review is not included.
Nine agents search, screen blind in pairs, extract and re-check every number; reviewers decide every conflict and the investigator signs the flow.
"Effects of high-dosage tutoring on middle-school math achievement", protocol REG-2026-0318, registered March 2, 2026, is one of the reviews Halden State University's evidence synthesis service runs. Here is Wednesday, October 21, screen by screen, in the working solution.
Every box is a count Priya can open: 6,412 records identified, 1,431 duplicates removed, 4,981 screened, 4,542 excluded with 61 awaiting a person, 378 reports sought, 369 assessed. A chip between the screening boxes reads "112 conflicts · 3 open", and the rail beside the flow holds the review question and the six inclusion criteria, each cited to its protocol page.
One next step: decide 3 screening conflicts, due October 26.
The Search agent re-ran the six saved strategies across five bibliographic databases and one trials register and found 86 new records. It removed 19 duplicates; the Screener agents voted on the rest: 58 agreed exclusions, 6 to full text and 3 conflicts. The increment sits on the top box of the flow as "+86", so the review stays a living systematic review without a manual update cycle.
For Moreau & Tanaka 2025, Screener A votes include at 0.82, quoting "two to four sessions a week"; Screener B votes exclude at 0.84, quoting "mean attendance was 2.3 sessions a week". The two run on different models, and their votes were sealed until both were in. The Adjudicator names the criterion they disagree on, C3 dosage, cites protocol page 5, and routes it to Dr. Imani Rhodes, who decides with a reason.
The Adjudicator has not decided any of them; they are yours.
The 58 records both agents excluded on October 1 wait as one batch, grouped by reason: 17 not a tutoring intervention, 15 outside grades 6–8, 10 below three sessions a week, 7 with no comparison group, 5 with another outcome, 4 of an ineligible design. Daniel reads a 10% random sample of six records in full before he confirms the batch.
At full text, both Screener agents included Vellacott 2023, a randomized trial from the October 1 search. The Checker agent re-derives every sample and effect size from the raw numbers and finds the abstract reports 212 students on page 1 while Table 2 on page 6 sums to 198: 101 tutored and 97 in the comparison group. Both passages light up beside the extraction form, and the study stays out of the synthesis.
The Extraction agent proposed 212. The Checker never overwrites a value; it flags.
Dr. Rhodes records 198 as the analytic sample: 212 were randomized and 14 left before the spring test, 6.6% overall with a 0.2-point difference between arms. The included box ticks from 40 to 41 with her initials, all 12 values on the form are verified, and the Checker's re-derived effect size, g = 0.27, matches the paper's Table 3.
Kessler & Ahn 2024 offered tutoring twice a week. Dr. Rhodes picks the protocol reason, dosage below 3 sessions a week (C3), and the exclusion waits for Daniel Ruiz to confirm it. The person who picks a reason never confirms their own.
For Vellacott 2023 the Bias agent drafts the RoB 2 domains with quoted passages: a computer-generated sequence held by the district research office (page 4), 91% of sessions delivered (page 7), 14 of 212 missing, 7 per arm (page 6). The analysis plan was not publicly registered, so the overall draft is Some concerns. Two reviewers confirm, and either can change a judgment with a reason.
The forest plot pools 36 studies at g = 0.29 (95% CI 0.22 to 0.36, I² = 41%, random effects) on the sample data. Beside it, the Checker's findings across the included studies: 7 found and 7 resolved, among them a standard error entered as a standard deviation and a sign reversed.
Priya prepares the export and the counts balance from 6,412 to 41. Dr. Okafor signs the PRISMA 2020 flow; the 27-item checklist reads 27 of 27 reported, and the AI-use statement for the methods names each agent, its prompt version, the 300-record calibration and the 112 disagreements decided by a second reviewer. The flow, checklist, search log, statement and audit trail export together for the journal.
The same review, seen by the people who carry it — what their day looked like, and what it looks like now.
Nine agents search, screen blind in pairs, extract and re-check every number; reviewers decide every conflict and the investigator signs the flow.
Reads the registered protocol and captures the review question and the inclusion criteria, each cited to its page, and spots amendments.
Re-runs the saved strategies monthly across five bibliographic databases and one register, removes duplicates and fetches full texts through the library link resolver, requesting interlibrary loan for gaps.
Votes include or exclude on every title and abstract against the criteria, quoting the abstract and citing the criterion.
The second, independent vote on every record, on a different model so the two screeners do not share blind spots.
Quotes both rationales side by side, names the criterion they disagree on and routes the conflict to the second reviewer.
Fills the data extraction form with every value pinned to its page and table: sample, dosage, grade, outcome, means and standard deviations.
Re-derives every sample and effect size from the raw numbers and flags any mismatch, lighting both spots in the paper.
Drafts risk-of-bias judgments by domain, RoB 2 for randomized trials or ROBINS-I for non-randomized studies, with the quoted passage behind each.
Drafts the PRISMA 2020 checklist locations, the flow export and the AI-use statement for the methods section from the record.
Decide, verify and sign. The agents propose; a named person decides.
Ask about any record, vote, value or count in the review. Every answer cites its source, and the assistant opens the screen where the answer lives.
Why did the screeners disagree on the 3 open conflicts?
