Where BPS AI Software publishes its point of view: benchmarking discipline from Error Bars, field notes from the PairCoder blog, and our own dispatches on AI in manufacturing. Every entry is something we wrote, not a link we picked up.
Benchmarking
Error Bars is a BPS team publication: every post runs a real eval panel with repeated runs, confidence intervals, and downloadable fixtures instead of a single vibe check. Receipts, not vibes. Read Error Bars →
Error Bars · July 27, 2026
The first release of The Battery, our public fixture set: four graded tasks you can point at your own coding agent this afternoon and grade with the same rubric we use against our production fleet.
Read on Error Bars →Error Bars · August 21, 2026
The second Battery drop, live in the public fixtures repo: a set of fixtures built specifically to check whether a model's stated reasoning holds up against what it actually did, not just whether the final answer scored well.
Read on Error Bars →Error Bars · coming soon
Additional Battery fixture drops are in progress. Material pending from the BPS team.
Error Bars · coming soon
A write-up of how we benchmark and route models across our own production agent fleet. Material pending from the BPS team.
PairCoder
PairCoder is our flagship product, and its blog is where David and Mike write about what actually happens running an AI coding agent on real work, under the BPS AI Software byline, not a marketing team's.
PairCoder Blog · David Wiens · February 15, 2026
Every major feature in PairCoder has a "we got burned" origin story. After 400+ tasks and 12 months of dogfooding, here's what survived contact with real work at real scale.
Read on the PairCoder blog →PairCoder Blog · Mike Doggett · March 6, 2026
After 12 months and 75K lines of AI-generated code, the job looks nothing like it used to: harder in different ways, not easier.
Read on the PairCoder blog →Articles & papers
Our own notes on applying AI in manufacturing, written up and kept here.
