Constraint Learning Loop

Constraint Learning Loop

Owner: Bill Eisenhauer Last Updated: August 6, 2026 Status: 🚧 Field hypothesis Contributors: Bill Eisenhauer

---

The durable capability is not knowing the constraint once. It is shortening the time required to find it, improve it, and notice that it moved.

The loop

flowchart TD
    A["Define accepted and owned"] --> B["Observe end-to-end flow"]
    B --> C["Locate the probable constraint"]
    C --> D["Protect and exploit it"]
    D --> E["Choose one bounded intervention"]
    E --> F["Predict the system effect"]
    F --> G["Run a learning window"]
    G --> H["Issue a learning receipt"]
    H --> I{"Did the constraint move?"}
    I -->|Yes| B
    I -->|No| C

Loading

In prose: define the outcome boundary, observe the complete flow, locate the likely constraint, protect it, make a bounded investment, predict the effect, run a fixed learning window, and record the evidence. Whether the constraint moved or not, begin diagnosis again.

1. Define the outcome boundary

Name the event that means work is accepted, in use, and owned. Avoid completion boundaries that stop at merge, release, framework availability, or partner handoff when later work is still required to realize value.

2. Observe the whole system

The 2025 Kanban Guide defines a minimum set of flow metrics: WIP, throughput, work-item age, and cycle time. Use those at a coherent system boundary, then add context-specific evidence such as rework, first-pass acceptance, adoption, ownership, and constrained human attention.

3. Separate signal from truth

Record both:

• perceived constraint — where current evidence says the limit is;
• confidence — how complete and discriminating that evidence is; and
• governing constraint — what later evidence shows actually limited outcomes.

Instrumentation can be the correct intervention even when it adds no delivery capacity, because it improves the expected value of the next decision.

4. Make an explicit prediction

Before the learning window, write down:

• the intervention;
• the mechanism by which it should help;
• the expected change in accepted outcomes;
• expected side effects in WIP, rework, attention, or cost; and
• where the constraint is expected to be afterward.

This makes the cycle falsifiable and reduces retrospective storytelling.

5. Issue a learning receipt

Field	Example question
Intervention	What changed in policy, capacity, automation, or ownership?
Mechanism	Why should this affect the constraint?
Prediction	What did we expect to govern throughput afterward?
Outcome	Did accepted and owned throughput change?
Burden	What happened to WIP, age, rework, and human attention?
Adoption	Who changed behavior and who owns the result?
Decision	Continue, change, stop, or instrument further?

Portfolio cadence

The loop can coexist with Sprints, continuous pull, quarterly planning, or another execution workflow. The learning window should be long enough for the mechanism to become observable and short enough to avoid protecting a stale theory.

sequenceDiagram
    participant L as Leadership
    participant T as Delivery teams
    participant R as Measurement rig
    L->>T: Fund one constraint hypothesis
    T->>T: Execute through the normal workflow
    T->>R: Emit flow, quality, adoption, and ownership events
    R-->>L: Return a learning receipt
    L->>L: Reallocate against the new constraint

Loading

In prose: leadership funds a bounded hypothesis, teams execute through their normal workflow, the measurement rig gathers evidence, and the resulting receipt drives the next allocation.

---

Related: Operating Hypothesis · Flow Discovery and Intervention Rig · Facilitator Guide