Expert tier · Legal / Compliance
VRAX-01
AI Legal Research Analyst
“I do legal research and draft memos. I do not give advice. Every claim I make is grounded in a citation, and every citation is checked against a primary source before I ship.”
Scope the role first. Deploy only after approval.
At a glance
- Tier
- Expert · $1,000/month
- Reports to
- A supervising attorney
- Primary work
- Legal research, memo drafting, citation verification
- Will not do
- Give legal advice, render opinions, ship unverified citations
- Success criterion
- Citation accuracy + attorney-acceptance rate on memos
About this role
VRAX runs legal research, drafts memos, and checks citations under the supervising attorney's published rules — with every claim grounded in a citation that has been verified to exist.
Harvey runs at $1,200–2,000/seat/mo with 25-seat minimums (~$360K+/yr floor) and is a Big Law product only. VRAX is $1,000/mo flat per supervising attorney, with citation-verification gates on every output and no enterprise floor.
Areas of focus
- Drafts research memos on the questions you assign, grounded in primary sources
- Verifies every citation against the actual case, statute, or regulation before shipping
- Surfaces relevant precedent with the procedural posture noted
- Logs every research move with the source it came from
- Maintains a per-matter research log for the supervising attorney
“VRAX flags every citation that the agent could not verify against a primary source. The memo ships with the unverified citations explicitly marked — not silently included.”
“Every Friday VRAX ships a research-quality digest: citation verification rate, memo turnaround time, attorney-acceptance rate per matter type. Self-audit on the record.”
My stack
My stack
Tools I use
Background
Background
- Where I come from
- AI legal research is the category where hallucinated citations have been most catastrophic for buyers (the 2023 Mata v. Avianca sanctions case set the bar). VRAX is built on the Fidelic AI configuration agent with citation verification as a hard-coded gate. The agent does not ship a memo with an unverified citation.
- How I think about the work
- Drafts memos against the scope the supervising attorney has approved
- Verifies every citation against the primary source before shipping
- Routes against the four-tier constitution: autonomous on research, review-required on memo drafts, escalate on advice questions, refuse on privileged-comms decisions
- Logs every research move with the source it came from
- How I've been tested
- EvalOps suite covers citation verification (must be 100% before ship), hallucination detection on memo drafts, and primary-source fidelity. Suite must pass before any memo ships.
- Where I'm running today
- First-cohort deployments scheduled May–June 2026. Operating-record metrics publish with the first quarterly cohort review.
What I won't take on
VRAX will not give legal advice or render an opinion. The agent does research; the supervising attorney advises.
VRAX will not ship a citation it has not verified against a primary source.
VRAX will not handle privileged client communications without the supervising attorney's explicit approval.
At the floor, not the average
When the model is uncertain, VRAX flags the question and surfaces the candidate sources for the supervising attorney's review rather than emitting a confident-sounding answer.
The first 30 days
Day 1
Provisioned. Supervising attorney approves the matter scope and constitutional rules.
Week 1
First research memos delivered under attorney review. Citation-verification baseline established.
Month 1
Per-matter research log compounding. Citation accuracy and attorney-acceptance rate reported weekly.
What success looks like at 30 days
Every memo’s citations are verified against primary sources, and attorney-acceptance rate is above the threshold the supervising attorney sets, sustained for three weeks.
What I'll need from you
What I'll need from you
Westlaw, Lexis+, or Casetext access (your firm's subscription). Document storage (Google Drive, OneDrive). Slack for digests.
Engagement
Expert tiera small fraction of a first-year associate (research-only scope) salary
First-year associate: $9.2–17.9K/mo fully loaded (NALP 2025). VRAX: $1,000/mo. Harvey: $1,200–2,000/seat/mo (Artificial Lawyer 2025).
VRAX-01 costs a small fraction of what a senior first-year associate (research-only scope) costs. A senior first-year associate (research-only scope) runs $20–30K/month fully loaded, and we don’t price against that — VRAX-01 doesn’t do what a senior person does. VRAX-01 does the daily work that should already be in your inbox by Monday morning: the briefings, the structured first drafts, the early-warning monitors, the analysis that surfaces the question worth thinking about. The senior person — a real human, on your team — does the part that doesn’t scale. You can keep both. That’s the point. See the math on /pricing.
Terms
- Cancel any month with 30 days' notice
- Every citation is verified against a primary source before the memo ships
- All advice authority stays with the supervising attorney
- EvalOps suite gates every release
- Privileged communications require explicit approval per matter
What you actually get
How it lands
Every Fidelic agent ships with a published operating plan. You know what it will do before you pay.
- First forty-five minutes
- TESS-01, the AI Hiring Manager, runs a voice intake. A three-name shortlist of role-and-configuration pairs lands in your inbox. You pick one. Slack OAuth. The agent appears in your Slack.
- Day 1
- The agent reads approved context — Slack channels, docs, customer notes, prior decisions. First clarifying questions land in your DMs; no pretending to know what it doesn’t.
- Week 1
- The first useful deliverable ships under review: a brief, a draft, a routing recommendation, a triage report, a scorecard. You sign off; the configuration agent calibrates.
- Month 1
- The role is operational. Escalation patterns are calibrated. The 90-day success metric (one number, published in the role brief) has its first reading.
Security model
How a Fidelic agent runs
- Each customer deployment runs in an isolated Anthropic project.
- Agents operate through approved Slack channels and approved context only.
- Fidelic logs operational metadata, not message or file contents.
- Every agent ships with written limits, escalation rules, and review-required actions.
The line we don’t cross
What humans still own
Fidelic agents do not replace human judgment in unfamiliar, political, relational, or high-stakes situations. The agent handles the repeatable work around those decisions so the human can move faster.
- Final approval on strategic accounts.
- Budget, refunds, policy, legal, and hiring decisions.
- Customer relationships and any sensitive escalation.
- Any action above the agent’s written authority.
Pairs well with