Contextual Default Retraction Dataset
Exploring how AI models reason and generalize across feature combinations and contexts.
Bird transition example · dry → wet → dry conditions.
Illustrative examples of how context changes an answer. These examples are separate from the growing catalog and random label templates.
Features
Contexts
Its feathers are dry throughout the attempt. Can this bird take off from level ground and fly 20 metres continuously without assistance?
* Assistant responses are illustrative and do not represent the actual dataset, factual truths, or normative judgments.
Features
Contexts
This person intentionally damaged someone else’s property without consent. There is no imminent serious danger or official duty, so no exception applies. Under the supplied hypothetical rule, is this person guilty?
* Assistant responses are illustrative and do not represent the actual dataset, factual truths, or normative judgments.
Features
Contexts
At a temperature suitable for its species, with appropriate food and hydration supplied, can this mammal survive 7 days in a safe enclosure? Keep its age, health, and energy reserves unchanged.
* Assistant responses are illustrative and do not represent the actual dataset, factual truths, or normative judgments.
Features
Contexts
The plant stays in a 10-litre pot with suitable well-drained soil, no pests, and no external water reservoir. Without watering or rain, can it survive 3 days at 20°C and 50% humidity, with 12 hours of bright indirect light daily?
* Assistant responses are illustrative and do not represent the actual dataset, factual truths, or normative judgments.
Features
Contexts
On a dry, level track at 20°C, can this person run 100 metres continuously at 8 km/h without walking breaks or external assistance? Keep their training, mobility, and fatigue unchanged.
* Assistant responses are illustrative and do not represent the actual dataset, factual truths, or normative judgments.
| Version | # Domains | # Queries / Domain | # Contexts / Domain | # Entities / Domain | # Features / Entity | Total |
|---|---|---|---|---|---|---|
| v1 |
v1 contains 5 contexts per domain; the examples above show 4. Feature counts refer to the 7 experimental judgments per entity.