Risk Inference Guide

Risk inference turns relevant, sourced text discovered during a search into explainable screening and Client Risk Rating signals. It complements direct watchlist and configured-list matches rather than replacing them. Access: Requires the Admin role or above. In the sidebar, go to Administration > Configuration, then open Risk Inference under Screening Behaviours. Risk Inference is workspace-scoped, so a Calibration workspace can be tuned without changing Live.

​

What risk inference does

Minerva combines structured records with public information found while a search runs. Open Source Intelligence (OSINT) is sourced public-web information. When Open Source is requested, Minerva retrieves that information in real time and carries sourced occupation, organization, notes, document titles, industry, employer, and other business context into the resolved profile. Risk inference evaluates the resolved profile context from all applicable sources. Open Source supplies most newly discovered live-web context, but it is not the only source of text. Minerva does not treat every web mention as fact: an inference finding is an explainable risk signal to review against the source, the identity evidence, and the surrounding context. Minerva derives risk from text through three separate signals:

• Risk inference (this page) evaluates the resolved profile context for PEP, Criminal, and High Risk Industry signals.
• Adverse media categories define the risk topics an article may carry once analysis has flagged it. For how articles are found and scored, see Adverse Media.
• Role-aware adverse media judges whether a flagged article appears to involve the screened subject or only mentions a similar name.

The three inference types have different outcomes: Direct PEP or Criminal watchlist hits and configured-list results remain separate evidence. Disabling text inference does not disable those sources. When matched, PEP, Criminal, and High Risk Industry can each activate a dynamic Client Risk Rating factor with a default weight of 0.25, described in the CRR model as a 25-point factor. The final normalized score does not always rise by exactly 25 because other factors, normalization, and overriding criteria can affect the result.

​

Which engine to use

Each channel has its own PEP, Criminal, and High Risk Industry settings. The workspace UI calls them strategies; this guide uses engine. The available engines are:

​

When Decision Classifier runs

The classifier is deliberately limited to searches where it has source-backed context to evaluate. It does not run on every search: This limit keeps the classifier focused on searches where Minerva has live, source-backed context to evaluate. A PEP-only or Criminal-only search is still useful for list screening; it is simply not a classifier workload.

​

How context-aware classification improves on keywords

A keyword engine decides from words. A name, brand, or ordinary phrase can look exactly like a risk term even when the surrounding context says otherwise. The Decision Classifier reads the complete potential-match context, including occupation, organization, industry, notes, source lineage, and other supported fields together, and classifies the subject against the labels the workspace configures. That difference is most visible in four places:

• Name and brand traps. A company called “The Senator Group” is not a senator, and a music company called “Ministry of Sound” is not a government ministry. The classifier can consider what the entity actually is rather than matching the word alone.
• Paraphrase and translation. A public role described in another language, or in narrative notes rather than a title field, can be understood without requiring an exact English phrase.
• Long and mixed profiles. The classifier can weigh a current role against stale, pruned, or conflicting evidence instead of treating every matching token as equal.
• Context-dependent exclusions. A regulated bank or insurer, a privatized company, or an honorary title can be handled through label descriptions and instructions rather than an ever-growing list of suppressions.

The implementation is intentionally abstracted for administrators and analysts: Minerva uses a large language model to make a short, constrained classification from evidence Minerva already retrieved. The model does not browse the web or create new facts at classification time, and the analyst-facing result is presented as Risk Inference with the selected category or tier and a profile-context rationale.

​

POC results

In Minerva’s internal POC, 600 multilingual profiles covering PEP and High Risk Industry (HRB) were built from real public records and deliberately included short, long, paraphrased, translated, and lexical-trap cases. The same corpus was scored with Keywords and with the Decision Classifier: Accuracy is the share of profiles with the exact expected label. Macro-F1 balances precision and recall across all labels. Negative FP rate is the share of expected-negative records incorrectly marked as risky: this was 1 of 60 PEP negatives and 0 of 60 HRB negatives. The single PEP false positive was a privatized state-owned enterprise landing at tier 3. The POC ran with the earlier rubric wording. That finding led to the current default wording, which requires current state ownership or control for a role held now and excludes a company merely because it was formerly state-owned. The POC is a synthetic benchmark built from public facts, not a production accuracy guarantee; validate changes against your own representative cases.

