EvidenceLab

Terminology guide

What the labels mean.

Use this lexicon while reviewing a decision. Definitions describe how the current beta uses each term.

Where variables come from

The synthetic retention example starts with predefined variables, including onboarding, outreach and support headcount. In your own investigation, variables can be entered manually, imported from JSON, transferred from the guided outline, or proposed through optional AI drafting. Review every proposed variable and its settings.

The map displays the variables and relationships currently recorded in the investigation. It does not automatically discover them from your objective. This beta does not record the origin of every individual variable, so editing an example or importing a file can mix sources.

Variables and control

Variable
A factor or outcome recorded in this investigation, such as onboarding workflow or customer satisfaction. Its presence in the map does not prove it affects the objective.
Explore from
Choose an existing variable as the focus of the map. The list is the current variable inventory; choosing a focus does not generate variables or predict outcomes.
Influencing
A variable you have classified as potentially influencing the objective. This is a role in the investigation, not an established causal claim.
Controllable
Marked as technically changeable by the organization in this investigation. Being able to change it does not automatically mean you are permitted to change it. The controllable group records its intended role; the technical-control checkbox is a separate setting.
External
Not marked as directly controllable in this investigation. It can still be inside the organization and may change indirectly. For example, outreach can affect support workload even when workload is not a direct control.
Affected
An outcome that may change downstream of an action. A variable can belong to more than one group: influencing, controllable and affected.
Permitted
Permission to change this variable is enabled for the current investigation. Technical control, fixed status and other constraints must still be checked.
Not permitted
Permission to change this variable is not enabled. The engine blocks actions that list it as directly changed. This setting does not prevent indirect effects; it also does not independently verify an actual company policy.
Fixed
Must remain unchanged as a direct decision variable during this investigation. Fixed status blocks an action that directly changes it, even if permission and technical control are enabled.
Variable ID
The stable identifier used to connect variables, relationships and actions. For example, outreach is the ID of Customer outreach. Reusing one variable in several branches does not create several different variables.

Relationships and evidence

Relationship
A recorded, directed connection from one variable to another. Incoming branches show what may influence the selected variable; outgoing branches show what it may affect. Arrows alone do not establish causation.
Positive / negative effect
Positive means the target variable may increase; negative means it may decrease. These labels describe direction, not whether the change is desirable. Unknown or mixed directions can also be recorded.
Hypothesis
A proposed explanation or effect that needs testing. A hypothesis by itself does not supply a supported outcome estimate for ranking actions.
Assumption
An input accepted for a what-if scenario. To use it for conditional comparison, link it to an approved scenario evidence record and enable scenario assumptions in Workspace settings. Approval does not turn an assumption into a measured fact.
Observed
Recorded as measured or directly observed, with a supplied source. An observation does not by itself establish what an intervention will cause.
Inference
A conclusion drawn from supplied evidence. Its strength depends on the evidence and reasoning; the label itself does not verify either.
Unknown evidence
A value, source or justification has not been established. Unknown is not zero, and it is not evidence that an effect is absent.
Evidence register
The record of claims, sources, source kinds and evidence status. Action estimates refer to these record IDs. Labels are supplied by users and are not independently verified by EvidenceLab.
Intervention evidence
Evidence about what happens when an action is introduced. For non-scenario effect checks, the current engine requires a referenced source marked as an intervention; simply marking an estimate observed is insufficient. You must judge whether the study actually supports the claim.
Outcome estimate / range
An entered expectation for an outcome, optionally with low and high bounds. Example values are fabricated assumptions. Ranges are scenario bounds unless empirically justified; they are not automatically confidence intervals or probabilities of success.
Shared variable / feedback loop
A shared variable appears in multiple paths. A feedback loop returns to an earlier variable in a path. The tree shows references to keep these connections finite; it does not simulate the loop over time.

Objectives and limits

Objective
The primary outcome you want to improve, its success metric and whether to increase or decrease it.
Baseline / target
Baseline is the starting value. Target is the desired value. A feasible action can still fall short of the target.
Time horizon
The time window stated for the objective. The current engine checks whether entered implementation duration exceeds it. You must separately check outcome-measurement timing, such as waiting for a day-90 retention cohort to mature.
Hard constraint
A structured rule an action must satisfy, such as cost <= 200000. A supported violation makes an action infeasible; missing supporting values or evidence leave the check unknown.
Guardrail
A protected outcome or resource limit, such as satisfaction change >= 0 or added support workload <= 8 hours per week. Like a hard constraint, a supported violation blocks the action.
Soft preference
A desirable condition, such as delivery within 30 days. A miss is reported but does not by itself make an action infeasible. The current engine does not use soft preferences to calculate a weighted score or break a ranking tie.
Metric identifier
The key linking a rule to an action estimate. cost and days are built in; other rules must match an action effect key, such as satisfactionChange. A written resource note alone does not become an enforced rule.
Percentage points
The difference between two percentages. Moving from 85% retention to 89% is an increase of 4 percentage points; 89% remains 1 percentage point below a 90% target.

Actions and results

Action / intervention
A candidate change you enter and evaluate. Record directly changed variables separately from outcomes it may affect, plus costs, duration, estimates and supporting evidence.
Feasible
No blocking violation or required evidence gap was found under the entered rules and accepted evidence. It does not mean the action will work in practice or meet the objective target.
Infeasible
At least one blocking check fails, such as a fixed variable, missing permission or a supported budget or guardrail violation. A better outcome estimate does not override that failure.
Unknown feasibility
No blocking violation has been established, but a required value or supporting evidence is missing. It is not confirmed feasible. A conditional result may still point to the option while showing what needs checking.
Conditional recommendation
A candidate to validate under explicitly accepted assumptions, unresolved feasibility or unranked alternatives. It is a direction for investigation rather than a promise of success.
Best-supported decision
The leading action among the alternatives evaluated under the current inputs and decision rule. It does not establish a global optimum or independent scientific validation.
Evidence gap
The system cannot support a reliable selection with the current inputs. Possible reasons include incomplete inputs, overlapping ranges, contradictory rules or no eligible action. Inspect the individual results to identify the cause.
Trade-off / opportunity cost
A cost, side effect or benefit forgone by choosing an action. The engine enforces numeric trade-offs only when they are entered as matching structured constraints and action effects.
Validation / reversibility
Validation describes what to test or measure next. Reversibility describes whether you can undo or limit the action. Writing a plan does not mean the test has been run.

Open the decision workspace · Return to guided start