How to Evaluate and Implement an Early-Warning Model for Legal Deadline Loss

The strongest signals usually come from the interaction between the publication and the process context. A notice that appears routine in isolation may become urgent when it follows a recent judgment, a request for evidence, a change in representation, or repeated procedural movements within a short period.

Useful publication-level features include the act type, court, chamber, jurisdiction, date and time of capture, date and time of publication, presence of an explicit deadline, names of parties and attorneys, and whether the text contains terms associated with appeal, response, payment, evidence, hearing, compliance, or procedural regularization.

Timing features are especially valuable. Calculate the interval between the movement timestamp and publication timestamp, the interval between publication and first human review, and the number of business days remaining after the item enters the feed. A short remaining window combined with no assigned owner is a high-value early-warning signal.

Process history adds another layer. Consider the number of movements in the previous 7, 14, and 30 days, the frequency of changes in the responsible attorney, the number of previous publications that required action, the existence of open tasks, and whether earlier alerts were ignored or completed late.

The model should also account for operational context. A high-volume retail portfolio with 5,000 monitored names may have recurring patterns by court, claim type, outsourced firm, region, or attorney. These patterns should be used carefully, because a historical association can reflect workload imbalance rather than legal complexity.

Data quality is more important than model sophistication. Before training anything, check duplicated publications, inconsistent process identifiers, missing timestamps, variations in court names, incorrect attorney associations, and records where the final outcome is unknown. The Brazilian General Data Protection Law, Law No. 13,709/2018 should guide the treatment of personal data, access controls, retention, and documented purposes.

A useful first version can be built from structured feed data without sending every publication to a large language model. Rules can identify explicit dates, procedural verbs, urgent act categories, and missing ownership. Only the smaller set of ambiguous or high-risk items needs deeper language analysis.

For example, a normalized table could contain one row per publication with fields such as process_id, publication_id, court, instance, published_at, captured_at, act_type, attorney_id, party_id, text, deadline_date, review_status, task_created_at, and task_completed_at. Historical process features can be calculated in a separate table and joined by process_id and publication date.

A simplified PostgreSQL query might look like this:

SELECT
 p.publication_id,
 p.process_id,
 p.court,
 p.act_type,
 EXTRACT(EPOCH FROM (p.published_at - p.movement_at)) / 3600 AS hours_movement_to_publication,
 COUNT(m.movement_id) FILTER (
 WHERE m.movement_at >= p.published_at - INTERVAL '30 days'
 AND m.movement_at < p.published_at
 ) AS movements_30d,
 COUNT(t.task_id) FILTER (
 WHERE t.status IN ('open', 'overdue')
 ) AS open_tasks,
 CASE
 WHEN p.deadline_date IS NOT NULL
 THEN p.deadline_date - CURRENT_DATE
 ELSE NULL
 END AS calendar_days_remaining
FROM publications p
LEFT JOIN movements m ON m.process_id = p.process_id
LEFT JOIN tasks t ON t.process_id = p.process_id
GROUP BY p.publication_id, p.process_id, p.court, p.act_type,
 p.published_at, p.movement_at, p.deadline_date; 

In production, replace CURRENT_DATE with the scoring timestamp and calculate business days using the team’s approved calendar. The query is illustrative, but the principle is important: every feature should be reproducible, timestamped, and explainable to the lawyer who receives the alert.

A Looker or similar BI model can expose the same features for quality checks. Track missing values by court, duplicate rates, median capture delay, publication volume by day, and the percentage of alerts with a linked task. These controls often reveal problems before model metrics do.

A streaming architecture scores a publication soon after it enters the feed. It is appropriate when the remaining response window is short, when the legal team needs immediate escalation, or when the process must be synchronized with an external task system in near real time.

Batch scoring runs on a schedule, such as every 15 minutes, hourly, or several times per day. It is simpler and less expensive, and it can be sufficient when publications arrive in predictable volumes and the team has enough lead time to act on the result.

Many legal operations benefit from a hybrid design. Use event-driven ingestion for new publications, apply inexpensive rules immediately, and send only ambiguous or high-risk items to a predictive model or language model. Recalculate portfolio-level features in batches and refresh daily dashboards on a fixed schedule.

This architecture also controls LLM costs. First filter the Pocket Jurídico feed using act type, explicit deadline indicators, court, process status, and recent history. Then invoke ChatGPT, Claude, Gemini, or another approved model only when the text requires summarization, deadline extraction, or contextual interpretation.

Store the model version, prompt or rule version, input identifiers, output, confidence, and reviewer decision. Do not allow an AI-generated date to become a final legal deadline without human validation and a clear source link to the publication.

For governance, define service levels before production. For example, a high-risk alert may require delivery within five minutes and review within 30 minutes, while a medium-risk item may be included in the next operational queue. The service level should be based on deadline exposure, not on technical ambition.

Consider a retail company monitoring 5,000 names across several outsourced law firms. The portfolio receives 18,000 publications per month, with 12% requiring an action, 2,160 actionable items, and an estimated baseline missed or materially late action rate of 1.8%. That represents approximately 39 high-consequence events per month, although the exact definition should be validated against the company’s records.

Suppose a pilot ruleset and predictive model achieve 88% recall for the selected risk class, 42% precision, and a median lead time gain of 1.4 business days. The system would generate approximately 1,800 alerts if applied to all actionable items, so the legal team should first limit alerts to the highest-risk 25% or route medium-risk items to a queue rather than an immediate escalation.

If the controlled workflow reduces the missed or materially late rate by 35%, the portfolio could avoid roughly 14 such events per month. The financial value depends on the cost of each event, but a conservative internal estimate might include outside counsel work, rework, escalation time, operational disruption, and expected exposure rather than assigning an unsupported legal-loss figure.

Assume the organization spends 12 minutes reviewing each high-risk alert and 3 minutes dismissing or triaging each lower-risk item. At 450 prioritized alerts per month, review effort is about 90 hours. If pre-filtering reduces language-model calls by 70%, infrastructure and inference costs can remain modest compared with the staff time and risk reduction created by the workflow.

This is a planning simulation, not a performance promise. The correct decision depends on the company’s baseline, the cost of review, the distribution of deadlines, and whether law firms actually create and complete tasks after receiving warnings. A pilot should replace these assumptions with measured results before a full rollout.

The decision threshold should be explicit. Choose a higher recall configuration when the consequence of missing an item is severe and the team can absorb review volume. Choose higher precision when alerts compete with urgent legal work, but add a separate safety queue for uncertain publications so that precision does not conceal blind spots.