Why a Boundary Record Matters
A model boundary is not fully communicated by a diagram or a list of state variables. A reviewer also needs to know why particular processes are represented, what has been left outside the model, which quantities enter from outside, and what the model is intended to support. Without that information, a modeling choice can be mistaken for a biological claim, and a later model change can be difficult to interpret or reproduce.
A boundary record is a structured account of those choices. It does not replace the model equations, evidence review, or decision analysis. Its purpose is to connect them: a reader should be able to follow a boundary choice from the decision question, through its rationale and evidence, to the model implementation and any review of its consequences.
The record documents the boundary rather than deciding it from scratch. Earlier material on decision questions, mechanistic boundaries, external forcing inputs, and evidence-informed complexity supplies the reasoning for those choices. Here the emphasis is on making that reasoning explicit and preserving it as the model develops.
What to Record
Use specific, testable descriptions rather than broad labels such as “immune system included” or “other biology omitted.” For every entry, identify the model element or process, its status, the reason for that status, and the consequences for interpretation. Where possible, point to a source or analysis using a stable identifier, such as a document title and section, an experiment code, or a model version.
| Record field | What it should make clear |
|---|---|
| Intended use and limits | Which decision or prediction the model supports, in what population and context, and what it is not intended to support. |
| Included processes | Which biological processes are represented, at what level, and which model states or outputs represent them. |
| Excluded mechanisms | What is outside the model, why it was omitted, and whether its effect is believed to be negligible, captured indirectly, or unresolved. |
| External inputs | Which quantities are supplied to the model rather than represented as evolving internal states, including their source, units, timing, and allowed range. |
| Assumptions and rationale | What is taken as fixed or simplified, why the choice is reasonable for the intended use, and which evidence or judgment supports it. |
| Change history | What changed, when, why, who reviewed it, and what implementation or decision-relevant consequences were checked. |
Keep three categories distinct. An assumption is a condition adopted to define or simplify the representation, such as treating a baseline drive as constant over a specified time window. An exclusion is a mechanism not represented within the model boundary. An external input is a quantity supplied to the model from outside, rather than generated by its internal dynamics. A quantity can be an input in one model and an internal state in another; the record should describe its role in this model, not imply that it is inherently one or the other.
An exclusion is not automatically a claim that the mechanism does not exist. Record whether it was excluded because evidence suggests little relevance to the decision, because it is not identifiable from available evidence, because its effect is represented by a simpler component, or because the question is outside the intended use. If the reason is uncertain, say so. A documented uncertainty is more informative than a confident but unsupported rationale.
Worked Example: Recording a Small Mechanistic Boundary
Consider a hypothetical model intended to compare two intervention schedules for their predicted effect on a tissue biomarker over 14 days. The team has chosen to represent drug exposure, receptor signaling, cell activation, and biomarker production and clearance. The following entries show how to turn those choices into an auditable record.
| Entry | Record |
|---|---|
| Intended use | Compare the two specified schedules for predicted tissue biomarker response over days 0–14 in the defined study population. Not intended to predict long-term clinical outcomes. |
| Included processes | Drug exposure, receptor signaling, cell activation, and biomarker production and clearance. Each is linked to its model component and output in model version 0.3. |
| Excluded mechanism | Adaptive immune-cell recruitment is not represented as a separate process. Available evidence review E-04 did not establish its contribution to the 14-day biomarker comparison; its relevance remains uncertain. |
| External input | Baseline inflammatory drive is supplied as a fixed input for each simulation. Its source, units, and value are recorded with the scenario data; it is not estimated as a changing internal state. |
| Assumption | Baseline inflammatory drive remains constant over the 14-day simulation. This is a simplification for the defined comparison, not a claim that the drive is biologically constant in general. |
The use statement constrains interpretation: a prediction about the 14-day biomarker is not automatically a prediction about a clinical outcome or a later time horizon. The exclusion entry also avoids an unsupported inference. It says that recruitment is not separately represented and explains the decision context; it does not say recruitment is absent. Finally, the input and assumption are separated: the drive is supplied from outside, and its constancy during the simulation is an assumption about how that input is used.
