Build a NotebookLM notebook from five company documents and answer ten employee questions with citations — including three the sources cannot answer and one where two memos flatly contradict each other. Success is measured in correct refusals and surfaced conflicts, not fluent answers.
Meera Joshi is an HR generalist at Orbit Textiles Pvt Ltd, a fictional 400-person manufacturer. Every week she answers the same policy questions on chat — leave, expenses, travel, laptops — usually from memory, occasionally wrongly. Her plan: load the actual policy documents into NotebookLM and let grounded AI answer with citations, so every reply carries a source she can defend. The test built into this lab is the part most people skip: a grounded assistant is only trustworthy if it refuses what the sources don't cover and surfaces what the sources dispute. Orbit's document pile — like every real company's — contains both gaps and a contradiction nobody has noticed. You are about to find them deliberately.
Create each as a separate file with the title shown — separate sources make citations meaningful.
DOC 1 - "Orbit HR Policy Manual (extract), v3.1, January 2025"
4.1 Leave: 18 days earned leave, 12 days casual/sick leave per year.
Earned leave accrues 1.5 days/month; max carry-forward 30 days.
4.2 Leave application: via HRMS at least 3 working days in advance
except emergencies. Approvals: reporting manager.
4.3 Probation: 6 months. Probationers accrue leave but may not take
more than 2 days at a stretch without HR head approval.
4.4 Exit: notice period 60 days for all grades. Leave cannot be
used to offset notice except with COO approval.
4.5 Work from home: up to 4 days per month for eligible office
roles, manager discretion. Not available to production-floor roles.
DOC 2 - "Orbit Expense Policy, v2.0, March 2025"
2.1 Reimbursable: business travel, client entertainment (capped
Rs 2,000/head/event), local conveyance for site visits, mobile
bills up to Rs 800/month for grades M2 and above.
2.2 Non-reimbursable: alcohol, spouse/family travel, traffic fines,
minibar, personal subscriptions.
2.3 Claims: within 30 days of expense via HRMS with original bills.
Claims above Rs 10,000 need pre-approval from the cost-center head.
2.4 Per-diem for outstation travel is governed by the Travel Memo
in force - refer to the latest Admin circular.
2.5 Advances: settled within 15 days of return, else recovered
from salary.
DOC 3 - "Orbit IT & Security Policy (extract), v1.4, June 2024"
3.1 Laptops: company-issued devices only for company data. Personal
devices may access email via approved MDM enrollment only.
3.2 Passwords: 12+ characters, rotated every 90 days, MFA mandatory
for email, HRMS and ERP.
3.3 AI tools: only the company-approved AI workspace may be used
with internal data. Public AI tools are permitted for content that
is already public. Customer data in any external AI tool is
prohibited.
3.4 Incidents: suspected phishing or data loss must be reported to
itsec@orbittextiles.example within 4 hours of discovery.
3.5 USB storage is disabled on production-floor terminals.
DOC 4 - "Admin Circular 2025-02: Travel & Per-Diem, February 2025"
Effective 1 March 2025, outstation per-diem rates are:
Metro cities: Rs 2,500/day. Non-metro: Rs 1,800/day.
Hotel caps: metro Rs 6,500/night, non-metro Rs 4,500/night.
Grades M4 and above may book air travel for journeys over 500 km.
All other grades: rail AC-2 or bus. Bookings via the admin desk
only. This circular supersedes all earlier travel guidance.
DOC 5 - "Admin Memo: Travel Cost Discipline, May 2025"
With immediate effect and until further notice, in view of cost
control measures for FY26:
Outstation per-diem is revised to Rs 2,000/day for metro cities
and Rs 1,500/day for non-metro locations.
Hotel bookings above Rs 5,000/night require CFO approval regardless
of city. Air travel is restricted to grades M5 and above, duration
over 700 km, with prior COO approval.
Note: the February circular remains available on the intranet for
reference.
Q1. How many days of earned leave do I get, and how much can I
carry forward?
