Pipeline & Workflow

ViceWire turns a raw article into a structured event payload through a multi-stage workflow. The pipeline does not jump straight from article text to a final label. It first determines which events are actually present, then enriches each confirmed event with timing, taxonomy, metadata, business-surface mapping, and directional layers.

​

The high-level flow

In simple terms, the workflow is:

​

1. Article cleaning and structuring

The pipeline starts by cleaning the raw source article and converting it into a structured markdown format. This step standardizes the document before classification and extraction. It gives later stages a consistent article shape rather than forcing every downstream step to work directly from unstructured source text. Most importantly, it reduces noise and improves downstream output quality.

​

2. Event-family identification and confirmation

ViceWire then determines which top-level L1 event families are actually present in the article. This happens in two stages:

1. Ranking: a first-pass ranking step narrows the label space to the most plausible event families.
2. Confirmation: a follow-on classification step determines which of those shortlisted families are actually supported by the article.

This confirmation stage gates the rest of the workflow. Later enrichment only runs for confirmed event families rather than every possible category.

​

3. Event enrichment

For each confirmed event family, the pipeline enriches the event with additional structured layers. That enrichment can include:

• timing and editorial_form
• event_sub_types and event_stage_statuses, where the taxonomy supports them
• event_family_metadata
• affected_business_surface

These layers are attached only after the event family has been confirmed, so later classification and metadata extraction run inside an event-specific context rather than across every possible category. For field definitions and payload semantics, see the Event Model, Affected Business Surface, and Directional Layers.

​

4. Directional layers and FinBERT

After the event has been identified and structured, the pipeline attaches directional layers. That includes:

• event-family-specific directional or impact fields inside event_family_metadata
• finbert, computed per confirmed event family from event-specific context

This ordering matters because ViceWire first identifies the event and only then attaches directional interpretation and sentiment within that event context. For field semantics and interpretation, see Directional Layers.

​

5. Final payload assembly

The final step assembles the external payload by combining article-level context with the event objects produced during the earlier stages. If the article confirms multiple L1 event families, the final payload includes multiple event objects under the same article-level container. For the external payload shape and field definitions, see Event Model.

​

6. Evaluation, escalation, and production push

Before the payload is pushed to production, it is run through an evaluation stage. That stage is designed to catch cases where the output may require additional review, including situations where the extracted structure may be incomplete or contain errors, or where the result falls outside the current limits of the system. If the payload passes evaluation, it can be pushed to production. If it does not, it can be escalated for human review. For more on quality checks, review logic, and current system limitations, see QA, Evaluations & Limitations.

​

Why the workflow is staged this way

The order of operations is deliberate. ViceWire does not start with sentiment and work backward. It first determines what happened, then where the event sits in time and taxonomy, then what structured facts belong to it, and only then attaches directional layers. This staging keeps interpretation anchored to the event itself rather than flattening the article into one broad document-level view. It allows timing, taxonomy, metadata, business-surface mapping, and sentiment to be attached within the context of a specific event family.