MedGaugeCapabilities / C · Data and program
C.3 Data augmentation Active
Given a labelled item, the agent produces variants (disease line, scale, defect type, reordering) that preserve the label. Graded by confirming that the reference solution still scores full marks on every variant.
Evidence: none. No benchmark card lists this capability yet. No level is issued: methodology 0.1.0 sets no level thresholds.
- Domain
- C · Data and program
- Evidence form
- program
- Veto constraints
- —
- Depends on
- C.1 Data processing
- Required by
- E.5
- Internal task families
- —
- External anchors
- —
- Version
- 0.1.0 · 0.1.0
Definition in detail
Variants of a labelled item extend a benchmark without new expert labelling, and test whether a result depends on surface features. They are only valid if the correct answer is unchanged.
Inputs a task gives
- A labelled item with its reference solution and grader
- The list of allowed variant operators, such as disease line, scale, defect type and reordering
- Constraints stating what must stay invariant
Artifact required
A set of variant items, each recording the operator applied, the fields changed and any reference update the operator requires.
How it is graded
A program runs the reference solution on every variant through the grader; it must score full marks on each. Variants must differ from the original in the way the operator states, and must use only allowed operators.
In scope
- Applying allowed operators while keeping the label correct
- Recording what each variant changed
Out of scope
- Generating new records from rules - graded under C.2
- Auditing the grader itself - graded under C.7
Typical failures
Expected failure patterns; not measured rates.
- Reordering material in a way that changes a temporal relationship the label depends on
- Variants that differ only trivially from the original
- Leaking the expected answer into the variant text
Benchmarks
No benchmark card lists this capability.
Grading references
No grading reference is mapped to this capability.