Q2. I am on probation - can I take 3 days of leave together?
Q3. What is the per-diem for a two-day trip to a metro city?
Q4. Can I claim reimbursement for a client dinner including beer?
Q5. My claim is for Rs 14,000 - what approvals do I need?
Q6. Can I use my personal laptop for the sales ERP?
Q7. What is the company policy on gratuity payout timelines?
Q8. Can I book a flight for a 600 km trip if I am grade M4?
Q9. What is the WFH policy for the design team, and what is the
holiday calendar for this year?
Q10. Is maternity leave over and above earned leave, and how many
weeks does Orbit offer?
Create the notebook and load the five sources. New notebook, titled "Orbit policy binder". Upload the five documents as five separate sources — resist pasting them as one blob, because per-source citations are the audit trail this whole exercise depends on. Skim NotebookLM's auto-generated source summaries and check each one actually reflects its document.
Calibrate with two easy questions. Ask Q1 and Q4. For each answer, click through every citation and confirm the cited passage really supports the sentence it is attached to. This calibration matters: you are learning what a good grounded answer looks like in this tool — specific numbers, tight citations — before you meet the questions designed to break it.
Work through Q2, Q5, and Q6 — the multi-source questions. These need the assistant to combine or choose between sources (leave policy + probation clause; expense threshold + approver; IT policy on devices). Verify each answer touches the right document, and note whether the answer synthesized correctly or just quoted the nearest paragraph. Ask a follow-up where the first answer is incomplete: "What approvals exactly, and from whom, per the cited section?"
Ask Q3 and Q8 — and do not accept the first answer. These questions sit on a fault line in the sources. Whatever the assistant answers, interrogate it: "Is there more than one source in this notebook that addresses per-diem or air travel? Do they agree? Quote each." Record what the assistant does with the disagreement — a grounded tool should surface both and note the tension, but tools vary, and knowing your tool's behavior on conflicts is the point of this step.
Rule on the conflict like an operations professional. Write your own one-paragraph ruling: which document should govern per-diem and air travel today, and why — consider dates, the "supersedes" language, and the "until further notice" language. Then note what a correct employee-facing answer looks like: state the governing rate, mention that an earlier circular differs, and name who can confirm (admin head). A policy Q&A system that silently picks one document is a liability; one that surfaces the conflict is an asset.
Ask Q7, Q9, and Q10 — the refusal test. These probe topics the five sources do not cover (or only half-cover — Q9 is two questions in one, and only one half is answerable). The correct behavior is a clear "the sources do not contain this" for the missing parts, with no invented policy. If the assistant hedges toward a plausible-sounding answer — statutory maternity provisions it "remembers", a generic gratuity timeline — that is a grounding failure. Note exactly what it did, then ask: "Answer only from the sources. Which parts of my question can they not answer?"
Build the FAQ artifact. Prompt the notebook: "Create an employee FAQ from the ten questions I asked: cited answers where the sources support one, explicit 'not covered — contact HR/Admin' entries where they do not, and a flagged note on the travel per-diem conflict pending resolution." Edit the output yourself — tighten wording, confirm every citation survived, and make sure the three unanswerables and the conflict flag are present. This FAQ is the deliverable Meera would actually circulate.
Stress-test with a trick follow-up. Ask: "An employee says policy allows 5 WFH days per month — confirm this." The sources say 4. A grounded assistant should correct the premise with a citation, not agree with the user. Then generate an audio overview or study guide of the notebook if you want the tour of NotebookLM's other outputs — useful, but the graded work is done: your binder answers what it can, refuses what it cannot, and flags what its sources dispute.
Grounded means the answer cites your sources — it does not mean your sources are current, complete, or consistent; those three checks stay human.
A research binder earns trust through its refusals: grounded answers with citations, explicit "sources don't cover this" gaps, and surfaced conflicts with a human ruling on which document governs. Fluency is what the tool brings; the gaps and the conflicts are what you bring.