The defense of slow hiring is always the same sentence: 'we're being thorough.' But thoroughness isn't measured in weeks elapsed — it's measured in signal collected. A six-week process built on unstructured chats and brand-name resume screens gathers less real evidence than a well-designed 72 hours, because most of those six weeks were calendar gaps and most of the conversations measured charm. Speed and rigor only trade off when the process is badly ordered. Order it well — strongest, cheapest signal first; every step able to end the process — and you get to a confident yes or no in days, with more evidence than the slow version ever collected.
What actually predicts performance
Strip vetting to what carries predictive signal and you're left with a short list. First, shipped-work review: examining something real the candidate built — a repository, a deployed feature, a system they can walk through commit by commit — and probing the decisions inside it. Work they actually shipped predicts work they will ship. Second, the structured past-decision interview: not 'how would you design X' hypotheticals, but 'walk me through the last time your model degraded in production — what did you check first, what did you find, what would you do differently.' Real events have detail, trade-offs and scars that invented answers can't fake. Third, reference specifics: a former colleague answering 'what did they ship, what happened when it broke, would you hire them again for this exact scope' — questions with checkable answers rather than tone-of-voice vibes. Everything on this list examines evidence of past output. That's the pattern, and it's what the rituals below lack.
The rituals that don't, and what they cost
| Ritual | What it claims to measure | What it mostly measures | Days it typically adds |
|---|---|---|---|
| Unstructured 'culture fit' chat | Team compatibility | Similarity to the interviewer, social ease | 2-5 per round, and it invites bias |
| Whiteboard algorithms for applied AI roles | Technical depth | Interview prep recency and composure under staged pressure | 3-7 including scheduling |
| Long unpaid take-home projects | Real-world skill | Free time; filters out the busiest, often strongest candidates | 7-14, plus review lag |
| Panel round N with no new question areas | Extra confidence | The same signal again, plus everyone's calendar | 5-10 for alignment alone |
| Resume pedigree screening | Capability floor | Brand names, which routinely mismeasure builders from unconventional paths | 0 — but silently deletes strong candidates |
Sequencing for fast kills
Ordering is where speed is actually won. The principle: run the cheapest step with the strongest elimination power first, so the process spends its expensive hours only on people likely to pass. Shipped-work review comes first precisely because it's asynchronous — no calendars — and decisive: a portfolio with nothing real in it ends the process in thirty minutes of one reviewer's time, and a strong one focuses every later conversation on the candidate's actual decisions instead of generic question banks. The structured interview comes second: expensive in senior attention, so it should only see pre-filtered candidates. References run in parallel with the final conversation rather than after it, because they're cheap and occasionally decisive. Compare this with the common inverted order — recruiter chat, hiring-manager chat, panel, then finally looking hard at work — which spends its priciest hours on its least-filtered population and discovers the fatal gap in week four instead of hour four.
- Step 1 — shipped-work review (async, ~30-60 min of reviewer time): kills weak fits in hours, arms the interview for strong ones.
- Step 2 — structured past-decision interview (60-90 min): probes the reviewed work plus the two or three competencies the scorecard names.
- Step 3 — team conversation and references, in parallel: fit questions with the people they'd work with, while two specific references are checked.
- Every step has kill authority. A step that can't end the process is a formality, and formalities are what slow processes are made of.
The 72-hour arc, laid out honestly
Here's what a 72-hour vetting arc actually looks like — with its preconditions stated, because the honest version is more useful than the impressive one. Hour 0-4: brief and scorecard confirmed, shipped-work review of the shortlist completed, bottom half cut. Hour 4-24: structured interviews booked into pre-reserved slots for the survivors. Hour 24-48: interviews conducted and scored against the anchored rubric; references initiated for the front-runners. Hour 48-72: team conversation, references returned, debrief against the scorecard, decision. The preconditions doing the heavy lifting: candidates come from a pre-vetted pool (a bench where baseline screening happened before your role existed), interviewer slots were reserved before the arc started, and the scorecard was written when the role opened. Without those, the identical sequence takes one to two weeks — still a large win over six — because sourcing and calendars re-enter the clock. The 72-hour number is real, but it's the product of preparation, not haste; anyone promising it without a pre-vetted pool is describing the haste version.
Calibration: how fast stays consistent
The legitimate worry about fast vetting isn't rigor per candidate — it's consistency across candidates and interviewers. Slow processes hide their inconsistency in long timelines and big panels; fast ones have less room to average out an interviewer having a bad day. The answer is calibration infrastructure, which is lighter than it sounds. Anchored rubrics: every scored dimension gets concrete behavioral descriptions of what a 2 and a 4 look like, so scores mean the same thing across interviewers — 'a 4 on debugging: independently traced a production model failure to root cause, and can name what they'd instrument to catch it earlier.' Paired scoring on a rotating basis: two interviewers score the same session independently, and gaps above a point trigger a five-minute reconciliation — divergence is where miscalibration announces itself. And a monthly miss review: every 90-day struggle or early exit gets traced back to its vetting scores, asking which step should have caught it. That last loop is what makes the system self-correcting — the fast process learns from its errors on a monthly cadence, which is more than most slow processes can claim on any cadence.
- Anchored rubrics: behavioral descriptions per score level, written once per role family, reused everywhere.
- Paired scoring, rotating: independent scores, then reconcile gaps over one point. Twenty minutes a month per interviewer.
- Miss review: trace every disappointing outcome back to its vetting scores; adjust the step that should have caught it.
- New interviewers shadow-score two sessions against a calibrated colleague before their scores count.
