Demos are not proof.

Any tool can look impressive on a curated demo. The question that matters is how it performs on a matter it has never seen — where the right answer is known and hidden. That is what CloseVector's validation program is built to measure.

A curated demo is overfit to easy facts.

A vendor demo is built around documents the vendor chose. It shows the tool succeeding on facts that were selected to make it succeed. That is a useful sales exercise and a poor measure of trust: it says nothing about recall on the oblique, buried, low-prevalence evidence that actually decides cases.

Public datasets do not close the gap. The Enron corpus is old, narrow, and has very likely been absorbed into model training, so a strong score on it proves little about how a system performs on a matter it has never seen.

A demo measures the demo. It does not measure your matter.

Sealed synthetic matters with hidden ground truth.

Corporate physics

A synthetic matter is a structured world — org chart, custodians, incentives, timeline, and document families — not a pile of fake emails. The facts behave the way they would inside a real organization.

Planted facts, not keywords

Evidence is planted as facts spread across documents, including oblique and coded language, not as a single target phrase a tool can pattern-match. The test rewards genuine retrieval, not lexical luck.

Hard negatives

Innocent look-alikes are planted alongside the real evidence to punish shallow pattern-matching. A system that grabs everything that looks relevant fails as clearly as one that misses the signal.

A sealed answer key

The ground truth is hashed, access-controlled, and never seen by the system in advance. The evaluation is one-shot, against an answer the system cannot study beforehand.

Measured against a known answer.

These are the measures the validation program is designed to produce against hidden ground truth — the same dimensions that govern the e-discovery workflow on the evidence engine. They are described here as what the program measures — not as scores, and not as a claim that the system has been validated.

Material evidence recall
Designed to measure how much of the planted material evidence the workflow surfaces.
Contradiction recall
Designed to measure how reliably the workflow surfaces evidence that contradicts a stated theory.
Privilege detection
Designed to measure how completely privilege candidates are flagged before any general report.
Citation fidelity
Designed to measure whether every asserted fact traces to a real, correct source document.
Unsupported-assertion rate
Designed to measure how often an output makes a claim no source supports.
Timeline reconstruction
Designed to measure how accurately the workflow reassembles the sequence of events.
Witness identification
Designed to measure how completely the key custodians and actors are surfaced.
Processing-exception accuracy
Designed to measure whether what could not be processed is reported, not hidden.

Bounded claims, or none.

CloseVector will publish a performance number only after a sealed, one-shot evaluation against hidden ground truth — stated together with the scenario family it was run on and the limits of what it shows. Not before.

Until that evaluation exists, what you see here is a validation program and a methodology, not a performance claim. A tool that is careful about what it will not yet claim is telling you something about how it will treat your matter.

This page is the proof of the discipline. The number comes after the sealed run, with its limits stated — or it does not come at all.

See how we measure — and what we won't claim until we can.

Bring a matter type. We will walk you through the validation methodology and exactly where its limits are.

Or reach the team directly: contact@closevector.ai