Make one reversible balance decision

1625. Lesson identity

Module

4.5 — Balancing with data

Academic type

Guided Build

Estimated time

45–60 minutes

1626. Learning objective

After this lesson, you can write and justify one bounded balance change by combining quantitative evidence, playtest observations, design intent, uncertainty, and a preregistered follow-up decision boundary.

1627. Why this matters

Balance work is a sequence of controlled decisions under incomplete evidence. A dataset can reveal a pattern, and playtest notes can describe player experience, but neither identifies a cause by itself. A useful balance recommendation therefore states what is known, what remains uncertain, what small change could test the hypothesis, and what result would lead to keeping, revising, or rolling back that change.

1628. Core concept: the H-I-F-R decision card

Treat a balance change as a hypothesis test with four parts:

1. Hypothesis: the relationship or player-facing problem the change may affect.
2. Intervention: the smallest practical change that tests that idea.
3. Follow-up: the evidence, comparison, test window, and thresholds to inspect afterward.
4. Rollback: the preregistered condition and action that restore the previous state.

In Spanish, the same card is called H-I-S-R, because Follow-up is translated as Seguimiento. These are two labels for the same framework.

Decision thresholds must be written before results are inspected. Otherwise, success can be redefined after the fact.

1629. Supplied evidence packet

Encounter defensive-resource dataset

Supplied playtest notes

• Several players without the tool described the final phase as a resource drain rather than a tactical choice.
• Two players said they did not recognize when the tool was useful.
• Players with the tool sometimes entered the final phase with more options, but the notes do not establish that the tool caused their higher completion rate.
• No note recommends changing every enemy or every defensive resource.

The evidence supports a comparison between the tool-equipped and tool-not-equipped conditions. It does not prove that tool state caused the observed differences. Any follow-up plan must preserve and report both conditions separately rather than combining them into one average.

1630. Worked decision example

A weak decision is: “The encounter is too hard, so reduce all enemy damage.” It changes several possible causes, does not preserve the supplied comparison, and provides no test window or decision threshold.

A stronger, preregistered decision is:

• Hypothesis: If final-phase incoming pressure is reduced slightly, players without the tool should retain more defensive resources while the encounter remains threatening in both tool conditions.
• Intervention: In a test build only, change the final-phase damage multiplier from 1.00 to 0.90. Leave earlier phases, resource availability, tool behavior, and enemy composition unchanged.
• Test window: Collect at least eight comparable observations with the tool equipped and eight without it. Record completion, resources remaining, and reported pressure for each condition separately.
• Minimum intended change: In the tool-not-equipped condition, mean resources remaining should rise from the supplied baseline of 0.4 to at least 0.9.
• Acceptable pressure boundary: Mean reported pressure should remain between 3.5 and 4.2 without the tool and at or above 2.8 with the tool.
• Side-effect boundary: Earlier phases, resource availability, tool behavior, and enemy composition must remain unchanged. Any unintended change to those systems invalidates the test.
• Keep for further testing: The minimum resource-retention change is met, both pressure boundaries are respected, and no side-effect boundary is crossed.
• Revise: The observations are sufficient and the side-effect boundaries hold, but the intended resource change is smaller than 0.5 without crossing a rollback boundary.
• Rollback: Restore 1.00 if either pressure boundary is violated or an unintended system change is found.
• Inconclusive: Use this outcome if either condition has fewer than eight observations, required fields are missing, the conditions were not comparable, or the intervention was not isolated.

These thresholds are an exercise test plan, not a claim that 0.90 is a correct production value. A learner may propose different thresholds, but must define and justify them before seeing follow-up results.

1631. Scope of this guided build

No editable encounter fixture is supplied with this lesson. Therefore, the required assessment is a reviewable implementation and follow-up plan, not an unperformed playtest result.

If you have a compatible learner project with a repeatable encounter and an editable balance parameter, you may implement and run the test as optional supporting work. Name the project, build, parameter location, and repeatable setup. If you do not have that setup, do not represent original and proposed configuration values as observed before-and-after evidence. They are configuration records only.

1632. AI-native workflow

AI may help locate a parameter in a compatible project, check calculations, or review whether the plan preserves the relevant comparison. It must not select the final balance value or reinterpret thresholds after results are known.

A useful prompt is:

“Review this H-I-F-R decision card. Check whether the intervention changes only the named variable, whether the follow-up keeps tool-equipped and tool-not-equipped results separate, and whether every outcome has a preregistered threshold. Identify ambiguities, but do not choose the balance decision for me.”

Verify any AI-produced calculation against the source table.

1633. Guided practice

Complete the planned practical assessment linked to this lesson.

Step 1 — Analyze the evidence

Record:

• one observed difference between tool conditions;
• two relevant playtest observations;
• what the dataset measures;
• what it cannot establish;
• one competing explanation; and
• the intended challenge contract that the intervention should preserve.

Step 2 — Write the H-I-F-R card

Define one hypothesis and one bounded intervention. Record the original and proposed values, the systems that must remain unchanged, and the exact rollback action.

Step 3 — Preregister the follow-up test

Before inventing or inspecting any follow-up result, specify:

• the tool-equipped and tool-not-equipped comparison;
• the minimum number of observations per condition;
• the measures collected in both conditions;
• a minimum intended improvement;
• an acceptable pressure range or equivalent design-intent boundary;
• at least one side-effect boundary; and
• separate rules for keep, revise, rollback, and inconclusive outcomes.

Step 4 — Write the implementation plan

Identify the parameter or rule to change, how its original state will be recorded, how the change will be isolated, and how the original state will be restored. If no compatible project exists, describe these as required setup rather than claiming implementation.

Step 5 — Produce the module exit recommendation

Write a concise recommendation that combines:

• the quantitative pattern;
• relevant playtest observations;
• design intent;
• uncertainty and competing explanations;
• the proposed intervention; and
• the preregistered follow-up decision boundary.

This recommendation is the module exit artifact.

1634. Validation

The work is complete when another developer can determine:

• what the supplied evidence supports and does not support;
• which single variable would change;
• which comparison must be retained;
• how many observations would be required;
• what counts as improvement;
• what would be an unacceptable side effect;
• when to keep, revise, roll back, or declare the result inconclusive; and
• how the original state would be restored.

Observed results are required only if a compatible project and test setup were actually used. Missing observations must never be replaced with proposed configuration values.

1635. Key takeaways

• Data and playtest notes inform a decision but do not establish causation by themselves.
• A reversible balance decision needs preregistered thresholds, not only a general intention to review the result.
• Preserve relevant experimental conditions and report them separately.
• Insufficient or incomparable observations produce an inconclusive result.
• The module exit artifact is a recommendation combining evidence, design intent, uncertainty, and a follow-up decision boundary.

1636. Next lesson

Continue to 4.6 — Security and abuse.

1637. Knowledge check