Every question a user asks moves through five stages before a response returns. Here's what each one does, and why.
01
Input validation
A context-isolated pre-screening model checks for injection, scope, and routing.
02
Context retrieval
Relevant bill text, case law, and constitutional text are pulled from a controlled corpus.
03
Response generation
The executor model drafts the answer using IRAC — with no conclusion.
On demand
04
Constitutional advisor
For constitutional questions, a cross-model advisor cross-checks the draft.
05
Neutrality audit
A separate auditor scores balance, editorializing, and source grounding.
Stage 01 · Pre-screening
Catch the bad question before it reaches the executor.
A context-isolated pre-screening model examines every incoming question. It catches prompt injection attempts, off-scope queries, and category misroutes before they ever reach the executor. Security and scope enforcement live here — not downstream, where they'd be slower and leakier.
What it does not do: moralize. The pre-screener does not refuse political opinions because the system has opinions — it refuses them because the system is not designed to hold any.
Stage 02 · Retrieval
The executor never answers from general knowledge alone.
Before generating a response, the system retrieves relevant source material from a controlled corpus — bill text, Supreme Court precedent, and constitutional text. This is Retrieval-Augmented Generation: look it up first, then reason. It is the single largest reason Mandate's hallucination rate sits where it does.
Corpus composition. 500+ analyzed bills including 20+ landmark pieces of legislation · 70+ Supreme Court cases · the U.S. Constitution and its 27 Amendments. All primary source, all version-controlled, all citable back to a link. See the RAG deep-dive below.
Stage 03 · Drafting
Legal reasoning as scaffolding. No conclusion at the end.
The executor model builds the answer using a legal-reasoning framework called IRAC, with one deliberate modification: no conclusion. The structure forces each component — issue, provision, both cases, context — into its own slot, at equal length, before the response can be finalized. See the IRAC deep-dive below ↓
Stage 04 · Advisor
Cross-check before confident hallucination becomes a problem.
For constitutional questions specifically, a cross-model advisor is invoked on demand to cross-check the generated response against enumerated powers, the Bill of Rights, and relevant precedent. It exists for one reason: confident hallucination of constitutional interpretation is the failure mode that makes legal AI dangerous. Routing a second model at this point catches contradictions the executor can't see.
Stage 05 · Auditor
Three dimensions. All three must pass.
A separate cross-model auditor reads the final response and scores it on three dimensions: argument balance, editorializing, and source grounding. Responses that fail are rewritten before the user sees anything.
Audit verdict
PASS
Cleared for send
Argument balance0.96
Editorializing0.94
Source grounding0.98
Users never see an unaudited answer. The audit is not a soft pass — a failure returns the response to the executor with the specific failure mode as a correction signal. The loop can run more than once.
Deep-dive · Retrieval-Augmented Generation
The AI looks it up before answering.
General-purpose models generate from training memory. On legal and political topics, that approach hallucinates at 58–88%. Retrieval changes the order of operations — source first, then reason.
Without RAG
Baseline
Who voted for HR 1024 and what did the CBO score?
Generated from training memory
— 58–88% hallucination rate
With RAG — Mandate
Production
Who voted for HR 1024 and what did the CBO score?
Retrieval · controlled corpus
HR 1024 §3(a)CBO/59-2024H.Roll 412W&M markup 08-25Medicare Part D 2022KFF brief 09-25Dobbs v. JacksonArt. I §8S 910 analogCommerce clauseAm. XIVHR 2055 dup.
Generated with citations
✓ Every claim traces back
Deep-dive · Legal-reasoning framework
Balance by design, not by hope.
We borrowed a framework from legal pedagogy because balance by intent doesn't scale. The structure enforces it — equal slots, equal length, no final verdict. Here's what the executor produces for every answer:
I
Issue
Stated neutrally.
What question does this raise? No editorial framing, no loaded verbs.
R
Rule · the provision
What does the bill actually say?
Quoted or precisely paraphrased from the bill text. The reader sees the actual language before any interpretation.
A₁
Application · supporter case
Strongest FOR argument.
Presented at full strength. The best version of the argument, not a sanitized one.
A₂
Application · opponent case
Strongest AGAINST argument.
Identical styling and length to the Supporter case — neither is default.
+
Context · precedent (Mandate addition)
Case law, comparable legislation, constitutional text.
Cites prior Supreme Court cases, comparable bills, and the foundational text — the Constitution and its Amendments. Factual only.
C
Conclusion · deliberately omitted
You decide.
Standard IRAC ends with a conclusion. Ours doesn't. This isn't a missing feature — it is engineered not to.
Mandate presents the case. You form the view.
IRACIssue — Rule — Application — Conclusion, the framework taught in every US law school's first semester.
Honest bounds
Limitations.
01
Outputs are not guaranteed fact.
Retrieval reduces hallucination. It does not eliminate it. Treat every response as a well-sourced draft, not a verdict.
02
Every citation links to primary source.
We encourage verification. If a link is broken or a quote looks wrong, that's a bug — tell us and we'll fix it.
03
When the system doesn't know, it says so.
Rather than fabricating a confident answer, the auditor returns a "not enough in-corpus material" state.