📊 Metrics

Measure only what helps the team make decisions.

Meeting-to-Production Lead Time (median per week) lower is better

Illustrative example data.

This is the number the whole model is judged on: how long a client need takes to reach production.

Flow metrics

The core delivery metrics, measured per squad over rolling four-week windows:

Metric	Definition	Direction	Target
Meeting-to-Production Lead Time	Client meeting transcription → running in production	Down	< 24–48 hours for standard changes
Deployment frequency	Production deployments per week	Up	Multiple daily deployments
Change failure rate	Deployments causing rollback, flag shutoff, or incident	Down	< 1%
Time to restore	Detection of production fault → mitigation	Down	< 15 minutes

Metrics per use case

Flow metrics measure delivery end to end. Also measure individual AI tasks, such as test creation or documentation updates. Track these values per work item:

Measure	Question it answers
Time spent	How long does this subtask take for one work item?
Iterations	How many attempts does it take to get it right?
Human intervention	How often does a person have to step in?
First-pass success	How often is it right the first time?

Use existing issue, branch, change-request, and CI timestamps—not manual timers. For token spend, attribution, and financial ROI, see Cost Management. The charts below are illustrative.

Automated test creation (Code)

Generate tests with the code. Track the time spent and whether better context and skills reduce it.

Where the time goes in one work item lower is better

Illustrative example data.

Test creation is the biggest slice here, so it is worth measuring on its own.

Time spent on test creation, per work item (median per week) lower is better

Illustrative example data.

A falling line suggests better context and skills. If it stays flat, check for missing test instructions. If it rises, check for vague acceptance criteria.

Fix loops (Code)

Average fix loops per work item lower is better

Illustrative example data.

Above 3 loops, the work item was probably too big or too vague when it left Plan.

Backlog issue creation (Triage)

Hours from the end of a meeting to a ranked backlog issue lower is better

Illustrative example data.

This is the speed of turning a client conversation into logged, ranked work.

Deterministic controls on the first run (Prove)

Pull requests passing all controls on the first CI run higher is better

Illustrative example data.

A low rate points to weak automated checks in Code or missing context.

Time to restore (Learn)

Minutes from detecting a fault to mitigating it lower is better

Illustrative example data.

The same targets as the flow metrics apply. The use case view shows which part of the work is improving.

AI autonomy & leverage metrics

Metric	Definition	Watch for
AI code generation share	Percentage of code, tests, and docs written by AI	Target 100% (zero human syntax typing)
First-pass green rate in Code	Work items where the agent's automated checks and tests were green on the first run	< 80% indicates Persistent Context drift or missing instructions
Deterministic control pass rate	Pull requests passing all deterministic controls on the first CI run in Prove	< 80% indicates weak automated checks or missing context
Fix-loop count	Auto-fix loops required before a green pass	> 3 loops indicates poor task decomposition in Plan
Acceptance rate	Generated artifacts merged without substantial rework	Falling indicates Persistent Context drift
Rework rate	Accepted output corrected within 14 days	Rising indicates validation is too shallow
Human rework rate	Pull requests where a human had to alter code syntax	Must trend to 0%; fixes belong in context, not code
Automated validation turnaround	Time from the deterministic pass to the recorded Prove verdict	Rising indicates a slow assessment or missing evidence; track requested human checkpoints separately
Assisted share per stage	Share of each stage's output produced with AI	Output concentrated in 4 — Code only
Context hit rate	Sessions where committed Persistent Context was sufficient	Falling indicates artifacts missing or stale

Outcome metrics

Metric	Definition
Hypothesis validation rate	Validated / total hypotheses closed
Feature deletion rate	Invalidated features removed within one cycle
Time to evidence	Backlog issue created → first real usage signal

A rate near 100% may mean the team only tests safe bets. Expect some hypotheses to fail, and remove invalidated features.

Quality metrics

Metric	Definition
Escaped defect rate	Defects found in production / total defects found
Coverage on changed lines	Per pull request, non-decreasing
Mean review depth on high-risk changes	Time spent reviewing, weighted by blast radius

What is deliberately not measured

Not measured	Why
Lines of code	AI makes it meaningless and perverse
Story points velocity	Measures estimation, not value
Assistant suggestion acceptance in the editor	Correlates with nothing that matters

Review

Review metrics monthly. Remove any metric that has not informed a decision in two quarters.