EvidenceGrid
Production acceptance test · July 2026

Quote and page verification, measured on a fixed PDF set.

EvidenceGrid was tested on 11 research PDFs and 88 extracted fields in the production workflow. The result below covers upload, extraction, source matching, repair, and conservative Synthesis output.

11 / 11supported PDFs extracted
98.86%quote and page accuracy after repair
4 / 4repair attempts succeeded
0severe fabricated synthesis claims
Acceptance result

Every threshold passed on this test set.

Rates describe one fixed production acceptance run. They should be read with the scope and limitations below.

MetricResultRateTarget
Supported PDF uploads11 / 11100%95% or higher
Extraction jobs completed11 / 11100%95% or higher
Quote location after repair87 / 8898.86%95% or higher
Claimed page accuracy after repair87 / 8898.86%95% or higher
Repair success4 / 4100%90% or higher
Strictly supported synthesis statements3 / 3100%90% or higher
Method

A source-first acceptance sequence.

  1. 01

    Upload representative PDFs

    The set included files up to 15.54 MiB and one scanned document within the declared 50 MiB limit.

  2. 02

    Extract and independently locate evidence

    Each generated quote and claimed page was checked against independently extracted PDF text.

  3. 03

    Repair uncertain fields

    Four failed initial matches were reprocessed at field level and then checked again against the source.

  4. 04

    Audit retained synthesis statements

    Three retained items were manually checked against every supplied quote and page. All three were strictly supported.

What was checked

The extracted value and its source stay together.

The acceptance test evaluated continuous quote location and claimed page accuracy. The product exposes the same quote, page, match state, and Open page action to the researcher.

One verified evidence cell
Quote matched
Scope

What this benchmark does not establish.

Run your own evidence-linked review.

Start with 10 papers free