Slow news fortnight (post-funding-secured, pre-fall-launches), so a working dispatch from the interview trenches, where I’ve spent this quarter as loop lead for our next eng cohort, the work-sample experiment, two years matured, with data worth reporting.
What’s held: work-sample auditions beat whiteboard algorithms by every measure we track, signal quality, candidate experience scores, and (the one that matters) six-month on-the-job correlation. What’s surprised me: the strongest signal in the whole loop is the debugging session. We hand candidates a real (sanitized) broken service from our incident archive and watch them navigate. Not solve; navigate. Do they form hypotheses or thrash? Read logs or guess? Say “I don’t know, here’s how I’d find out”? Five years of this blog is essentially that skill journaled (from the first reflog panic to the Thai-cave rehearsals), and watching candidates perform it live has taught me it’s rarer than algorithmic fluency and vastly less practiced. The industry drills LeetCode; the job is 70% forensics. We interview for the drill.
Also logged: the US Open began this week with a roof on Ashe and Serena chasing the record books post-maternity; the group chat’s Fortnite receipts have been formally framed; and Fantasy PL draft night concluded with Null Pointer Exception drafting, for the first time in six years, zero goalkeepers before the final round. Character development.
TIL: structured debriefs. Interviewers write independent verdicts before discussing, or the first loud opinion becomes everyone’s memory (the stale-cache lesson, for hiring). Consensus is a race condition; sequence your writes.