D4 — Behavioral, Written

40 minutes. Three prompts, roughly 12 minutes each, plus 4 to re-read.

Written rather than spoken, deliberately. Writing exposes whether the content is there. Delivery can paper over a story with no decision in it, and delivery is separately trainable later. Right now we are measuring raw material.

Table of Contents

• What Is Being Measured
• Prompt 1: The Architecture Decision
• Prompt 2: The Disagreement You Lost
• Prompt 3: Where Is AI Headed
• The Structure To Use
• Submission

What Is Being Measured

The reported signal for these rounds is technical leadership, architecture decisions spanning teams, and driving consensus under pressure — with concrete tradeoffs, not soft-skills answers (../research/source-report.md rows 33–36).

There is a second finding that raises the stakes: reported sources name the values / culture round as the most common failure point at a peer lab. That is a remarkable claim about companies whose technical bars are this high, and it means behavioral preparation is not the part you do in week eleven.

And a third: AI-lab levelling is compressed — an "L5 Senior" title reportedly carries Staff-equivalent scope (../research/findings.md). So these are graded at Staff altitude. A story where you personally made a good technical call on your own service is a Senior story. A Staff story has other teams in it, has people who disagreed, and has a decision that was expensive to get wrong.

Prompt 1: The Architecture Decision

"Tell me about an architecture decision you made that affected teams beyond your own. What was the decision, what were the alternatives, and how did it turn out?"

Write 400–600 words. Your answer must contain, explicitly:

• The decision itself, stated in one sentence in the first 60 words. Not the background. The decision.
• The constraint that made it hard. If there was no constraint, it was not a decision, it was a preference.
• At least two alternatives you seriously considered, and the specific reason each lost. "It wouldn't scale" is not a reason. "It required a synchronous cross-region call on the read path, which would have put p99 above our 200ms budget" is a reason.
• Who disagreed and what they wanted instead. If nobody disagreed, the decision was not consequential enough to be a Staff story — pick a different one.
• How you actually got alignment. Not "I explained my reasoning." What did you do? Wrote a doc? Built a prototype that settled it with data? Conceded something to get the important half?
• The measured outcome, with a number.
• What you got wrong. Every real decision has one. Omitting it reads as either dishonesty or as not having looked.

Draw from your actual history — multilingual search and recommendation, streaming, networking, enterprise infra, cloud. Ranking-pipeline redesigns, index-serving migrations, and embedding infrastructure decisions are all excellent Staff-altitude material if you write the tradeoff rather than the tour.

Prompt 2: The Disagreement You Lost

"Tell me about a technical disagreement where you did not get your way. What happened, and what do you think now?"

Write 300–500 words.

This is the highest-signal behavioral prompt in existence and it is the one candidates prepare least. It cannot be faked, because the failure modes are so visible:

What a strong answer has:

• A disagreement that mattered — where being wrong was expensive.
• Your position, stated as strongly as you actually held it. Do not soften it retroactively.
• Their position, stated fairly enough that they would recognize it. This is the single most discriminating element of the whole answer.
• What actually decided it, and whether the process was sound even if the outcome was not.
• How you behaved after losing. Did you commit or did you sandbag?
• Your honest current read: sometimes "they were right"; sometimes "I still think I was right and here is the evidence that accumulated"; sometimes "we were both solving the wrong problem." All three are strong. Only performed humility is weak.

Prompt 3: Where Is AI Headed

"Where do you think AI is headed over the next couple of years?"

Write 250–400 words. Reported as an actual recruiter-screen question (../research/source-report.md row 4).

Required structure — this is what separates a position from a newsletter summary:

1. A specific, falsifiable claim. Not "AI will transform industries." Something that could turn out to be wrong, and that you would notice being wrong.
2. Evidence. Something you have measured, built, or can cite with a number.
3. A falsifier. "Here is what would change my mind." Stating one is the strongest single move available in this answer, and almost nobody does it.
4. The connection to what you would build. Why this claim makes you useful here.

An example claim shape, which you should not copy but should match in specificity: "The binding constraint on useful AI over the next two years is inference cost and latency under agentic workloads, not model capability — because one user action now becomes tens of model calls, and decode is memory-bandwidth-bound so the cost curve doesn't fall as fast as capability rises."

Write yours from what you actually believe and can defend. A claim you cannot defend under one round of pushback is worse than a vaguer claim you can.

The Structure To Use

Not STAR. STAR was designed for behavioral interviews at companies that wanted to know if you were a good teammate, and it front-loads situation — the least interesting part — while burying the decision.

Use DTAO instead, and lead with the decision:

Context goes in a clause, not a paragraph. "On the multilingual ranking pipeline, we decided X" is enough situation-setting for any interviewer. If they need more they will ask, and them asking is good — it means they are engaged rather than waiting for you to finish.

Submission

Write into attempt-d4.md next to this file. Then:

1. Score against RUBRIC.md → Part 4, before reading the answer key.
2. Read ANSWER-KEY.md → D4 for the graded examples and re-score.
3. Note which of your 12–15 career stories these three drew on. That list becomes the seed of the story bank in ../tracks/behavioral/README.md.

On honesty: these are your stories. I will not invent them, embellish them, or let you present a Senior-scope story as Staff-scope. If the raw material for a required story category is genuinely absent from your history, that is a finding — the fix is to go acquire the experience or to find the closest real analogue and be straight about its scope, not to inflate the story you have.