# AssayScore: a guide to the calculation

Field guide — spoken transcript.

AssayScore. A field guide to the calculation.

AssayScore is a summary of selected delivery evidence from Assay, a system for software work carried out by agents under human governance. This standalone field guide is for engineering leaders and practitioners who need to interpret or check a report. We'll define each component, work through the calculation with synthetic numbers, and examine missing inputs and changing baselines. You'll leave knowing what supports a comparison, which caveats belong beside the score, and why it is separate from machine-learning predictions.

Assay organizes agent-operated software delivery under human governance. Work is defined in briefs: bounded requests with acceptance checks. Implementation, independent review and verification produce separate records. Statusgen is the command-line tool that turns those records into a work-status board and delivery reports, including AssayScore. Verification on main is the current delivery boundary.

Speed compares the window’s median authored-to-done lead time with individual completions in a trailing ninety-day baseline. Lower lead time earns a higher Speed score. With tenth- and ninetieth-percentile endpoints of two and ten days, a five-day median produces sixty-two point five. At least five baseline observations and a nonzero percentile band are required. Unfinished work needs a separate queue-age view.

Flow uses completed intervals: in-progress time counts as touch, while selected queued and acceptance states count as wait. Blocked and done intervals are excluded, and the current open interval isn't estimated. Thirty touch hours divided by forty counted hours gives seventy-five. This is a workflow-state proxy, not measured active labor.

Quality is first-pass yield multiplied by one hundred. A counted brief needs a resolvable merged pull request and readable reviews. Passing means no recorded request for changes, a last recognized verification verdict that isn't failure, and no unresolved finding naming the brief. Nine passes among ten counted briefs give ninety; excluded records still matter.

Value divides size-weighted completed briefs by the elapsed creation-to-closure hours of tracker issues awaiting a human decision. Small, medium and large briefs contribute one, three and eight points. The ratio is normalized against historical weekly ratios. This measures delivery relative to recorded decision delay. It doesn't measure human labor or customer value, and zero decision hours make it unavailable.

For four available component scores, multiply them and take the fourth root. With sixty-two point five, seventy-five, ninety and an assumed Value of seventy, the displayed result is seventy-three point seven. Components are rounded before combination. To reproduce the emitted composite, use those emitted component values; the article explains event-level rounding and baseline conventions.

When Flow is missing from that example, the three remaining components produce seventy-three point three using a cube root. The output marks the score incomplete. If all dimensions are unavailable, the score is null. Preserve the missing reasons. An unavailable input must never be displayed as a measured zero.

The more revealing synthetic example starts with Flow twenty and the other components eighty. The complete score is fifty-six point six. Remove Flow and the partial score is eighty. Nothing improved. Rolling baselines can also move while raw performance stays unchanged. Compare populations, available dimensions and baseline bands before inferring a trend.

Qualgen, Assay’s separate command-line repository-quality analyzer, studies Git change patterns and candidate defect origins. Its learned-risk library can evaluate individual changes using historical features and time-bounded labels. That is separate from AssayScore’s fixed formula. The broader machine-learning feedback cycle remains proposed; neither a prediction nor a composite score proves customer benefit.

On your own report, retain the reporting window, component states, raw inputs, observation counts and baseline bands. Investigate a component change before acting on the headline. Keep customer outcomes alongside delivery evidence when judging a repair. The article includes command examples and definitions you can use to audit that decision.

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