Definition
Proportion of sampled content blocks in AI-propagated submissions containing propagation errors: outdated references, incorrect version numbers, or superseded data carried forward because it pattern-matched.
The Measurement Gap
The NLP content propagation system reduces submission update cycle time by 70%. A manufacturing process change that took 3 writers 2 weeks now takes 2 days. But the system propagates text based on pattern matching and contextual similarity. When it misses, it carries forward outdated references that look right: a stability section citing "24-month data" when 36-month data is available, a manufacturing section with previous hold time parameters, a specifications section listing old impurity limits.
These errors are subtle: 95% of each content block is correct. The 5% that is outdated triggers agency queries, shifts submission timelines, and potentially delays product launches. Legacy metrics track query rate but do not connect queries to propagation errors as the root cause.
Decision enabled:
When SCPER exceeds 5%, mandatory expert verification of all NLP-propagated content blocks is required before submission. When SCPER exceeds 10% for a content type (e.g., CMC sections), that content type is excluded from automated propagation until the NLP model is retrained.
Formula
SCPER = blocks_with_propagation_errors / blocks_sampled
Where propagation errors include outdated references, incorrect version numbers, superseded data points, or other information carried forward without appropriate updating.
Worked Example:
A major submission update following a manufacturing site transfer. The NLP system prepopulated 187 content blocks across 4 regulatory submissions (FDA NDA supplement, EMA Type II variation, PMDA partial change, Health Canada SNDS).
Expert review found:
169 blocks fully current and accurate.
11 blocks with outdated stability data references (cited 18-month data when 24-month available).
4 blocks referencing old manufacturing site environmental classification.
2 blocks carrying previous process validation protocol number.
1 block citing superseded analytical method.
SCPER = 18 / 187 = 9.6%
Segmented by section: CMC content had 14.3% SCPER (changes most frequently, most version-sensitive references). Clinical had 2.1%. Nonclinical had 0%. The 11 stability blocks will likely trigger agency queries: each costing 40–80 hours of response time.
Sign Convention:
Lower is better. Zero means perfect propagation accuracy. Above 5% requires mandatory expert verification. Above 10% for any section type triggers NLP model retraining.
Proposed formula, pending validation pilots.