NNLS & Change-Point:
Prices are estimated using Non-Negative Least Squares to
ensure no negative values. Change-point detection
automatically splits data into periods when Cursor's pricing
changes.
Statistical Estimation:
We solve a system of linear equations where variables
represent token type costs and the result is the total Cost
from the CSV.
Confidence Intervals:
If data for a model is sparse or inconsistent (e.g., due to
CSV rounding), calculation accuracy decreases. In such cases,
intervals on charts will be wider, signaling potential margin
of error.
Caching:
The methodology separates costs for "cold" input, cache
writes, and cache reads, enabling you to see the real savings
from using Cursor Cache.