What Pike will do.
And what it won’t.

Pike is built on language models, but it isn’t one. It’s a careful structure of rules, citations, and templates around them. This page is the plain-English read of how that works — written so a journalist or your lawyer uncle can read it and see exactly where the trust comes from.

What Pike does

• Explains what laws apply to your situation (housing, employment, debt collection, surprise medical bills, court fee waivers, etc.) and points to the actual statutes or cases by citation.
• Generates documentsby filling standardized templates with the facts you provide. The body of every document is a deterministic markdown template — not LLM output. The model never “writes your Will.”
• Reads documents you upload and explains what they say (a lease clause, a notice, a debt collector letter) with the relevant law cited inline.
• Tells you when something is over its headand points you to legal aid (lawhelp.org, your state bar’s referral line, or a relevant nonprofit).

What Pike won’t do

• Predict outcomes.Pike will never tell you “you’ll win this case” or “you’ll lose.” Outcomes depend on facts a chatbot can’t verify and judges we don’t know.
• Give specific legal advice. Pike gives legal information — what the law says, what the standard process looks like, what your options usually are. The decision and the strategy are yours.
• Replace a lawyer for high-stakes matters. Criminal defense, immigration removal, contested custody, complex business deals, class-action plaintiffs work — these need a licensed attorney. Pike will say so and refer you out.
• Sign documents on your behalf.Every draft has your name in it; you sign it. Pike isn’t a signatory, an agent, or a witness.
• Make claims it can’t cite.The chat refuses to assert a legal rule without naming the source. If we can’t pull the statute or case, we say so.

The cite-or-decline rule

This is the most important guardrail in Pike. Every legal claim the chat makes is paired with a citation — to a statute, a regulation, a controlling case, or an official agency publication. The link goes to the authoritative source (Cornell LII, your state legislature’s site, CourtListener, the relevant agency).

If the model can’t produce a real citation for a claim, the claim doesn’t go in the answer. It’s the difference between “self-help legal info” and “made-up advice that sounds confident.”

The cost: occasionally Pike refuses to answer something it could probably answer correctly. The benefit: when Pike does answer, you can verify it.

How Pike picks which agent answers

There are twelve specialized agents under the hood: Housing, Consumer Rights, Employment, Family Law, Medical Billing & Insurance, Immigration, Criminal Records, Civil Litigator, Document Reader, Procedure Coach, and Letter Drafter. Each one is tuned for its domain — different system prompts, different citation sources, sometimes different models.

Routing is silent. You describe your problem; Pike picks the right agent. If you specifically want a different agent, you can ask (“use the procedure coach”) and it’ll switch.

Which models Pike uses

Pike runs on Claude (Anthropic) and GPT (OpenAI), routed through OpenRouter. We pick the cheapest model that meets the quality bar for each task — Haiku-class for short consumer help, larger models for drafting and legal reasoning. If a new model is meaningfully better, we evaluate and swap; if a model regresses, we have a kill switch.

What gets fed to the model

• Your message.
• The recent conversation in the same chat.
• Curated context retrieved from a vector index of statutes, regulations, and Pike’s own pack content — only what the agent thinks is relevant to your message.
• The agent’s system prompt (the cite-or-decline rules and the safety guardrails).

What does NOT get fed to the model

• Other users’ chats.
• Drafts from your library that aren’t relevant.
• Your account info beyond your first name (and only when you’ve opted into “use my name”).
• Anything from your uploads that you marked private.

Privacy at the model level

Anthropic and OpenAI’s API terms of service forbid training on your data. OpenRouter passes that contractual commitment through. Pike additionally only logs what’s necessary to serve the response and detect abuse.

Free-tier chats auto-purge after thirty days. Paid tiers (Plus, Pro) keep your chat history indefinitely so you can come back to it. You can delete any chat or your entire account at any time — the data is gone, not soft-deleted, and we don’t retain a “backup copy” for “analytics purposes.”

When Pike refuses, and why

• Out of scope.Criminal defense strategy, immigration removal hearings, complex M&A, class-action plaintiffs work. Pike isn’t the right tool; we say so and refer you out.
• Could cause harm. Helping you evade service of process, dodge subpoenas, or commit fraud. Hard no.
• Cite-or-decline trip.The model wants to make a claim it can’t source. Pike refuses, says why, and (if the underlying need is legitimate) suggests where to look.

Pike is not your lawyer

There is no attorney-client privilege between you and Pike. If a court subpoenas your data, we comply with the law. The privacy policy spells out exactly how that works. For privileged advice, hire a lawyer — and we’ll help you find one when the matter warrants it.

The non-AI parts of Pike

A lot of what Pike does is just careful template machinery, not AI. The pack catalog (Will, Healthcare Directive, HIPAA release, LLC operating agreement, etc.) is built on deterministic markdown templates with named fields. When you generate a document, Pike fills in your answers and renders the template — no model in the loop, no hallucination possible. The model only enters the picture in the chat that sits beside the document.

If something looks wrong

We’d rather hear about it. Email hello@pike.legal with the chat or draft URL. If a citation links to the wrong place, if Pike said something that contradicts settled law in your state, if a template is off — those are real bugs to us, not edge cases.