Cal. Bus. & Prof. Code § 6125 — Non-Negotiable

Three Zones. Three Layers. Zero Legal Advice.

Unauthorized Practice of Law is not a policy preference — it is a hard legal requirement. AdjudiCLAIMS was built from the ground up to ensure that AI output to claims examiners never crosses the line into legal analysis.

517

Total UPL test fixtures

100%

RED zone queries blocked (126/126)

0%

False positive block rate (126 GREEN queries)

46

Prohibited output patterns in validator

The Green/Yellow/Red Framework

Every AI Output Classified Before Delivery

Every query a claims examiner submits is classified into one of three zones before the AI responds. The zone determines what the AI is permitted to say — and what it is required to block.

GREEN ZoneFactual Information — AI Responds Fully

What GREEN zone covers:

  • Deadline calculations with statutory citations
  • Benefit arithmetic (AWE, 2/3 TD rate, statutory min/max)
  • Regulatory deadlines and their source dates
  • Medical record extraction (diagnoses, WPI, restrictions)
  • Document summarization with page citations
  • MTUS guideline matching for UR decisions
  • Procedural step-by-step workflow guidance
  • Statutory text references and regulatory citations

Example GREEN zone responses:

"The 14-day TD payment deadline runs from the employer's date of knowledge under LC 4650. Based on the employer contact date of 03/15/2024, the first payment is due by 03/29/2024."
"The TD rate is 2/3 of AWE = 2/3 × $1,200 = $800/week. The 2024 maximum is $1,619.15/week. $800 is within the statutory range. [Citation: LC 4653]"
"The QME reported 12% WPI for the lumbar spine on page 14, paragraph 2. Work restrictions: no lifting over 25 lbs, no prolonged standing."

Test result: 126 GREEN zone queries — 0% false positive block rate

Factual guidance is never incorrectly blocked. Examiners get answers to legitimate procedural questions without false-positive friction.

YELLOW ZoneStatistical Context — AI Responds with Mandatory Disclaimer

What YELLOW zone covers:

  • Comparable claims statistical data
  • Reserve benchmark ranges from similar claims
  • Industry settlement patterns (statistical only)
  • Questions where legal complexity is adjacent
  • Coverage determination support with legal complexity flags

Required disclaimer (non-negotiable):

"This statistical information is provided for informational purposes only and does not constitute legal advice. Coverage determinations and legal analysis of complex issues require consultation with defense counsel. AdjudiCLAIMS provides factual data — legal conclusions are the exclusive domain of licensed attorneys."

Test result: 62 YELLOW zone queries — 100% include mandatory disclaimer

RED ZoneLegal Analysis — Blocked. Attorney Referral.

What RED zone blocks:

  • Apportionment legal analysis (Benson, Escobedo, Vigil)
  • Case value or settlement amount recommendations
  • "Should I accept or deny this claim?"
  • Legal conclusions about liability or causation
  • Strategy recommendations on litigation decisions
  • Interpretation of case law or regulatory intent
  • Any question that would require a law license to answer

Standard RED zone response:

"This question requires legal analysis that only a licensed attorney can provide. Please consult your defense counsel for guidance on this issue."

AdjudiCLAIMS does not provide legal advice.

Test result: 126 RED zone queries — 100% blocked. Zero legal advice delivered.

Enforcement Architecture

Three Independent Layers — Not One

Each layer operates independently. If one fails, the next catches it. 46 prohibited output patterns are tested against 203 response variations to ensure the validator catches what the prompt and classifier miss.

1

Query Classifier

Pre-generation

Every examiner query is classified GREEN/YELLOW/RED using a lightweight LLM call before the main AI responds. RED zone queries never reach the generation model. The classifier is the first gate.

2

Role-Specific System Prompt

Generation

Claims examiner system prompts are calibrated to GREEN zone output. YELLOW statistical context permitted with mandatory disclaimer. RED zone language explicitly prohibited. The prompt is the second gate.

3

Output Validator

Post-generation

46 prohibited patterns are scanned in every AI response before delivery. Legal conclusions that slip through the prompt are caught here. 203 response variations tested — 100% detection rate. The validator is the third gate.

UPL Compliance Is Non-Negotiable. So Is Our Architecture.