A frontier lab’s careers page is the worst place to learn what that lab is hiring for. By the time a role is publicly listed, the lab has usually been looking for that person — or someone like that person — for between four and nine months, has already had three or four candidates land in late stages, and has, in most cases, already madethe hire and just hasn’t taken the listing down. The public surface is the slowest signal in the system.
What we actually look at — when a candidate asks us “is Lab X really hiring on the post-training team right now” — is a stack of five signals that we’ve calibrated over the last eighteen months of placing engineers into frontier labs. None of them, on their own, is decisive. Combined, they’re usually within a quarter of where the lab actually lands. Headcount is too coarse. Listings are too lagging. Compensation is too noisy. The signal is in the combination.
Below are the five, in the order we weight them, with the heuristic we use to translate each one into a read. We’ll close with the scorecard we run when a candidate asks us to assess a specific lab.
The headcount delta \u2014 actual versus announced.
Most frontier labs announce a headcount target — either explicitly in a fundraise communication, in a press cycle (“we’re tripling research”), or implicitly via leaked seat counts. The gap between announced growth and observed growth is one of the most honest signals a lab emits. A lab that announces 200 hires and shows up six months later having grown by 30 is either being selective or, more often, has hit an internal slowdown that the public messaging hasn’t caught up to.
// How we read itWe track three things: the absolute team-page delta (which underreports), the LinkedIn employee-count delta (which overreports because it includes contractors and renames), and the count of new engineers who appear in commit logs or paper authorship in a given quarter (which is the most accurate, because it counts people who actually shipped something). The three together give a tight range. If the announced growth implies a hire rate of ten per month and we’re seeing two per month show up in commits, the lab is hiring much slower than it’s communicating.
If observed quarterly growth is below 60%of announced quarterly growth, treat the lab as “soft hiring”: public listings are real but the bar has tightened and time-to-close has roughly doubled.
Where the equity refresh band reset.
Frontier labs do an annual or semi-annual reset of their equity bands — the dollar value attached to each level’s RSU or PPU grant. The direction and magnitude of that reset is the lab’s own internal forecast of what it will have to pay to hire and retain in the next twelve months. A lab that quietly raises the L5 refresh band by 40% has decided, on the basis of internal comp data we don’t see, that the market is moving faster than its prior plan assumed.
// How we read itWe see this in two places. First, in network conversations: refresh letters arrive in late Q1 and Q3, and engineers in our network share the deltas anonymously. Second, in the comp ranges quoted on closed offers during the two months after a refresh — the floor of the range typically tracks the new band within a few weeks. When we see refresh-band increases land asymmetrically — research bands moving more than applied bands, or senior bands moving more than mid bands — that’s a directional read on where the lab is feeling pressure.
An L5/L6 refresh increase above 25%in a single cycle is the lab telling its own people: “we are about to lose you to a competitor if we don’t.” Read this as: they are about to be aggressive on inbound.
The signing-bonus pattern.
The size of signing bonuses is a noisy signal. The shape— which seniorities are getting them, against which counter-offers, with which clawback structures — is much sharper. A lab that is paying signing bonuses to candidates with no counter-offer is hiring against a deadline. A lab that’s only paying them on competitive counters is operating from a position of strength. A lab that’s offering signing bonuses tied to one-year vesting cliffs rather than four-quarter linear repayment is signaling that it expects retention difficulty.
// How we read itThree sub-patterns are worth watching. (1) Signing bonuses to non-FAANG candidates.When a lab starts offering bonuses to candidates whose only counter is a startup that hasn’t raised, it has run out of patience. (2) Signing-bonus parity across seniority.If juniors are getting 60% of what mid-level engineers are getting (rather than the typical 25–35%), the lab is having a junior-pipeline problem. (3) Clawback-back-loaded structures.Aggressive clawbacks are usually a sign that prior hires didn’t stay, which is itself a forward-looking read.
If signing bonuses are appearing on offers with no counter-offer attached, treat the lab as “behind plan” on this team — they will move fast on a strong candidate.
The listings-to-roles ratio.
Every frontier lab fills some fraction of its roles through internal transfers and quiet-network referrals, and the rest through public listings. That fraction is one of the most informative — and least public — numbers a lab has. A lab that’s filling 80% of senior IC roles through internal moves and referrals is signaling that its bar on external candidates is exceptionally high, that it values internal context, or both. A lab that’s listing every role publicly is either growing too fast for the internal pipeline to keep up, or has lost confidence in that pipeline.
// How we read itWe back this out by counting the team-page additions in a quarter and comparing them to the number of public listings that closed in that quarter on the same team. The closing date is approximate (we use the date a listing disappeared from the careers page), but the ratio is stable. A team where new arrivals exceed closed listings by more than two-to-one is hiring mostly internally; a team where closed listings exceed new arrivals is either rejecting heavily or has open seats for longer than the cycle.
If a team’s listings-to-roles ratio is below 0.5, the public listing is decorative — the actual hire is being routed through the network. Tell candidates that referral-strength matters more than application quality on that team.
