# When a higher score hides less evidence

Case walkthrough — spoken transcript.

AssayScore explained. Understand the number before you act.

A rising delivery score can change where leaders invest attention, even when the underlying work hasn't improved. Assay coordinates software delivery by agents under human governance; AssayScore summarizes selected delivery evidence. This video explains its four components and uses synthetic numbers to show how the formula works. You'll see why missing data can make the headline look better, what the score cannot tell you about customers, and what to check before comparing two results.

Assay is an agent-operated delivery pipeline with human governance. A brief is a bounded work request carrying requirements and acceptance checks. Separate agents implement, review and verify it. Statusgen, Assay’s command-line board and reporting tool, reads the work records and computes AssayScore. People retain consequential decisions.

The components are Speed, Flow, Quality and Value. They describe completion time, selected workflow-state durations, recorded first-pass outcomes and completed work relative to decision delay. They use related but different populations. A score of eighty isn't an industry percentile, and it doesn't mean customers received eighty percent of the value they needed.

For Speed, compare the reporting window’s median completion time with the scope’s trailing ninety-day lead-time distribution. In this synthetic example, the tenth percentile is two days and the ninetieth is ten. A five-day median scores sixty-two point five after inverse normalization. The article specifies the exact percentile and rounding conventions.

Flow is touch time divided by touch plus waiting, multiplied by one hundred. Today, statusgen uses completed workflow-state intervals as the proxy. Thirty hours in progress and ten in selected waiting states produce seventy-five. That doesn't establish thirty hours of active work. Blocked states and unfinished intervals aren't counted in this component.

Quality ninety can mean nine passing briefs among ten readable, linked records. It tracks recorded delivery friction, not all escaped defects. Value relates size-weighted completions to decision-queue dwell hours, normalized against history. Decision delay isn't human effort. For the calculation, assume Value seventy; that number does not measure customer benefit.

Multiply sixty-two point five, seventy-five, ninety and seventy. Take the fourth root because four components are available. The displayed AssayScore is seventy-three point seven. The geometric mean gives a weak dimension more influence than an arithmetic average would, but it can't repair misleading inputs or prevent gaming.

If Flow becomes unavailable, the implementation leaves it out and takes the cube root of the remaining product. This example becomes seventy-three point three, marked incomplete. Missing is not zero. A measured zero would make the product zero; an unavailable measurement changes which dimensions the score summarizes.

Now use another synthetic case: three components at eighty, with Flow at twenty. The complete score is about fifty-six point six. Lose Flow’s source, and the remaining-component score becomes eighty. No delivery behavior improved. The stronger-looking number reflects less evidence. Its incomplete state belongs beside the headline.

Qualgen is Assay’s command-line analyzer of Git change patterns and candidate defect origins. Its learned-risk library concerns individual changes. AssayScore uses a fixed published formula; it isn't a machine-learning prediction. A rolling historical baseline can also change Speed while raw lead time stays constant. Check that benchmark as well as missing inputs. A higher score alone cannot prove an intervention helped.

For your next score change, check the available dimensions, sample counts, exclusions and historical baselines before comparing headlines. Then inspect the component that changed and choose an investigation. Keep that decision with the component history. The companion article gives the exact formulas, runnable commands and current measurement limits.

For more Assay explainers, like and subscribe. You'll find the complete guide at Assay dot finance.