3. What Round 3 can be

Evidence-based shapes for a third live after two lives, India SDE I / UTA / AUTA. Not a Job 10454435 prediction. Recruiter framing can be wrong. Labels: IE-asked.

What this chapter is not

It is not “Amazon will ask Unique Paths II.” It is not “your R3 is the Bar Raiser.” It is not the OA. It is a map of other loops that already had two lives and then another hour — plus later-slot DSA that still belongs in your hands because Fluency and HM rounds keep mixing code with talk.

HR has told candidates the next hour would be LLD or “three LPs only.” LC 7406809 is the cautionary AUTA post: framed as LP, still got DSA. Stay in Live Code posture until they close the editor.

Shapes that actually showed up

Shape	What they did	Evidence (bible)	Note for you
HM after two DSA	Project / POC / LP depth; sometimes one more coding	LC 7563011; LC 8362604 (Anagrams variation + Combine Garlands)	Closest “true third live” family
GenAI + one DSA	LLM use / verify, then heap or NGE	LC 7850431 k-th largest; LC 7724048 Next Greater	Have both mouths ready
Labeled Fluency, actually DSA	Distance K, O(n), no parent map	LC 7623949; intern echoes NGE / min-heap	Write the tree. Essay later
BR after two lives (R3 skipped or BR is the third)	LP-heavy; sometimes still coding	Reddit AUTA Chennai 1qj304z LP-only BR; LC 6806195 Nice Subarrays on BR	Do not drop DSA
OOD in a later slot	Bookstore word-count; Spring login layers at HM	Shiwangi R3 UTA; Bhavya HM login layers	Not Login Tracker
Tech-stack grilling	Kafka, B vs B+	igreaper — OA-as-R1, likely not fresher AUTA	Do not fake Kafka; see ch. 10

Grid path-count then max-reward, encoded-string frequency, and “subtract a digit of n” are on LC 6800853 AUTA. Flag OA-as-R1: that “R3” may be your already-finished second live. Still practice the DP. Do not tell yourself Amazon scheduled it as your third hour.

Most likely 60-minute reconstructions

These are reconstructions from comparable first-hand posts, not a schedule Amazon sent you.

A. Hybrid LP + one medium DSA (walk in assuming this)

00–10  intro + current role (IQVIA two systems)
10–25  two LPs, why-chains (deadline / feedback / didn’t know)
25–50  one DSA: tree/graph, heap, grid DP, or monotonic stack
50–60  complexity, one Java/CS question, your questions

B. HM / project deep dive, little or no code

00–10  intro
10–40  two or three LPs + one project, 10x and trade-offs
40–55  GenAI should / should-not + how you verify
55–60  your questions

C. Fluency-titled hour

00–15  how you use GenAI (cite-or-abstain, RLS SQL, no LLM on PPE)
15–50  DSA that may be Distance K or k-th largest — treat as a coding round
50–60  wrap

Second research dumps that claimed “100% behavioral R3” exist. Some GFG BR hours are LP-only. Your AUTA-labeled sample is mixed. Prepare A as default, B and C as first pivots.

What to walk in ready for

One clean Java DSA from brute to optimal, narrated — tree/graph or grid DP or heap. Chapter 05.
Distance K without a parent Map, because Fluency was that problem once.
Two failure-shaped STARs: Argus 73% mAP as mistake / negative signal; “no missed ERC on the resume” plus how you would flag a stale costmap. Chapter 08.
IQVIA Hybrid vs vector-only as Backbone; Ylogx RLS vs app-only as Ownership. Chapter 09.
GenAI should / should-not in under two minutes. Chapter 07.
Do not trust “this round is only LP.”

If they ask what changed vs R1/R2: same Live Code, possibly fewer DSA and deeper why on the same story. Difficulty is not proven higher. Seniority of the interviewer may be. Keep Java warm.