Match Scoring Guide

Match scoring controls how broadly or narrowly Minerva keeps potential screening matches as they move through candidate retrieval, entity resolution, and final result filtering. Access: Requires the Admin role or above. In the sidebar, go to Administration > Configuration, then open Match scoring under Screening behaviours. Use this guide when you need to:

• understand what the Match scoring page controls
• choose between the built-in scoring presets
• tune pre-resolution and post-resolution score thresholds by workflow
• understand how criteria match scores and score labels should be interpreted
• decide how strict News retrieval and article analysis should be before adverse-media data reaches the Minerva engine
• tune News geography bias, name filtering, article volume, and location inference by screening workflow
• understand how role-aware adverse media fits beside match scoring
• review, audit, or roll back match scoring changes

​

How Match Scoring Works

Every screening result passes through three broad stages:

1. Candidate retrieval: Minerva gathers source records from the selected feeds.
2. Pre-resolution scoring: each source candidate is scored against the searched subject before entity resolution.
3. Post-resolution scoring: Minerva clusters likely duplicate source records, scores the resolved candidate, and filters the final returned result set.

The Match Scoring configuration controls the second and third stages through:

• pre-resolution thresholds for each feed: Sanctions, PEP, and News
• one post-resolution threshold for the workflow
• News geography bias, which controls how subject geography is used during News article retrieval
• News name filter, which controls how strictly article evidence must match the subject name
• Max requested articles, which caps the News article candidates requested for downstream analysis
• News location inference, which can infer article-linked subject locations before News geography scoring

The configuration is workflow-specific. Minerva currently separates:

​

Understanding Scores

Minerva uses normalized match scores from 0.00 to 1.00.

• 1.00 means the compared values are effectively exact matches.
• Values closer to 1.00 are stronger matches.
• Values closer to 0.00 are weaker matches.
• A threshold is the minimum score required to keep the candidate at that stage.

In result details and API responses, field-level criteria match labels are interpreted as: The overall candidate score is not a simple average of the visible field labels. The engine scores the candidate using weighted evidence such as name, aliases, date of birth, location, identifiers, and feed-specific evidence. Missing or unreliable fields can therefore change how much a candidate depends on name and other available signals.

​

Pre-Resolution vs Post-Resolution Thresholds

Pre-resolution and post-resolution thresholds answer different questions.

​

Pre-Resolution Thresholds

Pre-resolution thresholds decide whether a raw feed candidate is strong enough to continue into entity resolution. Each workflow has separate pre-resolution thresholds for:

Raise a pre-resolution threshold when a feed is sending too many weak candidates into the pipeline. Lower it when analysts are missing plausible source records before entity resolution has a chance to group and enrich them. Example: raising the News pre-resolution threshold for Ongoing monitoring makes Minerva reject weaker adverse-media candidate records earlier, before they can be clustered with other source records.

​

Post-Resolution Threshold

The post-resolution threshold is the final score a resolved candidate must meet before it remains in the returned result set. Post-resolution scoring happens after Minerva has clustered likely duplicate or related source records for the same candidate. This final score is what controls the analyst-visible result list after entity resolution. Raise a post-resolution threshold when resolved matches are still too broad after clustering. Lower it when entity resolution is correctly grouping evidence, but the final result list is hiding candidates your team wants to review.

​

Built-In Presets

Minerva includes three built-in presets. Built-in presets stay fixed, and workspace presets can be created from a custom draft.

​

Balanced

Balanced preserves the legacy scoring defaults and strict News retrieval behavior. Use Balanced when you want Minerva’s standard screening posture and do not have evidence that a specific workflow needs broader or narrower matching. Balanced uses strict News geography bias, strict News name filtering, a retrieval cap of 300 articles, and News location inference off.

​

Narrow

Narrow raises thresholds and keeps strict News retrieval controls. Use Narrow when review volume is too high and your program can tolerate less borderline recall in exchange for fewer weaker matches. Narrow uses strict News geography bias, strict News name filtering, a retrieval cap of 200 articles, and News location inference off.

​

Wide

Wide lowers thresholds and broadens News retrieval. Use Wide for calibration reviews, higher-risk cohorts, or workflows where the cost of missing a plausible match is higher than the cost of reviewing additional candidates. Wide uses broad News geography bias, broad News name filtering, a retrieval cap of 600 articles, and News location inference off.

​

News Controls For Adverse Media

Each screening workflow has its own News controls. This lets you keep onboarding or monitoring conservative while making direct API or risk-assessment searches broader when a specific process needs more adverse-media recall. These settings sit before final match scoring. They change which articles are retrieved, which retrieved articles are admitted for adverse-media analysis, and whether article-linked locations are added before geography scoring. The downstream adverse-media model and match scoring can only evaluate the article candidates that survive this stage.

​

News Geography Bias

News geography bias controls how subject geography is used during upstream News article retrieval. Strict is the Balanced and Narrow default. Wide uses Broad. Use Strict when the searched subject has reliable city or state evidence and common-name noise is a concern. Use Broad when relevant articles often mention only a wider region or country, or when Strict appears to miss plausible adverse media. Use None for targeted investigations where location evidence is weak or misleading and you are prepared to review a larger candidate set.

​

News Name Filter

News name filter controls how strictly retrieved article evidence must match the searched subject name before the article reaches adverse-media scoring. Broad name filtering can find articles that Strict misses, but it can also admit more same-name or near-name articles. It is most useful when paired with representative export analysis or analyst review examples.

​

Max Requested Articles

Max requested articles caps how many News article candidates Minerva requests and reviews for article matching, sentiment scoring, and adverse-media risk classification. Supported values are 100 to 750 articles. Balanced uses 300, Narrow uses 200, and Wide uses 600. Raise this value when a subject has many neutral or unrelated articles and relevant adverse-media articles may appear later in the provider result set. Lower it when common-name searches are creating too much latency or when a workflow does not need broad article recall.