Who\u2019s still there \u2014 retention of the load-bearing engineers.
Every frontier-lab team has between three and eight engineers who are load-bearing on the actual work: the ones who lead the post-training run, who own the eval stack, who shipped the agent harness, who hold the institutional memory of the last training collapse. The retention rate on this small group, over a rolling twelve months, is one of the cleanest forward-looking signals a lab emits. If two of the five leave in a quarter, the team is going to be in a rebuild for nine to twelve months whether the lab acknowledges it publicly or not.
// How we read itWe don’t try to identify these people by titles — titles lag. We identify them by paper authorship, by commit-message signature in open repos, by talk circuits, and by who our network candidates name when we ask “who do you most want to work with on that team.” When one of these named engineers leaves, we know within days. When two leave in a quarter, we re-rate the team’s near-term hiring as “rebuilding” — which means the lab is hiring for that team aggressively at the senior level, slowly at the junior level, and almost not at all at mid-level until the rebuild reaches the next layer.
Two load-bearing departures in a quarter, on the same team, is the cleanest “this team is going to over-hire seniors and freeze mid-level” signal we’ve found. Candidates in the network who fit a senior bar should be told to move now.
// PART 2 · THE SCORECARDHow we combine the five.
The five signals don’t combine linearly. We weight them roughly as shown in each card — headcount delta a little above equity refresh, signing-bonus pattern in the middle, listings-to-roles below that, and load-bearing retention as the lagging anchor that re-rates everything else. But the weights are not the interesting part. The interesting part is what the combination looks like, because each pair of signals tells a different story.
A lab that has a tightening headcount delta anda flat equity refresh is in a quiet slowdown — the public messaging may still be growth, but the lab is internally rationing seats. A lab with a strong equity refresh and visible signing bonuses on no-counter candidates is hiring aggressively against a hard deadline. A lab with a low listings-to-roles ratio andstable load-bearing retention is selectively hiring through the network — public applicants are mostly not getting in. The combinations are the read; the individual signals are noisy.
The scorecard below is the table we run when a candidate asks us to assess a specific lab. Each row is one signal. Each cell is the read we’d give that signal for a hypothetical Lab X in early 2026 — the actual numbers are network-internal and we don’t publish them, but the shape of the scorecard is portable. We score each signal as + (hiring strongly), 0 (steady), or − (slowing or rebuilding).
| Signal | Score | Read |
|---|---|---|
| Headcount delta vs. announced | + | Observed growth tracking 85% of announced. Hiring real, not soft. |
| Equity refresh direction | + | Mid-level band reset by ~30% in last cycle. Lab expects competitive pressure. |
| Signing-bonus pattern | + | Signing bonuses landing on candidates with no counter \u2014 hiring against a deadline. |
| Listings-to-roles ratio | 0 | ~0.7 on this team. Public listing is real, but referral routes still close fastest. |
| Load-bearing retention (rolling 12mo) | 0 | One named engineer departed in the prior quarter; two new senior arrivals netted out. |
The combined read on a scorecard like this — three positives, two neutrals, no negatives — is what we call actively hiring, network-led. The translation we give the candidate: apply, but also work the referral path; the public funnel will move, but the referral funnel will close two weeks faster, and the offer at the end will be on the upper end of the band because the lab is paying for speed.
A different shape — three negatives on retention, headcount, and listings, with positives only on signing bonuses and refresh — is the shape of a team rebuilding while spending. We tell senior candidates to move; we tell mid-level candidates to wait a quarter; we tell junior candidates not to bother.
// PART 3 · ONE LAST NOTEThe signal we deliberately don’t use.
Recruiter outreach intensity. It looks like a signal — if a lab’s recruiters are messaging your network heavily, something must be happening — but it is the noisiest data we collect, and we’ve stopped weighting it. Recruiter quotas swing on a quarterly cadence that doesn’t correlate with what the team actually closes. A team can be in a rebuild and have recruiters at full pace because the recruiting org is centralized and rebalancing slowly. A team can be on a hiring sprint and have recruiters quiet because the team is staffed up on a single trusted partner. Whatever the recruiter outreach reflects, it isn’t what the team is hiring for this quarter.
If you take only one habit from this list, take this one: treat the public surface as the slowest, noisiest version of the signal. Read headcount, refresh, bonuses, listings ratios, and named-engineer retention, and weight each against the others. The frontier lab will tell you a lot more about its trajectory than its careers page does — you just have to read the things the careers page doesn’t say.
// Read next
Frontier hiring · May 2026 snapshot
What the trailing twelve months of network-verified offers say about who’s moving where.
READ →// COMPSigning bonuses became normal in 2026
The structural pressures behind the rise of the signing bonus — and how to negotiate one.
READ →// INSIGHTSFrontier lab insights
Lab-by-lab profiles, hiring posture, comp ranges, and team trajectories — updated quarterly.
READ →Want the scorecard for a specific lab?
Network members get the latest lab-by-lab scorecards across the five signals, refreshed quarterly. Apply to the network and ask for the one that’s relevant to your search.