Security & defensibility
The defensibility is the point.
Regulators now warn that generic AI legal interpretation risks the unauthorized practice of law, and that brokers must supervise how AI is used. ComplianceGPS is the supervised lane — citations, authority precedence, disclaimers, and human deferral, enforced in code and in continuous testing. Here is exactly how.
The defensibility stack
The obligation register
Every document is read once and its requirements extracted into a structured, deterministic register — each entry traced to its source. Answers are reasoned over organized facts, not recalled from a model’s memory.
Cited, or no claim
Every factual statement traces to an obligation the system retrieved. If it can’t cite a source, it doesn’t say it — it defers to your broker or counsel instead of guessing.
Authority precedence
When rules conflict, higher authority wins: federal, then state, then association, then MLS, then your office. The conflict is named and the controlling rule applied — not silently resolved.
Shadowing
When your organization publishes an official policy, it supersedes members’ personal copies automatically — so the office speaks with one voice and nobody answers from an outdated manual.
The audit trailComing
Every question and its cited answer is recorded — the supervision-diligence record a broker wants at E&O renewal or a commission inquiry. A one-click admin export view is coming.
Training-off attestation on the MCP
When you connect the platform to your own AI client over MCP, a per-client training-off attestation gate keeps your data out of any training set. The compliance tools stay locked until the client confirms.
The invariants we build on
These are the binding rules of the product — carried on every release and enforced by tests, not slogans. Written for buyers here; enforced in the codebase verbatim.
Locked by default
The API refuses every request without a valid, active membership. No token means no signal of any kind — access is checked at the moment of each request.
No guessing in retrieval
Documents are found by deterministic search, never by a model deciding what you’re allowed to see. The AI reasons over what it’s given; it never picks the data.
Row-level tenant wall
The database itself enforces that one organization can never see another’s data. A cross-organization isolation test suite runs on every change, forever.
Server-held privilege
The keys that can read across the system live server-side only; sensitive operations run under tightly scoped execution, and who you are is stamped on the server — never taken on trust from the client.
Answers, never exports
The tools answer questions — there is no “dump everything” endpoint, no raw-file download, and batch and rate limits with an audit trail on every call.
Your source files stay private
Uploaded PDFs live in a private store, reachable only by their owner through short-lived signed links. They are never committed to a repository and never served back by a tool.
No verbatim copyrighted text
Copyrighted source language is never reproduced word-for-word — at extraction or in an answer. Obligations are paraphrased and cited to their exact provision.
No citation, no claim; never legal advice
Anything outside the registry is an honest deferral, never a guess. Every output cites its sources and carries a not-legal-advice disclaimer — verification and final judgment stay with you and your broker or counsel.
Nothing is ever destroyed
Superseding or removing a document never deletes history. Removed means not-active at the database, and answers can still be resolved as of any past date.
You confirm every change
Nothing is silently added, replaced, or shadowed in your live corpus. Every change to what the system answers from is confirmed by its owner first.
One source of truth
Derived artifacts are regenerated, never hand-edited. Your documents are the truth for your data; nothing important is stored in two places to drift apart.
Your data never trains a model
Tenant data is never used to train models. Bring-your-own-AI keys are encrypted and never logged, and the MCP's per-client training-off attestation gate is live.
Monitors watch, never touch
The automated systems that watch for law and rule changes can detect and flag — they can never mutate your corpus on their own.
Public and private, kept apart
This marketing site is a separate deployment from the product. The application itself is not indexable — your workspace never leaks into a search engine.
Observability from day one
Every automated job reports success and failure to an internal event log with alerting — the system is watched as a build rule, not an afterthought.
Least privilege, proven
Database access is granted explicitly and narrowly — never inherited by default — and an automated test fails the build on any surplus permission.
No accidental exposure
Database extensions are installed away from any publicly reachable surface, closing a whole class of “left the door open” misconfigurations by construction.
Identity checks that actually check
Every function that acts on a caller’s behalf verifies the real, JWT-backed identity — and ships with a test proving it denies the wrong caller and still serves the right one.