A founder's rolodex lasts about two years.

Every young defense company has a dated hiring cliff: the month it must start recruiting strangers and finds out whether it can.

Each point is the share of everyone hired in that year of the company's life who had previously worked at the same employer as a founder, at the same time, overlapping by a month or more. The two lines are the same calculation under a strict and a broad definition of what counts as a shared employer; the truth sits inside the band. Both fall by roughly the same factor, which is the point. Year one against year three is significant at p < 0.001 under both (χ² 58.3 strict, 101.1 broad). Group sizes: 234, 805, 980 and 636 hires.

SOURCE / METHOD — Network hire = worked at a founder's prior employer with overlapping dates. A founder's own company is stripped from their history, so internal promotions do not count. Restricting the cohort to companies under four years old means the whole roster is the founder-era roster, so company age cannot masquerade as founder pull.

Founder relationships are one of the main doors through which commercial engineers enter defense at all — and that door closes on a two-year timer.

Count of network hires by whether the shared prior employer sits inside defense. Strict definition; the broad definition gives 58% outside. The largest shared employers are Google, SpaceX, Mandiant, the US Navy, Parrot, Anduril and 6 River Systems — a mix that is majority commercial.

The network is a leadership instrument, not a staffing one. Build repeatable process for the junior roles first, because those are the ones it was never going to cover.

Share of hires at each level who had worked with a founder. Strict definition. Director and above against entry level is significant at p < 0.001 (χ² 18.6 strict, 40.0 broad). Founders spend scarce network capital exactly where interviewing fails worst: senior roles that turn on trust and are ruinous to get wrong.

Founding team size is a hiring capability, not a governance detail. A solo founder should expect to buy recruiting earlier than a team of four.

Share of all staff who had worked with a founder, by how many founders the company has. Strict definition. Monotonic, and one against four-or-more is significant at p < 0.001 (χ² 16.4 strict, 27.1 broad).

The founding-mafia story that dominates defense tech coverage describes the tail, not the industry.

Number of companies at each level of network hiring, strict definition, of 80 measurable. The distribution is bimodal rather than a spread: a small set staffed largely from one prior team, and a majority built from strangers on day one. 9 companies account for half of every network hire in the cohort.

Founders are not working a broad address book. They are extracting a slice of one team.

Single prior employers supplying five or more people to one young company, broad definition since several involve large firms. Ten employers account for 53% of every network hire found. Armadin's 54 come from one security organisation that Google acquired: founder Evan Pena led Google's global red team, and Armadin's own chief executive, Kevin Mandia, previously ran Mandiant.

Anyone reading a lift-out as evidence of momentum, or as a governance red flag, is reading noise.

Across 80 companies (70 with funding data). The company that brought the whole crew across is neither more nor less likely to be growing, or funded, than the one built from strangers.

Cohort. All 119 companies in Arena's tracked defense tech universe founded in 2023 or later. Neo-primes are excluded by construction, since none were founded in that window. 81 had both an identifiable founder and at least five other staff; 80 remain measurable under the strict definition. Founding years come from Crustdata company records, not inferred from founder start dates — that proxy was tested and rejected because it dated Hermeus, founded 2018, to 2026.

Strict against broad. The strict line removes shared employers where being there at the same time tells you nothing: firms of 10,000 or more staff, national armed forces, and universities. Headcounts were verified by company enrichment rather than assumed. Manual inspection of the largest excluded lane, Armadin and Google, confirmed a genuine Mandiant red-team lift-out, so the strict figure understates.

Control. A permutation test: founder sets are shuffled between companies 120 times and the metric recomputed. Null mean 0.54% strict and 0.96% broad, against observed 6.4% and 10.8%. Lift over chance is 11.9× strict and 11.3× broad. The two definitions agreeing almost exactly is the strongest evidence the effect is not an artefact of where the line is drawn.

Spot check. Fourteen detected relationships were drawn at random and read by hand against the underlying profiles. That found three defects, all repaired before publication: one-word “Cofounder” titles were being read as hires rather than founders; investors, advisors and board members (502 records) were sitting in the staff denominator; and zero-length overlaps counted as shared time. Repairing them raised the year-one figure and every significance test. Median verified overlap is now 13 months, and 60% of matches overlap for a year or more.

What we removed. Two candidate findings failed testing and were cut rather than softened: an operations-versus-engineering gap (χ² 0.1 under strict, not significant) and a relationship with funding raised, whose direction flipped between the two definitions.

Limits. LinkedIn is self-reported and undercounts very early hires at young companies. Overlap at a shared employer is necessary but not sufficient evidence that two people knew each other, and at a very large employer it is weak evidence for any individual — which is what the strict line exists to test. Founders who have since left are still counted as founders. Person-level rows stay private; only aggregates are published. Every figure here is scoped to these 81 companies, not to defense tech as a whole.