ThinkR dropped React 104 days after choosing it to train six R developers

The skills were supposed to arrive through the reading. February's record priced React at twice the effort of the same work in R and accepted that delay as the cost of getting six R developers fluent in a modern web stack ^[1][8]. The Architecture Decision Record put the team's comfort with React at "junior" and called it "a deliberate investment, not an asset we already hold" ^[9]. By June, with two milestones shipped, the team's account is that the code was arriving faster than it could be read, and so the fluency never accumulated ^[2][3].

As evidence about AI-assisted work, this is one team of six, one internal staffing tool, and a retrospective written by the people who made both decisions ^[8][5][20]. The authors do not pretend otherwise. They wrote both decisions down with their criteria, so the sequence can be checked ^[19], and they are plain about the standing of the argument: "A decision justified by learning is only valid if the learning happens. That is a hypothesis, not an asset, and it can be checked," the post said ^[14].

The two decisions sit 104 days apart, 19 February to 3 June ^[18]. Those 104 days are the entire observation window, and two React milestones were delivered inside it ^[2].

What I find most useful here is the comparison. Three options went into February's table: all JavaScript on React and Supabase, all R on raw HTMX over raw plumber2, and a hybrid ^[7]. Keeping the existing Shiny application and adding what it lacked was not among them. A project brief written a month earlier had already fixed the target stack under the heading "Technical constraints": "Target stack: exit Shiny, REST architecture, classic web site" ^[10].

The post draws the distinction itself. Google's API quotas are imposed on the team; exiting Shiny was a choice, and filing a decision under constraints takes it out of the debate ^[11]. Version 1 of the tool is a 1,200-line read-only Shiny app that has run for years ^[5]. The post says Shiny could have handled all three of version 2's additions: writing to Google Calendar, a simulation mode, historised indicators ^[6][12]. So the baseline that would have told them what the premium actually bought was never costed. When the file was reopened in June, the team went back over every criterion one by one and left the perimeter unquestioned ^[13].