​

News Location Inference

News location inference reviews matching risk-classified News articles and extracts locations that appear tied to the screened subject. Those inferred article locations can then provide more geographic evidence before News match scoring. Location inference is off in all built-in presets. Turn it on when article text often identifies the subject’s relevant city, state, region, country, operations, arrest, investigation, or other subject-linked geography more clearly than the original search request. It can help distinguish broad same-name article pools from articles tied to the subject’s actual geography. Location inference is bounded and runs only after article retrieval and local adverse-media classification. Neutral articles are not enriched. The latency impact is therefore highest for searches that produce many risk-classified articles.

​

How The News Controls Combine

The controls are applied in sequence:

1. News geography bias and News name filter shape the provider query and local article-admission checks.
2. Max requested articles limits how many candidates can be retrieved and reviewed.
3. The adverse-media model classifies the retained articles for sentiment and risk.
4. News location inference, when enabled, enriches risk-classified articles with subject-linked locations.
5. News pre-resolution and post-resolution thresholds decide which candidates remain visible after scoring and entity resolution.

For conservative production monitoring, prefer changing one control at a time and measuring alert volume before broadening another control. For one-off risk assessments, broader retrieval and location inference can be appropriate when the analyst expects to inspect the evidence directly.

​

Role-Aware Adverse Media

Role-aware adverse media is a separate screening configuration that runs after adverse-media articles have been retrieved and analyzed. It helps identify whether a negative article appears to involve the screened subject or is likely only a name mention. Use Match Scoring when you need to tune candidate thresholds or News retrieval and article-analysis behavior. Use Role-Aware Adverse Media when you need to show or filter confident article-subject relevance signals after adverse-media analysis has already run. For details, see the Role-Aware Adverse Media Guide.

​

Request-Level Overrides

Match scoring is the workspace default. It applies when the screening request does not provide an explicit override. Core screening logic gives request-level values precedence over workspace configuration. In practice:

• a feed-level match_threshold overrides feed-stage thresholds for that feed
• a feed-level pre_resolution_match_threshold overrides the workspace pre-resolution threshold for that feed
• a global match_threshold acts as a legacy override and takes precedence over post-resolution threshold defaults
• a global post_resolution_match_threshold overrides the workspace post-resolution threshold
• global or News feed-level news_geography_bias overrides workspace News geography bias
• global or News feed-level news_name_filter overrides workspace News name filtering
• global or News feed-level max_requested_articles overrides the workspace article cap for News
• global or News feed-level news_location_inference overrides workspace News location inference

This is important for API users. A workspace can keep a stable default while a specific API integration or investigation flow requests a narrower or wider threshold, broader News retrieval, or location inference for a single submission.

​

Recommended Tuning Approach

Use this sequence when calibrating match scoring:

1. Start from Balanced and collect examples of false positives, plausible missed matches, and high-volume review cohorts.
2. Tune one workflow at a time. Onboarding, monitoring, API, and risk assessment often have different tolerance for review volume.
3. Tune one feed at a time for pre-resolution thresholds. Sanctions, PEP, and News have different data density and false-positive patterns.
4. Adjust pre-resolution thresholds before post-resolution thresholds when the issue is feed-specific noise.
5. Adjust post-resolution thresholds when entity resolution is grouping candidates correctly but the final result set is too broad or too narrow.
6. Tune one adverse-media retrieval control at a time. Moving geography bias from Strict to Broad or None, turning on Broad name filtering, or raising max requested articles changes the upstream article pool, not only the final score filter.
7. Enable News location inference only where the extra subject-linked geography is worth the added latency.
8. Save a clear change description so history shows why the calibration was deployed.
9. Use history and rollback if the change moves review volume, latency, or recall in the wrong direction.

General tuning guidance:

• higher thresholds are stricter and usually reduce analyst-visible matches
• lower thresholds are broader and usually increase recall and review volume
• lowering News thresholds, setting geography bias to Broad or None, setting name filtering to Broad, and raising max requested articles at the same time can create a large increase in adverse-media candidates
• higher article caps and News location inference can increase News search latency, with the largest impact on high-volume or risk-heavy article sets
• monitoring changes should be made carefully because they can change recurring alert volume across an existing profile population
• direct API changes should be coordinated with API consumers, especially if those consumers already submit request-level overrides

​

Reviewing And Confirming Changes

When you click Review changes, Minerva shows a grouped confirmation view before anything is saved. The review dialog shows:

• the number of changes and sections affected
• the selected preset change, if applicable
• each workflow with changed pre-resolution thresholds
• the changed post-resolution threshold
• changed News geography bias, News name filter, max requested articles, and News location inference values
• an optional Change Description field

Use the change description to capture the reason for the calibration, such as a false-positive review, a model calibration review, a high-risk cohort exception, or a post-launch monitoring adjustment.

​

Deployment History And Rollback

The Match Scoring page includes a quick History drawer so you can preview recent deployments without leaving the configuration page. Use the drawer to:

• see the current live deployment first
• review prior threshold and News retrieval changes
• preview an older deployment on the main page
• open the full audit history page

Rollback restores a previous deployment by writing a new history entry. It does not delete prior history. The full history page includes:

• Changed: when the configuration was saved or rolled back
• Action: whether the event was an update or rollback
• Summary: the key threshold or News retrieval changes
• Changed by: the user who performed the change
• Actions: rollback entry points for older deployments

​

Related Guides

• Screening Guide
• Repeated Alert Suppression Guide
• Role-Aware Adverse Media Guide
• Workspaces Guide
• Bulk Actions Guide
• Risk Assessment Flow