Real-Data Realism Gate¶
IINTS now separates two questions:
- Plausibility: does a trace look physiologically plausible?
- Evidence readiness: is it strong enough for public results, local-AI training, or EUCYS-style claims?
The normal realism validator still returns likely_realistic, needs_review, or likely_unrealistic. The stricter gate in iints.data.realism_governance adds a second layer for research evidence.
Strict Gate¶
Use review_real_data_realism(report) after validate_realism_dataset(...).
The strict profile blocks traces when:
- the base verdict is not
likely_realistic; - any realism check failed;
- the trace is shorter than 20 hours for daily physiology claims;
- impossible values, rapid sensor jumps, or long gaps are present;
- time below 54 mg/dL is too high for a normal-reference trace;
- meals are present but insulin annotations were lost;
- no empirical reference profile was used;
- the reference envelope fails.
Data Evidence Strategy¶
The SDK registry now ranks real-data sources by evidence use:
tier_1_calibration_reference: strongest current candidates for simulator calibration and realism envelopes.tier_2_external_training_candidate: useful after import, MDMP certification, and external validation.tier_3_controlled_validation_candidate: strong independent validation once access is approved.demo_only: good for quickstarts, not final realism claims.
In code:
from iints.data.evidence import rank_real_data_sources
for source in rank_real_data_sources():
print(source["id"], source["tier"], source["use_case"])
Important Boundary¶
Passing the strict gate does not mean the SDK is clinically validated. It means the trace is better suited for research evidence than a basic synthetic demo trace.