From the wave-0 evidence ledger
Recorded claims
Each claim states what the inspected source says, at the recorded location—bounded by its scope and graded by its confidence. Nothing here is a synthesis across works.
Across three field experiments involving 4,867 software developers at Microsoft, Accenture, and an anonymous Fortune 100 company, the preferred instrumental-variable/treatment-on-the-treated estimate for actual coding-assistant use induced by randomized access under imperfect compliance was a 26.08% increase in completed tasks; this is not the randomized-access intention-to-treat effect.
Scope: Software developers and measured completed tasks in three firms; not all high-skilled work or firm financial performance.
Gains were larger for less-experienced developers, reinforcing an expertise-gradient pattern, but the study does not imply that every experienced developer or mature repository benefits—an important contrast with the METR result in AIORG-C023.
Scope: Three company contexts and their coding workflows; external validity bounded.
Boundaries
Limitations & independence
Recorded at coding time, carried with the work forever. A claim without its limits is not evidence.
Recorded limitations
- Software-development task family
- Tool use and outcome logging differ across firms
- Less-experienced-worker heterogeneity may not transfer to other professions
- Task completion is not firm profit or software quality over the long run
- Imperfect compliance means the preferred 26.08% IV/treatment-on-treated estimand is not the randomized-access intention-to-treat effect
Source independence
Academic/vendor-affiliated multi-company field experiments, including Microsoft, Accenture, and a Fortune 100 firm; product and employer involvement disclosed by setting.