PDIS / Technical paper
PDIS.
Engineering research / System study
2026 · Project paper
Palantir Foundry AIP · Multi-agent systems · Legal technology

Public Defender
Intelligence System

Case analysis, precedent retrieval, and risk assessment using an ontology-backed agent system on Palantir Foundry AIP.

ProjectPublic Defender Intelligence System (PDIS)
PlatformPalantir Foundry AIP

Abstract

Public defenders need to pull together case records, charges, evidence, prior decisions, and possible outcomes. That work often involves switching between separate sources. PDIS brings these tasks into one Foundry application. It uses a six-object ontology to connect case data and six agents to handle lookup, charge review, evidence synthesis, precedent search, risk scoring, and outcome estimates. The project site reports 93% grounding accuracy on a 15-case evaluation and a 2.1-second median response time across 100 queries in a staging environment. This paper explains the design, reported results, and limits of using automated predictions in legal work.

Ontology modelingAgent orchestrationHybrid retrievalHuman review

Overview

PDIS is a case-analysis workspace for public defenders. Its main idea is to make related case information available through a common data model, then let focused agents work with that information.

The application has five workspaces: Case Lookup, Charges, Evidence, Precedents, and Risk/Outcome. A separate chat interface lets users ask questions in natural language. The system routes those requests to the relevant agents and passes an explicit case identifier so results stay tied to the selected case.

What it is designed to do. Reduce the time spent finding connected records, surface evidence gaps, retrieve potentially relevant precedent, and explain the factors behind a risk estimate. The system should support an attorney's review, not replace it.

System architecture

The ontology acts as the shared structure for all agent workflows. It defines six object types and links records through case and client identifiers.

Six-object ontologyConceptual data model
CaseCase ID, filing date, jurisdiction, narrative, client link, narrative embedding
ClientDefendant details, prior record, conviction count, case link
ChargeStatute, severity, statute text, filing date, case link
EvidenceEvidence type, admissibility status, chain of custody, case link
PrecedentCitation, court, synopsis, tags, outcome, synopsis embedding
OutcomeVerdict, sentence type, sentence duration, case link

The site describes these object types in Foundry Ontology Manager. The diagram summarizes their fields rather than reproducing a verified schema export.

How a request moves through the system

Input
Workshop or chatUser selects a case or asks a question
Context
Ontology and case lookupLinked objects are retrieved by case ID
Analysis
Specialized agentsEvidence, precedent, risk, and outcome steps
↓   Results returned with case context and supporting data
Review in WorkshopAttorney inspects the records, reasoning inputs, and flags before deciding what action to take

The site describes agents as AIP Actions that chain object queries, transforms, filters, aggregates, and joins. It also describes event-based communication and logged input/output for audit review. These are architectural claims from the project page, not independently inspected Foundry workflows.

Six agent workflows

Each agent has a specific responsibility. Keeping these tasks separate makes it easier to understand which data informed a result and where an error may have occurred.

Case Lookup

Joins case, client, charge, evidence, and outcome records. Supplies the shared case context to downstream agents.

Charge Assessment

Reviews statute information, charge severity, prior convictions, and possible enhancements or sentencing ranges.

Evidence Synthesis

Groups evidence by type, checks recorded chain-of-custody fields, and flags possible admissibility issues for review.

Precedent Discovery

Combines structured court and tag filters with semantic similarity over case-law summaries.

Risk Scoring

Combines four documented risk factors into a score and links those factors back to source records.

Outcome Prediction

Finds similar past cases, summarizes verdicts and sentences, and highlights mismatches for human review.

Coordination example

For an outcome question, the site describes a workflow that first retrieves the case, then calls precedent discovery and risk scoring before returning an aggregate result. The chatbot passes the case ID to each agent rather than relying on an agent to infer the active case. That reduces the chance of mixing results from different matters.

The documentation also describes a 60-minute cache for case lookup and per-agent least-recently-used caches. Caching may reduce repeated data access, but stale records are a concern when evidence or charge information changes.

Precedent discovery and hybrid retrieval

Legal research needs two kinds of matching. Structured filters can narrow results by statute, court, or applicability tags. Semantic search can find similar fact patterns even when the language differs. PDIS combines both.

Reported precedent-ranking design
Structured filterUse tags, statutes, and court level to narrow candidates
Semantic searchCompare case narrative and precedent synopsis embeddings
Rank resultsCombine tag relevance and similarity; return cited records

The site specifies a 0.75 similarity threshold, relaxed to 0.65 when fewer than three results are found. It describes a score weighted 40% toward tag matches and 60% toward vector similarity, with an additional boost for matching tags.

Illustrative ranking rule: score = 0.40 × tag_match + 0.60 × semantic_similarity

The source does not specify the normalization of the tag-match term or exactly how the separate boost is applied. Those details matter because otherwise the numerical score cannot be reproduced from the description alone.

Query design

The documentation provides SQL-like examples for structured case search, vector similarity, and combined filtering. The example below illustrates the intended pattern, not a tested Foundry SQL dialect.

