Series: Solo + AI Delivery Efficiency — Part 5 of 6.

Before you quote the 8×

Here's where my own numbers break — and where the independent research is blunt enough that I have to address it head-on. A multiplier like "8×" is only honest if you say what it doesn't mean.

1. Commits aren't value, and the research proves it. Studies consistently find that activity metrics — commits, pull requests, lines of code — inflate under AI without proving output. A randomized controlled trial by METR even found experienced developers on mature codebases were slower with AI on measured tasks, while believing they were faster (a 2026 redesign softened that to roughly flat). Cross-developer data puts real productivity gains near 10%, not 800%. So I treat my 8× as a throughput proxy — and cross-check it against merged PRs and 771 test files, not commits alone.

2. The review bottleneck is real — so I built around it. One analysis found individuals completed ~21% more tasks with AI, but review time jumped 91% as nearly twice the PRs piled up. The gains drowned in the review queue. My answer was to gate every change through a PR before merge — turning the exact bottleneck the research flags into the control mechanism, not an afterthought.

3. Breadth is high; hardening is uneven. On my own honest grading: functional breadth 8.5/10, test coverage 8.0/10 — but production-readiness and operability sit at ~4/10. Security hardening and live payment/KYC rails are still ahead. The AI run bought breadth and speed and deferred some hardening. That's real debt, not a free lunch.

Overall: a real platform, ~6/10 complete, with hardening waves between here and go-live. Not "40 person-years of production-grade, done."

Velocity without this disclosure is marketing. The disclosure is what makes the velocity trustworthy.

One framing that matters: the external productivity figures measure task-level speed on existing tasks; my 8× measures throughput / units shipped on a greenfield build. They're different metrics — which is the point of this post, not a contradiction. Don't read them as directly comparable.

Sources

Figures are paraphrased; verify against the originals before quoting.

• METR randomized controlled trial (experienced developers, measured task time; 2025 result and 2026 redesign): summary
• DX cross-developer productivity data (~10% gains): shiftmag.dev
• Faros AI, AI Productivity Paradox (individual gains vs +91% review time): faros.ai
• Independent synthesis: softwareseni.com

Previous: Part 4 — What a Team Would Have Billed · Next: Part 6 — The Real Leverage