The single sentence I would give a partner this month
If you can only read one sentence about the engineering layer of the venture market in 2026, read this one: AI-native developer tools, agents, MCP servers, code-generation infrastructure, and on-device inference, sustained the highest median commit-velocity acceleration of any tracked sector for 11 of the last 12 weeks.
Every other line in this address is a footnote to that sentence.
What the panel showed
Across 350+ venture-backed startup GitHub organizations and 12 weekly observation windows ending May 5, 2026, the panel surfaced 47 fundraise-precursor profiles, orgs scoring 5 of 6 on the composite. Of those 47, 18 announced a Series A or Series B round inside the 90-day window. Another 11 raised silently, executed a strategic transaction, or shipped a major platform launch. The remaining 18 are either still in the open window or represent organizations where no public material event occurred during observation.
False-positive fraction this quarter: 22 percent, against our earlier working estimate (the SSRN dataset itself carries no funding-event labels). The narrowing is consistent with the 2026 panel having more contributor-quality stratification than the original 2025 cut. Lead-time band tightened by two days at the upper bound, the IQR is now 21 to 45 days for combined stages, with Series A clustering at 23 to 36 days and Series B at 38 to 52.
Three sector-level shifts you should plan around
First, AI-native devtools. The contributor-influx profile inside this sector has decoupled from the rest of the panel. A typical AI-devtools org now adds a senior contributor every 14 days during fundraise preparation, against the cross-sector median of 23 days. Operationally that means the panel's lead-time band is genuinely shorter for AI-devtools than for any other sector, closer to 17 to 32 days. If you cover this sector, your sourcing rhythm needs to compress.
Second, on-device inference infrastructure. The contributor-diversity Gini coefficient inside this sector dropped from 0.42 to 0.31 in three quarters. That's the broadening-team-composition signal, capital being deployed, infrastructure being built, the sector industrializing. The question this raises for any investor is no longer 'is this sector real' but 'who's the index-level exposure'.
Third, verifiable-compute infrastructure. New-repo creation rate inside this cluster is up 3.7x year-over-year. Most of the new repos are tooling, provers, verifiers, attestation libraries, language frontends, not flagship products. That is the signature of a category two to three years upstream of where AI-devtools were in 2024. If you take a 24-month view, this is where the panel disagrees with the headline narrative most sharply.
What I changed about the methodology this month
Two changes shipped to the public methodology in the last 30 days. One, the contributor-quality stratification now separates organizational committers (with @yourcompany.com email) from external contributors (drive-by PRs, dependabot, GitHub Actions bots). The composite score now requires at least two organizational committers in the 14-day window before flagging a contributor-influx signal. This drops about 8 percent of prior false positives on contributor-influx alone.
Two, the dependents-graph signal is now anti-spam-filtered. Organizations whose dependents-graph activity is dominated by automated mirror repositories no longer trigger the dependents-growth flag. This was a long-running false-positive source, especially on cryptography and ML-infra orgs that get auto-mirrored across academic forks. About 3 percent of prior false positives are eliminated.
Both changes are in the SSRN-mirror methodology paper at /methodology and re-run against the public dataset on Hugging Face Datasets.
What I'm watching next month
The thing I want to find out across the next 30 days of observations is whether the AI-devtools acceleration is sustained or reverting. The sector is hot enough that the question worth asking is not 'is the signal real' but 'is the signal saturating'. If by mid-June the AI-devtools median commit-velocity acceleration drops below the cross-sector median for the first time in a quarter, that's a structural top, not just a data fluke, and the panel will say so unambiguously.
Two other watches. Verifiable compute should pass AI-infra (not AI-devtools) in new-repo creation rate by July 1 if its current trajectory holds. And the on-device inference Gini coefficient should drop below 0.30 by the end of Q3 if the sector is continuing to industrialize.
If you read these monthly addresses to size sector bets, those three watches are the ones to mark on your calendar.
Closing, the standing offer
Every monthly address ends with the same standing offer to readers who've made it this far. If you want the 47 fundraise-precursor org names from this month, the ones that scored 5 of 6 on the composite and whose 90-day window is still open, they're inside the Insider Circle private Telegram, which is €197/month, locked while subscribed. The free Sunday digest gets you five names a week. The Custom Sector Sweep, €1,997 once, gets you a deep written report on any one of the three sectors above.
I publish this address on the first Wednesday of every month. Permanent canonical at /state-of-github with the previous twelve months indexed below. From May 2026 onward each address also ships as a 90-second synthetic-voice founder talk on the YouTube channel, same content, narrated. Read the written version, watch the talk, or both.
Talk soon, The Data Nerd