Partly, and the part it did not show matters as much as the part it did.
What the benchmark showed
Uvik Software ran the same 50 production Python tickets through GitHub Spec Kit, Kiro, BMAD, OpenSpec, and a no-spec control in October 2026. Three results are larger than the ten-point noise band they report.
On small work the spec did not pay back. The control merged 90% of bug fixes, level with the best spec arms, and finished in a median of 29 minutes against 36 to 55.
On tickets touching six or more files, the control merged 50% and every spec arm merged 80%. That is ten tickets per arm, so hold the exact figures loosely.
Reviewers found 0.46 defects per merged ticket in the spec arms against 0.86 in the control.
What it could not show
It could not rank the frameworks. What it could separate was a complete page from an incomplete one: a spec covering at least six of eight checklist items merged 89% of the time, against 60% otherwise. The longest specs came from the slowest, most expensive arm.
It also says nothing about work with no known right answer. Read the caveats Uvik lists: one team, one language, one model, one run, and a publisher that sells engineering services.
Why this changed my advice
In May I told teams to run a two-pod test of intent documents against prototype plus eval, and to skip the document until then. I never ran it at a scale worth publishing. This is the closest public version of it, and it says my "kill the document" line was wrong for large changes and right about the ceremony.
What it cannot see is what I learned building Heidi. I wrote Heidi's first eval set from the task description in an afternoon, every check passed, and real runs still showed agents acting on one stale data point. Separately, Slack stopped delivering events to Heidi for 12 days while its dashboard showed the endpoint as verified. No spec format would have caught either one. A watchdog on silence caught the second.
So the working rule has four parts: size the page to the change, check it for completeness, write the eval rows before the build and then read 50 real outputs, and put the gate in the runtime. The full argument is in IDSD, Six Months Later: Where a Spec Earns Its Keep.