How Havast works
One study PDF, extracted twice: once by you, once by the model, neither
able to see the other's answers. The two records are then reconciled into a single consensus
record, and every disagreement is settled by the reviewer, not by the model.
Study PDF
Reviewer
extracts in Havast
Artificial intelligence
reads only the PDF
Independent
neither record is visible to the other
Study characteristics
Study design
Cluster RCT
Total randomized
1,204
Outcomes
All-cause mortality · 12 mo
HR 0.82
Incident T2DM · per 1 SD
Odds ratio 1.18
Study characteristics
Study design
Cluster RCT
Total randomized
1,204
Outcomes
All-cause mortality · 12 mo
HR 0.82
Incident T2DM · per 1 SD
Odds ratio 1.81
Reconciliation
rows matched on outcome · timepoint · contrast
agree
agree
agree
disagree — reviewer decides
Cluster RCT · Cluster RCT
1,204 · 1,204
HR 0.82 · HR 0.82
Odds ratio 1.18 · 1.81
Consensus record
Study design
Cluster RCT
Total randomized
1,204
All-cause mortality
HR 0.82
Incident T2DM
Odds ratio 1.18
Faster
A second reviewer, available
on demand.
Fewer discrepancies
Caught and resolved before
they reach the dataset.
More accurate
Dual extraction catches what
single extraction misses.
