NewResearch: the same pull request data supports opposite conclusions

Engineering intelligence

Know whether the work was any good.

Merged this quarter112
  • Slop16
  • Real judgment11
  • Unremarkable85

Throughput tells you a team was busy. It cannot tell you whether anyone exercised judgment, and when producing is free that is the only question left worth asking.

The Slop Index

The number your other dashboard is scoring as productivity.

Slop is work that looks right, passes review, and costs you later. It does not show up in throughput because throughput counts it as a win. We compute it from what happens to a change after it merges, and we price it, so the conversation happens in a budget meeting rather than a retro.

  • Rework: merged work rewritten within ninety days
  • Review drag: reviewer minutes per merged change, by team
  • Compounding: complexity added that never comes back down
  • Orphaned change: work no author can account for at sampling
Slop IndexQ3 · all teams
61/ 100

Priced

$412k

  • Rework38%
  • Review drag27%
  • Compounding21%
  • Orphaned14%
The Judgment Rate

The only number that goes up when people think harder.

Judgment leaves a trace, just not in the event log. It is in what got rejected, what got redirected, and what was thrown away before it cost anything. We read the artifact rather than the metadata, which is why we can see a reviewer who changed the direction of a change and not merely its wording.

  • Redirections: review that changed what was built, not how
  • Rejections: work stopped before it entered the codebase
  • Abandoned designs: the cheapest possible failure, counted as a win
  • Concentration: whether judgment sits with three people or thirty
Judgment Rateby team · rolling 90d
22%

Avoided

$780k

  • Payments34%
  • Identity28%
  • Ledger19%
  • Growth8%

Growth ships fastest and thinks least. No throughput chart contains that sentence.

Delivery and review

The standard reporting, done properly.

Delivery, review health, quality and DORA, attributed to teams and initiatives rather than to individuals, and benchmarked against organisations we will name rather than an industry average nobody can audit.

  • DORA, with the distribution shown and not just the median
  • Review as a gate or a bottleneck, per team and per repository
  • Incidents traced back to the change that introduced them
  • Benchmarks against a named comparison set
OverviewBy team
Lead timep50
2.1d−18%
Change failure90d
7.4%+2.1pt
Review depthper PR
1.9−0.6
Deploysweekly
142+31%
Taste capture

Stop making the same correction twice.

Every review comment your team writes is a standard someone is holding in their head. We turn the recurring ones into context the agents actually receive, so the correction is made once rather than every sprint. It is the one feature here that reduces the work instead of reporting on it.

  • Recurring review comments clustered into rules
  • Rules shipped into the agent's context, not a wiki nobody opens
  • Measured by whether the correction stops recurring
Captured standards12 active
  • Error handling at boundariescaught 34×
  • No new global statecaught 21×
  • Migrations must be reversiblecaught 17×
  • Test the failure pathcaught 9×

Recurrence down 62% on captured rules since adoption.

See what your last quarter actually cost.

The rest of the platform

Behold Labs

Tell us a little more.