What’s changing
We’ve upgraded the model that scores your PRs.
Scores are now significantly more accurate — 3.3x reduction in average prediction error — especially for complex refactors and cross-file changes.
Every score also now includes a plain-language explanation of how the model arrived at its estimate.
What stays the same
Same scoring scale, same API, same reports, same dashboards.
A “3.0” under v2 still means three hours of expert engineer effort. Your integrations and alerts require zero changes.
You can expect a maximum of 5% difference in your overall metric scores with the new model, with most differences coming from a better estimate of bigger PRs.
What’s better
The model understands code at a deeper level:
- Complexity of each line changed
- Cross-file dependencies
It also learns from your team using reinforcement learning
When you override a score, the model adapts to your codebase’s conventions over time.
What you need to do
Nothing. Both models run side-by-side through the end of 2026, so there’s no hard cutoff, you can compare estimates at all times.
During rollout, you’ll have access to:
- A side-by-side toggle to compare v1 and v2 scores on any PR.
- V2 will also backfill your historical data, so you’ll always be looking at one consistent model — no split in your reporting.
- We also recommend hand-picking a few PRs your engineers know well so we can walk through the estimates together and validate accuracy against your team’s judgment.
Optional: score calibration
After rollout, we suggest running a calibration exercise with your team to fine-tune score thresholds to your engineering culture. This can be scheduled anytime.
Questions?
Reach out via Slack or at support@workweave.ai.Last modified on June 12, 2026
