5. Act, measure and learn¶
Module outcome: a learning contract fixing expectation, horizon, metrics, guardrails, comparison and review date before execution.
A decision does not end at choice¶
The cycle closes when the organization compares expected with observed, interprets deviation and changes the next decision. Without a record, outcomes are rewritten retrospectively and little is learned.
Fix the expectation before acting¶
Record lower, central and upper expected results; horizon; proposed mechanism; process measures showing action occurred; outcomes and guardrails; decisive assumptions; stop, hold, scale or reverse conditions; and available comparison or counterfactual. Precision must match evidence—a defensible range beats a fabricated decimal.
Decision quality and outcome differ¶
| Favorable outcome | Unfavorable outcome | |
|---|---|---|
| Sound process | Good decision and outcome; identify what is replicable | Good decision with bad luck or failed assumption; review mechanism and calibration |
| Weak process | Lucky outcome; do not institutionalize without understanding | Bad decision and outcome; correct process, governance and action |
Outcome-only evaluation rewards luck; process-only evaluation may protect decisions that never create value. Both dimensions are required.
Five metric families¶
| Family | Examples | Question |
|---|---|---|
| Latency | Signal-to-decision or action time | Did we decide in time? |
| Process | Adoption, compliance, override, escalation | Did the intervention occur as designed? |
| Outcome | Revenue, cost, retention, error, service | Did the priority move? |
| Guardrail | Equity, quality, risk, load, complaints | What worsened while something else improved? |
| Learning | Calibration, revised hypotheses, correction time | Is the next cycle better informed? |
Prudent attribution¶
A later change does not prove causality. Claim strength depends on design: historical comparison, comparable group, experiment, difference-in-differences, discontinuity or another suitable method. Without a defensible counterfactual, report a post-action observation, list concurrent changes, limit causal language, use multiple cycles where possible and preserve non-attribution.
Cycle record¶
Decision and date:
Chosen option:
Lower / central / upper expectation:
Executed action:
Execution fidelity:
Observed result:
Guardrails:
Comparison or counterfactual:
Deviation from expectation:
Alternative explanations:
Learning:
Change for next cycle:
Owner and review date:
Stress test¶
Can the metric improve without the real priority improving? Can execution capacity support the intervention? What external shock could explain the outcome? Does the window favor the preferred narrative? What result would abandon the option? Who reviews uncomfortable learning?
Important boundary¶
A demo can teach the method. Organizational-impact claims require a real decision, owner, observed action and sufficient evidence. The CDI-BoK keeps these learning classes separate.
Next: integrate the modules in a Decision Brief → or design the measurement system →