How the clinics decide what to say

1. Answers are grounded in official primary sources

Substantive legal statements come from official primary sources, not from a model's memory or from secondary summaries. For the California Employment Rights Clinic, that means the California Labor Code and Government Code (via the Legislature's official site), the Code of Civil Procedure, the Unemployment Insurance Code, and the agencies that administer these laws: the Department of Industrial Relations and Labor Commissioner, the Civil Rights Department, and the Employment Development Department, plus the relevant federal agencies (U.S. Department of Labor, EEOC, NLRB).

Before a pathway is published, each rule, deadline, agency, and official form is traced to a primary-source page and recorded in a source registry. Law-firm blogs, encyclopedic sites, and unsourced summaries are used, at most, to locate an official page, never as the basis for a legal statement.

2. Deadlines are not guessed by a language model

A missed deadline usually ends a claim no matter how strong it was, so deadlines get the most conservative treatment on the site. A filing deadline is never free-handed by an AI model. Where a reviewed rule exists, the clinic shows the verified statutory period and the event that starts the clock, and it repeatedly tells the reader to treat the earliest possible date as the real deadline and confirm it quickly.

The clinic deliberately does not compute a specific "your deadline is this date" result from a single input, because accrual dates, the continuing-violation doctrine, equitable and statutory tolling, delayed discovery, and administrative exhaustion routinely move the real clock. Presenting a computed date would risk giving a self-represented person false confidence and causing them to miss an earlier deadline. When a deadline cannot be verified to a primary source, the clinic says so rather than inventing one.

3. Every answer states its jurisdiction and review date

Each pathway states the jurisdiction it covers and the date its law was last reviewed. State modules outside California will be published only after review by counsel licensed in that state. A national entry page routes to reviewed state modules rather than generating fifty thin pages that pretend every state has been substantively covered.

4. A predictable answer format

Every substantive answer follows the same shape, so a reader knows what they are getting and what the tool will not do.

1. A preliminary read of what may be going on
2. The facts that would change the answer, and which are still missing
3. The rights that may be involved
4. The deadlines that may apply, with the caveat that only case-specific review confirms a date
5. The evidence to preserve
6. Practical next steps
7. The right form, agency, or procedure
8. When the matter genuinely needs a human lawyer
9. The official sources behind the answer
10. The date the underlying law was last reviewed

5. A second pass checks the first

The design separates the work rather than putting everything into one undifferentiated prompt: eliciting the facts, retrieving sources, doing the analysis, verifying citations, and producing the final answer are distinct steps. After a draft answer is generated, a verification step checks that each material legal statement is actually supported by a cited source, and looks for overstated conclusions, jurisdiction mismatches, unsupported deadlines, outdated sources, and invented agencies, forms, statutes, or cases. If material support is missing, the answer is revised or withheld before it reaches the reader.

Structured results from that verification feed the quality metrics on the impact page, without storing unnecessary identifying information about the person who asked.

6. The model that does the reasoning

The clinics run on an attorney-supervised reasoning model, routed through a single configuration point so the model can be upgraded in one place as better options become available. Substantive legal analysis is not silently downgraded to a cheaper model; lighter-weight processing is reserved for non-legal interface tasks like language detection, menu routing, and analytics classification. The point of naming this is accountability, not branding: a named California attorney designs the prompts, curates the sources, sets the guardrails, and reviews the output.

7. The review workflow, kept in the open

• Each source in the registry records what it is, its issuing authority, its jurisdiction, the topic, the official URL, the effective date, the date it was last reviewed, and its review status.
• A pathway goes live only after its rules and deadlines are traced to primary sources; items that cannot be fully verified are flagged for attorney confirmation and are presented conservatively or held back.
• The "law reviewed as of" date on each clinic is the honest date of the last review, not a marketing timestamp.
• Identified errors are corrected in the open under the corrections policy, and material corrections are logged.