One evening, one public repository, no interviews, no dashboards. This is what an engineering org looks like on film — read from nothing but its commit history.
PostHog's handbook states the company pays engineers near the top of market “because we believe there is a 10x difference between average and top engineers.” Almost every company believes some version of this. Almost none can measure it, and none of them publish the evidence.
PostHog accidentally does. Their product is open source, so their engineering org's complete work record sits in public. I pointed the instrument at it: 13,548 commits, 205 human contributors, 120 days. Five findings.
Define the core roster as everyone with ≥20 commits in the window: 92 people producing 96% of all output. The median core engineer shipped 94 commits. The top contributor shipped 975: a 10.4x spread, in an org whose comp philosophy is built on the number 10.
| Engineer | Commits / 120d | Active days | Volume |
|---|---|---|---|
| Engineer A | 975 | 58 | ██████████████████████████████ |
| Engineer B | 641 | 105 | ████████████████████ |
| Engineer C | 417 | 91 | █████████████ |
| Engineer D | 406 | 86 | ████████████ |
| Engineer E | 341 | 75 | ██████████ |
| Engineer F | 259 | 71 | ████████ |
| Engineer G | 215 | 55 | ██████ |
| Engineer H | 214 | 63 | ██████ |
| …84 more ≥20 commits | median 94 | — | ███ |
Engineer B is its own finding: 105 active days out of a possible ~120, meaning commits on 87% of all days in the window, weekends included. Volume gets celebrated; that kind of metronomic consistency is rarer and harder to hire for.
The top five contributors hold just 24% of output; the top twenty hold 48%. Half of everything shipped comes from outside the top twenty. That is a deep bench: the org would absorb the loss of any single engineer, including the 10.4x one.
Forget the star. PostHog's median core engineer ships 94 commits per 120 days: roughly one merged change per working day, sustained for four months. In most mid-market companies, that figure would make you the top performer on your team, possibly in the org. Here it makes you exactly average.
This is what "we pay 90th percentile" buys when it works: not one hero, but a median that would be a standout anywhere else.
Of 13,548 commits, just 44 are reverts, or 0.3%. Shipping this fast with this little walk-back means the safety rails (review, CI, flags) are real, not ceremonial.
Monthly output rose from 2,384 human commits in March to 4,310 in June; the best month in the window is the most recent one. Because PostHog also publishes its compensation philosophy (90th percentile; public market data puts their median US engineer around $255K), you can price the machine from the outside: roughly $23M in annual engineering payroll producing ~34,000 changes a year, about $675 per shipped change. Most CFOs cannot produce that number for their own company. It took one evening to produce it for someone else's.
The obvious suspicion about agent-era velocity is that review died. The GitHub API says otherwise: of the last 100 merged PRs, 95 carry an approving review, 9 are bot-authored, and the median PR goes from opened to merged in about one hour. The p90 is 14 hours, so even the slow tail clears in a day.
The mechanism is a flywheel, not heroics: the median PR is 4 files and 94 lines, which reviews in minutes, so nobody defers it, so merges are cheap, so people ship small, and the loop closes. Review latency may be the single best one-number gauge of an engineering org: trivial to measure, impossible to fake without actually fixing the system, and it predicts everything else on this plate.
Source: the public posthog/posthog repository, last 120 days, bots and automation excluded, merge commits excluded. Core roster = ≥20 commits. Roster anonymized as a courtesy, not for secrecy; the data is public, so every number above can be checked.
Limits, stated plainly: commit counts are not value. Line counts are inflated by generated code. A public monorepo shows most, not all, of an org's work (some contributors here are community, mitigated by the core cutoff). Cross-org comparisons must normalize for PR grain (one chunky PR elsewhere ≈ two or three PostHog-sized ones; both conventions squash-merge, so within-org ratios hold) and for stakeholder surface: a product whose customers are engineers pays its communication tax in writing, which compounds as docs and specs; orgs serving non-technical stakeholders pay it in meetings, which vanish. A fair examination prices that discount first and reports what survives it. And a healthy surface doesn't rule out deeper issues; this is the instrument's quick pass. The full examination reads work-type (who does load-bearing systems work vs. contained UI), per-person churn loops, fade patterns, manager-absorption signatures, and bus-factor concentration: the things that only show up when you read the history, not count it.
Why PostHog: they published the claim and the evidence. The verdict happens to be flattering; this is what a healthy org looks like on film. The instrument exists because most orgs are not this one.