Make the Rationale Traceable
A useful record has a chain of traceability: intended use → boundary choice → rationale and evidence → model implementation → consequence for interpretation. If any link is missing, a reviewer may be unable to tell whether a choice followed from evidence, convenience, or an undocumented judgment. This chain is also a practical test of completeness: each important boundary choice should connect to the model and to the scope of conclusions that can be drawn.
For example, “mechanism omitted because unnecessary” is not an auditable rationale. A stronger entry names the mechanism, identifies the decision context in which it was judged unnecessary, points to the supporting analysis or evidence, and records the remaining uncertainty. If there is no direct evidence, distinguish that absence from evidence of no effect. The record should not make the rationale appear stronger than it is.
A concise register is often easier to review than long narrative notes. Use stable identifiers for evidence and model versions, define abbreviations, and keep a controlled change history. A diagram can help show where processes and interfaces sit, but it rarely communicates the evidence, uncertainty, or reason behind an exclusion by itself. Conversely, a detailed narrative without links to model components may be hard to verify. Use the record and the diagram for complementary purposes.
Worked Example: Documenting a Boundary Change
In a hypothetical review of the biomarker model, new evidence suggests that treatment-related feedback may reduce the inflammatory drive during the 14-day window. In version 0.3, that drive was an external fixed input. The team evaluates a revised representation in version 0.4 in which feedback changes the drive internally. For a defined illustrative scenario, the predicted biomarker reduction relative to baseline is 40% under version 0.3 and 31% under version 0.4.
Change entry. “Version 0.4: represent treatment-related feedback on inflammatory drive as an internal evolving process; previously the drive was supplied as a fixed input. Reason: evidence review E-07 identified a plausible feedback effect within the intended 14-day window. The earlier fixed-input assumption is no longer used in this version.”
Numerical impact. The predicted reduction changes by \(31\%-40\%=-9\) percentage points. Thus, in this scenario, the revised model predicts a reduction 9 percentage points smaller than the earlier version. This difference is a model-version comparison, not evidence that the biological feedback caused a particular clinical outcome.
What to verify. Record which component and input definition changed, whether the same intervention schedule and other scenario settings were used in both runs, and whether the new implementation behaves as intended. Also note whether the changed prediction affects the intended comparison. If the decision conclusion changes, that is material impact; if it does not, the numerical difference still belongs in the change record. The reason for the change, the evidence identifier, and the comparison result should remain linked to versions 0.3 and 0.4.
Reviewing and Updating the Record
A boundary record is useful only if it stays aligned with the implemented model. Review it when a mechanism is added or removed, when an external input becomes an internal state or vice versa, when the intended use changes, or when new evidence alters the rationale for an existing choice. Do not silently overwrite an earlier rationale: preserve the previous version and state what changed.
Name the affected process, assumption, exclusion, input, or use statement, and identify the model version in which the change is proposed.
Link the change to a decision need, new evidence, or a documented limitation. Describe uncertainty rather than presenting a tentative rationale as established fact.
Verify that the model matches the record. Compare relevant outputs under appropriately matched conditions and record the effect on interpretation or the intended decision.
Assign a version or date, retain the prior entry, and record review responsibility. Update linked diagrams or specifications when they no longer match.
This process assumes that model versions and evidence sources can be identified, that an entry can be linked to the implementation, and that the intended use is stated precisely enough to judge relevance. If those conditions are not met, first record the gap and its consequence. For example, if a source is unavailable, mark the rationale as unsupported or pending rather than supplying a guessed citation.
Change review should be proportionate to the choice. A wording clarification that does not alter the model may need only a brief note. A new mechanism or a changed input-state boundary can alter predictions and therefore warrants an implementation check and a comparison relevant to the intended use. The point is not to require identical work for every edit; it is to make the basis for the review visible.
