Make dataset requirements honest

[5aumit]

Outcome

SetScout does not call a requirement “hard” unless the application can actually enforce it with reliable source data. Researchers can clearly see which requirements are supported by evidence, not met, partially met, or still unknown.

Scope

• Replace public-facing “hard constraint” language with truthful requirement language.
• Keep requirement checks tied to retrieved evidence.
• Do not imply that every natural-language request was used as a deterministic search filter.
• Define which future requirements could be enforced only when reliable structured source data exists.
• Add focused tests for the chosen wording and result contract.

Out of scope

• Building a general constraint engine.
• Adding metadata extraction for every source.
• Changing the search algorithm based on intuition.

Blocked by

Nothing.

[5aumit]

Current-code references

• The internal search specification calls min_samples and access hard_constraints: setscout/models/schemas.py:17-31.
• The evaluator reads those values only to add them to the model prompt: setscout/graph/nodes/evaluator.py:12-32.
• Result checks already have explicit evidence-oriented statuses: setscout/models/schemas.py:104-114.

The ticket should make the public language match this current level of enforcement.