​

Configure Decision Classifier

1. Open Administration > Configuration > Risk Inference.
2. Select the workspace and the screening channel: Onboarding, Ongoing Monitoring, Direct API Calls, or Risk Assessments.
3. Choose Decision Classifier for the PEP, Criminal, or High Risk Industry inference you want to configure.
4. Edit the Agent prompt, the Review prompt, and the Risk labels for that inference.
5. Select Review changes, add a clear change description, and save.
6. Test the changed channel in a Calibration workspace before applying it to Live.

When a channel has no stored classifier policy yet, Minerva shows a warning. Saving applies Minerva’s default policy for that risk type; the engine stays selected, and the policy can be edited after the first save. Figure: the review step when a channel first adopts the Decision Classifier. Figure: the Decision Classifier editor with prompts and PEP risk labels.

​

Prompts

Each risk type has two prompts. They serve different purposes and are evaluated together: The two prompts share an 8,000-character limit across both fields. Concise, specific instructions work better than a long list of examples or duplicated label text.

​

Write a good agent prompt

A strong agent prompt answers four questions:

1. What is being decided? Name the risk type and the decision the classifier must make.
2. What counts as evidence? Require sourced profile evidence and source lineage; say that a name alone, an association, or an unproven allegation is not enough.
3. How should conflicting evidence be handled? Prefer current, trusted evidence; treat pruned or historical evidence as supporting context, not current fact.
4. When should the absent label be selected? Say that insufficient evidence should resolve to the absence label rather than a speculative positive classification.

For example, a PEP agent prompt can establish that the classifier must choose the most specific qualifying tier and must never upgrade a senator, provincial legislator, or state-owned-enterprise executive to a higher tier. That instruction belongs in the prompt because it is a cross-label precedence rule; the individual tier boundaries belong in the label descriptions.

​

Write a good review prompt

Use the review prompt for instructions that are specific to your workspace and not already captured by the labels. Examples of appropriate review-prompt content include:

• How to treat a particular regulatory or licensing status
• Whether an honorary or ceremonial title should be treated as insufficient for a PEP tier
• A local definition of criminal activity or a high-risk business category
• Language-specific context that should not be interpreted literally
• A reminder to distinguish a charge or allegation from a conviction

Do not use the review prompt to repeat the whole label list, to contradict the agent prompt, or to add facts that are not in the profile. If a rule changes the meaning of a category, update the label description as well; the label is the definition the classifier chooses from, and the prompts should not disagree with it.

​

Risk labels and categories

Risk labels are the categories the classifier chooses from. Each label has a name, a description, a Marks this risk as absent setting, and, for positive PEP labels, a PEP tier.

​

Rules the editor enforces

• A policy must contain at least one label that identifies the risk as present.
• A PEP policy must contain at least one label that marks the risk as absent. Without an absence label, every profile would be forced into a tier.
• Every positive PEP label needs a tier from 1 to 4.
• A label with Marks this risk as absent cannot have a PEP tier.
• A policy can contain up to 40 risk labels.
• The agent prompt and review prompt together can contain up to 8,000 characters.

Figure: validation messages beside an incomplete classifier policy.

​

When to mark risk as absent

Use an absence label when the sourced evidence does not support the risk. It is a real classification outcome, not a “skip” or “unknown” bucket:

• Not a PEP means the profile does not show a qualifying public role. A name-only match, a non-public honorary title, or a title that does not independently satisfy a tier definition is not positive evidence. A current private-sector role does not erase a prior qualifying public office; PEP status persists after leaving office, so use Not a PEP only when no qualifying present or past office is supported within the applicable lookback period in the PEP Policy.
• No criminal risk means no documented conviction, charge, sanction, or official finding supports a criminal category. It does not mean the person has never been mentioned in an article.
• No high-risk business, if your High Risk Industry policy adds that absence label, means the sourced business or role does not fall into any configured high-risk category. It does not mean the company is low risk for every other reason. The default High Risk Industry labels are category-only; if none applies, no High Risk Industry finding is produced.