Illustrative hybrid query
SELECT p.citation, p.court, p.synopsis
FROM Precedent AS p
WHERE p.court = :target_court
  AND similarity(p.synopsis_embedding, :case_vector) > 0.75
ORDER BY similarity(p.synopsis_embedding, :case_vector) DESC
LIMIT 10;

Pseudocode for explaining the retrieval approach. Actual query syntax depends on the Foundry environment and configured search operations.

Embeddings are described as 1,536-dimensional vectors created at ingestion through an OpenAI connector. Precomputing them avoids repeated embedding work for unchanged records, though case updates need a process to refresh the stored vectors.

Risk scoring and outcome estimates

The Risk Scoring Agent assigns a numeric score from zero to 100 using prior record, charge severity, evidence quality, and historical outcomes. The project page gives the following weights:

Prior record
30%
Charge severity
30%
Evidence quality
20%
Historical outcome
20%

Weights reported in the project documentation. Bars show the contribution weights, not measured risk or model accuracy.

Risk = 0.30 × prior_record + 0.30 × charge_severity + 0.20 × evidence_quality + 0.20 × historical_outcome

To interpret this formula as a 0–100 score, the individual components must use compatible scales. The site does not provide the feature normalization rules, so the formula communicates the intended weighting rather than a fully reproducible scoring procedure.

Outcome estimation

The Outcome Prediction Agent finds five similar cases and reports the proportion with a guilty verdict. It also derives a sentence estimate from the neighboring cases. This is a summary of selected historical records, not necessarily a calibrated probability for the current defendant.

Decision-support boundary. A similarity-based estimate should never be treated as a recommendation to accept a plea, set bail, or choose trial strategy. Legal context, jurisdiction, selection bias, missing evidence, and client circumstances need attorney review.

Performance and optimization

The project page reports a reduction in median end-to-end response time from 4.2 seconds to 2.1 seconds. It attributes the change to precomputed embeddings, vectorized similarity operations, caching, and parallel work where agent tasks do not depend on one another.

Reported latency optimizationLower is better
Baseline
4.2s
Embeddings
3.1s
Vectorization
2.5s
Caching + parallel
2.1s

Project-reported timings. The combined effect of caching and parallel calls is not separately isolated by the published breakdown.

50%Reported latency reduction
2.1sMedian across 100 queries
3.2sP95 query latency

The site also reports a 4.1-second P99 latency and a cache with up to 1,000 entries per agent. It does not supply timing traces, hardware specifications, cache-hit breakdowns, or a repeatable load-test harness. These numbers should therefore be read as project-reported measurements.

Results and validation

The site says the team evaluated 15 representative cases spanning assault, theft, drug possession, and DUI. Each case passed through the six agents and was checked for record accuracy, charge classification, evidence findings, precedent relevance, risk factors, and outcome estimates.

93%Reported grounding accuracy
15Cases in evaluation
100Queries in timing sample

Its reported failure modes include missed precedent for less common statutes, evidence admissibility flags triggered too broadly, and risk scores that did not capture some mitigating factors. These are useful findings because they show where an apparently well-grounded answer can still be incomplete.

What the metrics do not establish

The documentation does not publish the case-level annotations, evaluator criteria, raw agent traces, or a breakdown of the 93% grounding figure. A 15-case sample is also too small to establish performance across jurisdictions, charge categories, and case complexity. Grounding accuracy is not the same as legal correctness or outcome-prediction calibration.

Deployment status

The project describes a working five-tab Foundry Workshop application in a staging environment, tested with domain experts. It says the system is ready for integration into production case-management tools. That is not the same as confirming an active production deployment in a public defender's office.

Limitations and safeguards

Legal decision support carries risks that ordinary document search does not. A missed precedent or inaccurate risk estimate can affect a person's defense. The next version should make those risks visible and keep a clear review process.

Traceable claims

Show which case objects, evidence records, statutes, and precedents support each output. Preserve source identifiers and timestamps.

Human approval

Keep attorneys responsible for legal judgments. Require explicit review before exporting or acting on predictions.

Data protection

Limit access to sensitive client records, log access events, and define retention and redaction rules.

Bias and calibration

Evaluate separately by jurisdiction, charge type, and available demographic groups. Historical verdict patterns may reflect unequal treatment.

Stale information

Invalidate cached analyses when case facts, statutes, or evidence change. Track source version and retrieval time.

Reproducible testing

Publish an anonymized test protocol, scoring rubric, and timing method before making broad accuracy claims.

Scope: PDIS should be treated as a research and workflow-support system. The reported predictions are not a substitute for legal research, attorney judgment, or independently validated decision models.

Conclusion

PDIS brings case records, evidence review, precedent search, and outcome summaries into one Foundry application. The six-object ontology gives its agents a shared way to access case information. Hybrid search supports both statute-based filtering and similarity-based research. The project reports faster query responses after removing repeated embedding work and reusing cached results.

The main next step is evaluation, not another agent. Stronger legal benchmarks, clear provenance for every recommendation, and calibrated uncertainty would make the system easier to assess and safer to use. For public defense work, the quality of the review process matters as much as the speed of the answer.