Fingerprint
Your selected vector is transformed, resampled, normalized, and projected into compact band keys.
Method, minus the lab coat
We search for resemblance.
Not explanations.
A high correlation answers one narrow question: did two transformed vectors move together over their shared coordinates? It does not tell us why.
Retrieve
Those keys vote for a bounded candidate set. The browser never scans the whole corpus.
Rerank
Only candidate values are fetched. Exact correlation, lag, overlap, and fragility are computed locally.
Confess
The receipt records every degree of freedom used to make the match look impressive.
Why SimHash?
After mean-centering and L2 normalization, Pearson correlation is a dot product. Similar normalized shapes point in similar directions, so random-hyperplane signatures are a compact approximation for candidate retrieval.
Approximation only chooses what to inspect. The coefficient shown on a result is recomputed from overlapping observations.
Why the warning labels?
Search enough series, transforms, and lags and chance will eventually hand you something spectacular. A conventional single-test p-value ignores the assignment process.
Later corpus releases will compare results with randomized queries searched through the exact same pipeline.
What does “precalculated” mean?
Gallery records are selected offline from one immutable corpus release, then their exact coefficient, overlap, fragility, source trail, and search freedoms are written together. Opening one does not run a fresh search or choose a result on the server.
Can dinosaur bins match modern years?
No. Historical centuries, geological age bins, months, and modern years are separate coordinate schemas. The engine compares only like with like and keeps dating precision, interval uncertainty, preservation bias, and collection bias visible.