If a role is recognized by the PEP Policy but is not covered by the default labels, add a custom label instead of forcing it into Not a PEP. Do not classify a relationship the PEP Policy places at tier 1, such as the spouse or partner of a head of state or government, as Not a PEP. The tier records relationship-based exposure, not an office held by that person. Absence labels are also how a classifier can clear a result. When the evidence is insufficient to support a positive category, the classifier should choose the absence label rather than guess. That is why the default PEP prompt explicitly tells the classifier to choose Not a PEP when the evidence is insufficient. Do not use absence labels to hide evidence that actually meets a positive definition. If the profile supports a qualifying role or a documented criminal finding, choose the positive label even if the result would be inconvenient. If the evidence is genuinely ambiguous, resolve the ambiguity through the label description and prompt; do not create an absence label whose description is broader than the policy.

​

Default PEP labels

​

Default Criminal labels

Criminal inference is multi-label: the classifier can select more than one category when the documented facts support more than one. It should distinguish a conviction from a charge or an allegation and should not classify a person solely because they are associated with someone else.

​

Default High Risk Industry labels

High Risk Industry is multi-label. The default policy uses category labels only; if no category applies, the result has no High Risk Industry finding. If your policy needs an explicit absence outcome, add a No high-risk business label and mark it as absent. A regulated bank or insurer is not high-risk business on that fact alone; classify the actual business, product, or role, not a word in the company name.

​

Choosing between Keywords and Decision Classifier

Keywords remain valuable when the policy is narrow, stable, and easy to express as explicit terms. The Decision Classifier is the better default when the evidence is contextual or the keyword policy is producing recurring false positives. A practical pattern is to use the Decision Classifier for the ambiguity-heavy channels and retain Keywords for a narrow, stable policy where explicit term evidence is required. Do not run both against the same risk type in the same channel expecting two independent findings; choose one engine per risk type and channel.

​

Keyword tuning

If you choose Keywords, the Matching Terms editor is designed for compliance users: it stores structured, plain-language rules instead of regular expressions. For each category, you can:

• Review the active Minerva built-in terms
• Add a local term or phrase
• Suppress a built-in term or add a false-positive suppression
• Choose whether any or all added terms must match
• Choose how an added term is recognized
• Restore the Minerva defaults

A suppression is most useful when the same harmless phrase repeatedly causes a known false positive. For example, an admin might suppress student senator in PEP tier 2 or charged with overseeing in Criminal inference after confirming that those phrases produce false positives. These figures show Keywords configurations tuned from evidence. The Matching Terms editor is available only when the channel uses Keywords. Figure: a PEP tier with a local term and a false-positive suppression. Figure: a Criminal category with a local term and a false-positive suppression. Figure: a High Risk Industry category with a local term.

​

Tune from evidence

Tune from a representative pattern rather than from one unusual result. A safe calibration sequence is:

1. Collect a representative set of false positives and known true positives, including different languages and profile lengths where relevant.
2. Identify the exact inference type, channel, label or term, and source field responsible.
3. Make the smallest label, prompt, term, or suppression change that addresses the pattern.
4. Test the changed channel in a Calibration workspace and compare result volume, true-positive retention, and CRR changes against the baseline.
5. Review the change summary and save a clear reason.
6. Switch to the Live workspace, deliberately reapply and save the reviewed settings, then monitor the affected workflow.

​

Review, history, and rollback

Risk Inference changes use the same workspace deployment controls as other tenant configuration sections. Before saving, Review changes shows:

• Each affected channel and inference engine
• Added, changed, or removed labels, terms, and suppressions
• Prompt changes
• Category and match-rule changes
• An optional change description

The history page records who changed the configuration, when it changed, the affected sections, and the saved description. Rollback restores a prior snapshot by creating a new history entry; it does not erase the audit trail. Figure: saved entries in the configuration history.

​

Related guides

• PEP Policy
• Data Feeds
• Risk Rating
• Screening Guide
• Match Scoring Guide
• Workspaces Guide