Discover · TAR/CAL

The statistics are the feature

Continuous active learning with a held-out control set, Wilson confidence intervals, a stopping rule on the pessimistic end of the interval, and an elusion test on the discard pile — a validation protocol you can paste into a declaration.

Where it starts

Review is the expensive part of discovery, and cutting it is only defensible if you can state — with confidence intervals — what the review you did not do would have found. Settle what 'responsive' means first; the protocol does the rest.

How it runs

01
Draw the control set
A simple random sample, reviewed once and never trained on. It is the unbiased estimate of prevalence — the console reports it as a Wilson interval: 'prevalence 2.31% [1.94–2.75%] → ~15,940 responsive in corpus'.
02
Seed randomly, then let CAL rank
The seed round is random so the model doesn't inherit a query's bias. Then batches of 200, refit every round, ranked relevance-first — the model keeps handing you the most-likely-responsive documents next.
03
Watch the recall estimate, not the document count
Each round reports recall and precision with 95% intervals and a control-set AUC computed only on held-out documents — an honest generalisation signal, drawn as a confidence band against the target line.
04
Stop on the lower bound
The run stops when the 95% lower bound on recall clears the target — conservative on purpose. A point estimate that touches 80% is not the same claim as an interval that sits above it.
05
Test what you are not reviewing
The elusion test samples the discard pile — documents neither reviewed nor in the control set — and reports the responsive rate with its upper bound. The stop decision is corroborated against the documents nobody opened.

Why it holds up

The control set is never trained on — ever. Prevalence and recall come from documents the model has not seen.

Wilson score intervals throughout, because the normal approximation misbehaves exactly where review decisions live: small samples, low prevalence.

Every round is recorded: reviewed, found, batch precision, recall with CI, control AUC — a round-by-round audit trail.

The methodology is written in prose as a disclosable validation protocol, fit to paste into a meet-and-confer letter or a declaration.

The help panel warns against the tool's own misuse: inconsistent coding produces a confident model that is wrong.

What you hand the regulator

A workflow that ends on a screen isn’t finished. This one ends in a document.

The validation protocol

The TAR result record — persisted, ledgered, and written for the person who has to defend the review, not the person who ran it.

prevalence with 95% CI · estimated responsive range
final recall and precision with intervals · review avoided %
elusion sample: size, responsive count, rate, upper bound
round-by-round table with control-set AUC
the stopping criterion, stated, and whether it was met
methodology paragraph in disclosable prose

Adjacent workflows

Early Case Assessment Search & Term Hit Reports Privilege & Redaction

See it run on your data scenario

The demo form asks which workflows you want to see — name this one and we’ll stage it.

Book a Demo All workflows

How it runs

Draw the control set

A simple random sample, reviewed once and never trained on. It is the unbiased estimate of prevalence — the console reports it as a Wilson interval: 'prevalence 2.31% [1.94–2.75%] → ~15,940 responsive in corpus'.

Seed randomly, then let CAL rank

The seed round is random so the model doesn't inherit a query's bias. Then batches of 200, refit every round, ranked relevance-first — the model keeps handing you the most-likely-responsive documents next.

Watch the recall estimate, not the document count

Each round reports recall and precision with 95% intervals and a control-set AUC computed only on held-out documents — an honest generalisation signal, drawn as a confidence band against the target line.

Stop on the lower bound

The run stops when the 95% lower bound on recall clears the target — conservative on purpose. A point estimate that touches 80% is not the same claim as an interval that sits above it.

Test what you are not reviewing

The elusion test samples the discard pile — documents neither reviewed nor in the control set — and reports the responsive rate with its upper bound. The stop decision is corroborated against the documents nobody opened.