AI configuration

Tune model parameters, matching weights, surfacing thresholds, and guardrails for the in-boundary triage assistant. Changes apply organization-wide and are recorded for audit.

Model parameters

In-boundary model choice and sampling controls for the triage assistant.

Model

All models run inside the authorized boundary. No complaint data is sent to external APIs.

Temperature0.20

Low values keep outputs deterministic and easier to audit.

Response diversity (Top-P)0.10

Kept low for consistent, repeatable classifications.

Maximum response length2048 tokens

Upper bound on tokens per AI response.

Matching weights

How the assistant weighs each signal when surfacing related cases and precedent.

Narrative similarity60

Semantic similarity between the complaint narrative and prior cases.

Statute keyword match40

Overlap on cited statutes, regulations, and controlling authorities.

Prior-outcome influence15

Kept low deliberately — historical dispositions reflect past decisions, not ground truth.

Institution profile35

Surfaces recurring institutions and systemic patterns for human review — not for automatic prioritization.

Recency decay25

How quickly older precedent loses weight relative to recent matters.

Surfacing thresholds

Confidence and risk posture that determine how borderline cases are routed.

Escalation threshold55%

Below this confidence, the case is routed to a senior reviewer — never hidden or auto-dismissed.

Risk sensitivityCautious is recommended — a wrongly dismissed complaint is the costliest error.

High-sensitivity categories

These categories always route to a person, regardless of confidence.

Sexual violence

Locked on

Restraint / seclusion

Locked on

Disability harassment

Locked on

Racial harassment

Locked on

Guardrails

Non-negotiable rules the model must follow. Applied to every AI response before it reaches a reviewer.

Human decision required

The AI can never close, dismiss, or transfer a case. Final disposition always requires an affirmative human action.

Always on

No data egress

All model calls stay inside the authorized boundary. Complaint data is never sent to external services.

Always on

Require citations on every recommendation

Each AI suggestion must reference a specific case, statute, or connected document.

Block unsupported suggestions

Hide any AI output that cannot be traced back to a document in the library.

Redact & flag PII

Automatically mask names, addresses, SSNs and other PII in drafts, and warn the reviewer.

Flag references to minors

Complaints referencing K-12 students require attorney sign-off before AI actions can be applied.

AI instructions

The base system prompt sent with every AI request. Set the tone, expected reasoning, and citation requirements here.

You are an OCR triage assistant. Ground every classification in the complaint narrative and connected precedent. Cite the statutes and prior case IDs that informed each recommendation. Never propose an enforcement action without a human reviewer in the loop.

Keep instructions to cite sources, stay grounded in the case library, and defer enforcement decisions to a human reviewer explicit.