Worked Example: Auditing an External Input Range
Suppose a hypothetical model accepts an external inflammatory-drive value of 1.0 arbitrary units in its reference scenario. The record initially gives no permitted range. During review, the team defines two additional scenarios, 0.8 and 1.2 units, and compares the predicted biomarker reduction: 34%, 30%, and 26%, respectively.
Range and arithmetic. The tested inputs span \(1.2-0.8=0.4\) units. Across those scenarios, the predicted reduction spans \(34\%-26\%=8\) percentage points. Relative to the reference prediction of 30%, the lower-input scenario differs by \(34\%-30\%=4\) percentage points, and the higher-input scenario differs by \(26\%-30\%=-4\) percentage points.
Record update. The input entry should name the source and units, give the tested range of 0.8–1.2 arbitrary units, and link the three predictions to the relevant model version and scenario settings. It should state that the predicted reduction varies by 8 percentage points across this tested range. This documents what was tested; it does not establish that the range contains every plausible biological value or that the model is valid outside the tested range.
This example illustrates why an input’s allowed range is part of the boundary documentation. A value without source, units, or scenario context is difficult to reproduce, while a tested range without a statement of its limits can be mistaken for a validated operating domain.
Common Failure Modes and Limits
Several documentation practices make a model harder to audit even when the underlying modeling work is careful:
- Listing omissions without reasons. A list does not explain whether a mechanism was judged irrelevant, was unresolved, or was outside the intended use. State the reason and its uncertainty.
- Treating absence of evidence as evidence of absence. If no relevant evidence was found, record that limitation; do not claim the mechanism has no effect.
- Using vague scope statements. “For biomarker prediction” does not specify a population, time horizon, or decision. State the context and the limits on interpretation.
- Leaving assumptions implicit. If a quantity is held fixed, say so, identify its role, and explain the context in which that simplification is used.
- Overwriting the change history. Replacing the old entry hides how and why the model changed. Preserve prior versions and link changes to their consequences.
- Documenting a diagram instead of the rationale. A diagram may show an omitted process but cannot, by itself, explain the evidence, uncertainty, or intended-use reasoning behind the omission.
Documentation cannot establish biological truth, compensate for weak evidence, or guarantee that the model is adequate for every use. A complete-looking record can still contain a poor assumption. Its contribution is narrower and important: it lets a reviewer identify the assumption, assess its support, understand the model’s stated limits, and see whether a change was evaluated consistently. A record should therefore distinguish what is implemented, what is assumed, what remains uncertain, and what is not claimed.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A model supplies a measured baseline signal as a fixed value. What two distinct entries should document its role and how it is treated over time?
- A mechanism is omitted because the team has not found evidence about its effect. Write a suitable rationale that does not overstate what is known.
- A revision changes a predicted response from 52% to 45%. What is the change in percentage points, and what information should accompany that comparison in the record?
- Why should an intended-use statement name a population and time horizon rather than merely say “predict response”?
- A diagram shows an excluded process but gives no reason for its omission. What information is missing from the boundary record?
Solutions. 1. Record the signal as an external input, including its source and units; separately record the assumption that it is held fixed over the specified simulation period. 2. State that the mechanism is not represented because its decision-relevant effect is unresolved or unsupported by available evidence, identify the evidence review performed if applicable, and note the remaining uncertainty. Do not claim the mechanism has no effect. 3. \(45\%-52\%=-7\) percentage points, so the revised prediction is 7 percentage points lower. Record the model versions, matched scenario conditions, changed component, reason and evidence for the revision, and whether the difference affects interpretation or the intended decision. 4. Population and time horizon define the context in which the prediction is intended to apply; without them, users may extend the conclusion to settings the model was not designed to address. 5. The record needs the reason for exclusion, its evidence or uncertainty, its relevance to the intended use, and any limitation on interpreting model outputs that follows from leaving the process out.