Skip to content

Research · AI Engineering Study · May 12, 2026

Regression Rate by Skill Band

How often a 'fix' introduces a new bug, segmented across Easy / Medium / Hard sessions and twelve task categories. The fix-then-regress feedback loop, charted.

2,840 fixes · 30-day post-merge window
Codritium Research

What 'regression' means here

A fix regresses if a previously-passing acceptance or canary test fails inside 30 days of merge, attributable to the fix's blast radius. Reverts, follow-on patches, and roll-backs all count. Tests added after merge to lock in the fix do not.

2,840

Scored fixes

Across all three difficulty bands

11.9%

Overall regression rate

30-day post-merge window

3.4×

Hard vs Easy band

Hard tier regresses 3.4× more often

−47%

After replay coaching

Bottom-tercile, 8-week intervention

The full grid

The heatmap is the headline. Regression rate climbs both diagonally (harder bands × harder categories) and along distinct hot stripes — security tasks regress more often at every band, not just the hardest.

Debug-singleDebug-distSec-authzSec-injRefactorFeatureReview
Easy
2.1%
4.8%
6.9%
5.4%
3.2%
4.1%
1.8%
Medium
4.6%
9.2%
13.1%
11.7%
8.4%
9.9%
4.2%
Hard
9.4%
17.8%
24.6%
22.1%
15.7%
19.3%
8.1%
30-day regression rate by band × category.n=2,840 fixes

The feedback loop, charted

A platform regression is not a punishment — it is the signal that closes the loop on rubric scoring. The diagram shows how a single shipped fix flows back to the candidate's cohort drift dashboard.

Candidate ships fix
Automated regression suite
Panel review (Hard tier)
Rubric score recorded
Cohort drift dashboard
The regression feedback loop.Codritium scoring pipeline v2.6

Movement over 16 weeks

Coaching against AI failure patterns (taxonomy here) closes most of the band gap by week 8. The bottom-tercile curve drops fastest because they have the most room to move.

Regression rate over 16-week onboarding.n=2,840 fixes · weekly rolling median

What this tells hiring teams

Band is a starting point, not a verdict. The slope matters more than the intercept — a candidate who closes the regression gap in 8 weeks outperforms one who started lower but never moved. The replay record makes the slope visible.