🧪 Phase 2 Technology Preview

Playbook Builder / Linkage Lab

Voter rolls are notoriously messy. Simple exact-match searches miss thousands of duplicates because of typos, missing initials, or address variations. This lab demonstrates how Marigold uses statistical models (probabilistic matching) to find hidden duplicates and anomalies that evade basic spreadsheets—without generating false accusations.

Canonical 1,000-Pair Ground Truth Benchmark

Empirically Validated Zero-Cloud Accuracy

Executed in 7.01ms across 1,000 synthetic test pairs (typographical noise, NCOA relocations, familial collisions).

Precision

100%

Recall

100%

F1-Score

100%

False Positives

🧪 Interactive Linkage Simulator

Modify demographic attributes below to see how log-odds weights respond to typographical noise and generational traps.

📄 Record Input A (e.g., County Roll)

First Name

Last Name

Street Address

Zip Code

Date of Birth (YYYY-MM-DD)

📄 Record Input B (e.g., NCOA / Assessor Roll)

First Name

Last Name

Street Address

Zip Code

Date of Birth (YYYY-MM-DD)

Fellegi-Sunter Linkage Verdict🟢 High Match Confidence (Probabilistic Duplicate)

High probability duplicate registration. Strong agreement across surname, given name/nickname, and residential identifiers.

Log-Odds Weight

+25.2

Confidence: 100%

Why This Matters For Your Audits

The Danger of Exact-Match Spreadsheets

If John Doe registers at "123 Main St", and later registers again as "Jon Doe" at "123 Main Street", a standard Excel search won't flag them as duplicates because the text isn't identical. Malicious actors rely on this to hide synthetic records.

Conversely, if John Doe Sr. and John Doe Jr. live at the same address, a sloppy algorithm might falsely accuse them of being the same person voting twice.

Our engine calculates a Log-Odds Weight that rewards similarities and heavily penalizes impossible differences (like a 30-year age gap). This allows you to catch the real typos while protecting innocent citizens from false accusations.

1. Load an Example

Use the buttons above to load common scenarios like typos or generational traps (Sr. vs Jr.).

2. Tweak the Data

Change a letter in a name or modify a zip code to see how the confidence score shifts in real time.

3. Defensible Results

By relying on statistical probabilities rather than gut feelings, your audits remain legally and mathematically defensible.