Deterministic Legal AI Controls

Definition

Deterministic legal AI controls are non-generative checks that decide legally sensitive facts, gates, warnings, or workflow eligibility before or alongside LLM output.

Current Synthesis

The episode makes deterministic controls a practical legal AI safety and cost pattern. Case Ready Intake AI uses Python for dates, scope, warnings, and pass/fail decisions so the LLM does not calculate legally sensitive thresholds or decide whether a user request is inside the product’s permitted role. The control layer also reduces cost by stopping or warning on out-of-scope submissions before the full LLM workflow runs.

Key Claims

• Date calculations and limitation-like thresholds should not be left to LLM generation when deterministic logic can check them.
• Pre-flight controls can reduce token spend by filtering unsuitable or out-of-scope matters early.
• Deterministic gates make the system’s refusal and warning behavior more auditable than prompt-only guardrails.
• Runtime QA can compare user input, generated output, and charter requirements before release.
• Deterministic controls are most useful when paired with human review and a documented product boundary.

Evidence

• Date and scope checks: EP 40: Governance First: The Architecture Framework That Makes AI Auditable, Defensible, and 99% Cheaper says the system never asks the LLM to calculate dates and uses Python to detect out-of-scope situations.
• Cost effect: EP 40: Governance First: The Architecture Framework That Makes AI Auditable, Defensible, and 99% Cheaper says front-door date checks prevented unnecessary full LLM runs.
• Prompt-injection and legal-advice boundary: EP 40: Governance First: The Architecture Framework That Makes AI Auditable, Defensible, and 99% Cheaper says the workflow fails when a user tries to push it into legal advice.
• Charter comparison: EP 40: Governance First: The Architecture Framework That Makes AI Auditable, Defensible, and 99% Cheaper describes a final check that output still fits the intended narrative, timeline, and evidence-list role.

Counterevidence & Qualifications

Deterministic checks are only as good as their encoded rules and input parsing. They may miss ambiguous facts, jurisdiction-specific legal differences, misleading user descriptions, or procedural exceptions. The episode does not provide benchmark data proving a 99% cost reduction or legal-risk reduction.

What Changed

• Initial concept created for the legal AI version of deterministic verification and workflow gating.

Related Concepts

• Deterministic AI Verification - broader deterministic-checking frame.
• Legal AI Verification And Auditability - auditability layer supported by deterministic logs and gates.
• Governance-First Legal AI - architecture pattern that uses deterministic controls.
• AI Workflow Triage - workflow-allocation principle separating code, LLMs, and humans.
• Legal AI Hallucination - failure mode reduced when sensitive checks are not delegated to generation.
• Human Judgment Under AI - human review remains necessary around deterministic gates.