All three came in with the October 1 search, and each disagreement sits on one criterion. Moreau 2025: Screener A read "two to four sessions a week"; Screener B read the mean of 2.3, and the protocol judges the mean delivered. Adeyemi 2024 reports grades 5–9 together, and the protocol needs grades 6–8 reported separately. Lindgren 2025 measures anxiety and course grades only. The Adjudicator has not decided any of them; they are yours.
Which studies have a number that does not match?
The Checker agent re-derived every sample and effect size across 41 studies and found 7 mismatches. Six are resolved, for example Harlan 2022, where a standard error was entered as a standard deviation. One is open: Vellacott 2023 reports 212 students in the abstract, but Table 2 sums to 198 (101 + 97). The study stays out of the meta-analysis until a reviewer records the right value.
Draft the AI-use statement for our methods section
Drafted from the record: two screening agents on different models voted blind on all 4,981 records; people confirmed every agreed exclusion and decided all 112 conflicts; every extracted value was verified by a reviewer and every effect size re-derived. It is on the Report screen, with model and prompt versions.
What did this month's search add?
On October 1 the Search agent re-ran the 6 saved searches and found 86 new records. It removed 19 duplicates; the screeners voted on 67: 58 agreed exclusions wait for a person's batch check, 6 went to full text, and 3 are conflicts. One of the six, Vellacott 2023, came from the trials register.
13 screens from the working solution, on its sample data. Pick one to see it large.
The thresholds every vote and value is checked against are settings. A change needs a reason, creates a version and goes on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Screener confidence | 0.80 | Below it, a record goes to a person even if both agents agree |
| Auto-merge duplicates at | 0.97 similarity | Pairs from 0.85 to 0.97 go to the librarian |
| Checker tolerance | Counts exact; g within ±0.02 | Changed only with a reason, versioned |
| Batch sample read in full | 10% | Of each agreed-exclusion batch, before a person confirms it |
| Monthly search | 1st of the month, 6:00 a.m. | Saved strategies, sources and schedule per review |
| Service levels | Conflicts 5 business days | Full-text decisions 10; Checker flags 5; escalation 2 days past, to the head of service |
| Reviewer notes and votes | Retained 7 years after publication | Then deleted, with a record of the deletion |
| Agents learn from this review | Off | Decisions in a review never train a model |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
Sources: Borah et al., BMJ Open 2017 (67.3 weeks on average from registration to publication; a mean of 1,781 records screened by a mean team of five) · Journal of the Medical Library Association, librarian time on systematic review services (14.7 hours per review across 319 requests) · PRISMA 2020 statement (flow diagram and 27-item checklist) · RAISE recommendations on reporting AI use in evidence synthesis (2025) · "target" = design goal, measured in the live solution · "demo" = seen in the working solution, on its sample data · "typical" = published figures · "estimated" = our arithmetic from published inputs. Halden State University, its review and the people named on this page are fictional — characters and sample data in the working solution; study names and results are sample data.
Systematic Review & Evidence Synthesis is systematic review software for library evidence synthesis services and review teams. Agents capture the criteria from the registered protocol, re-run the saved searches monthly, vote blind on every title and abstract, extract data pinned to page and table, re-derive every effect size and draft the PRISMA 2020 flow and checklist. People decide every conflict, confirm every exclusion and value, and the principal investigator signs the flow before it exports.
Here, two Screener agents on different models vote on every title and abstract against the registered criteria, each quoting the abstract and citing the criterion, without seeing the other's vote. A vote below 0.80 confidence goes to a person, agreed exclusions count only after a person confirms the batch, and every disagreement goes to a second reviewer. The agents screen a review only after a 300-record calibration pilot.
The PRISMA 2020 flow diagram reports how many records a review identified, removed as duplicates, screened, excluded and included, with the reasons for full-text exclusions. In this solution the flow is the home screen and is drawn live from the record, so every box opens to its records and the counts balance from identification to inclusion.
A living systematic review is kept current as new studies appear. The Search agent re-runs the saved strategies on the 1st of every month, removes duplicates and drops the new records into screening, and the increment shows on the top box of the flow, such as "+86" from an October search.
The Report agent drafts an AI-use statement for the methods section from the record: which agent did what, on which model and prompt version, the calibration result and the human checks. The 2025 RAISE recommendations ask reviews to report any automated step that makes or suggests judgments, and how people oversaw it; the statement exports with the signed flow.
Yes. The Adjudicator never decides a conflict, the Checker never overwrites a value and the Bias agent never confirms a judgment. A second reviewer decides every conflict, every full-text exclusion needs a reason picked by one reviewer and confirmed by another, any person can override an agent with a reason code, and the principal investigator signs the flow before anything exports.
On average, 67.3 weeks from registration to publication. The estimate for this solution is 50–70% less screening and extraction time, while keeping two decisions on every record; it is measured in the live solution.
The Agentic Solution Engine builds and deploys it from your requirements and documents — a registered protocol and a past review's records and search log — and it goes live once every quality gate has passed. We will walk you through it on one of your own reviews first.
Backlogs described in weeks, not years, with every line approved by an archivist.
1-min demoResearch Integrity ScreeningDuplicated figures fixed before submission, not retracted after publication.
1-min demoResearch Conflict of Interest & Research Security ReviewUndisclosed ties found before just-in-time, and explained by the researcher first.
1-min demoWe'll run Systematic Review & Evidence Synthesis on a protocol and a search export from your own service.