Trust, proven
Engine verification
Every headline number the Djehuti engine produces is checked against the free, industry-standard libraries on real public data. Below is a sample of the results and the speed. The complete log is available on request.
The speed
Same answers, a lot faster.
Measured on a 10-year-old AMD Ryzen 5 1600X (6 cores), pure AVX2, no cloud, no GPU. The engine returns the exact same numbers as the reference tools, in a fraction of the time.
faster than pandas on a real correlation matrix
177×177 Pearson, 2534 obs: 11 ms vs 137 ms, identical values
Monte Carlo simulations per second
50M N(μ,σ) draws in 234 ms, real DJIA parameters
correlations per second
full 177×177 matrix, same result as pandas
The matching
Checked against the tools your field already trusts.
A representative sample, one row per domain. Engine is the value Djehuti computed; Library is the value the reference tool computed on the same real data; diff is the absolute difference. Most agree to machine precision.
CHECK REFERENCE LIBRARY ENGINE LIBRARY diff RESULT Heston call (K100) QuantLib AnalyticHeston 10.39421857 10.39421857 5.0e-11 PASS SABR implied vol QuantLib sabrVolatility 0.02022475 0.02022475 0.0e+00 PASS Correlation sp~nq pandas / scipy 0.94957472 0.94957472 1.5e-13 PASS Tajima's D scikit-allel -2.53823864 -2.53823864 6.5e-14 PASS Fst (AFR vs EAS) scikit-allel 0.15620377 0.15620377 1.0e-12 PASS Gaussian process scikit-learn 19.41224870 19.41224870 7.1e-15 PASS Bayesian evidence scipy multivariate_t 1721.92097474 1721.92097474 4.0e-14 PASS HMM log-likelihood hmmlearn 8499.98751943 8499.98751937 6.4e-08 PASS EVT tail index scipy genpareto 0.31196273 0.31189952 6.3e-05 PASS Random forest acc scikit-learn RF 1.00000000 1.00000000 0.0e+00 PASS reference libraries: QuantLib, scikit-learn, scikit-allel, statsmodels, hmmlearn, scipy, numpy, pandas data: real public sources only (market prices, Treasuries, 1000 Genomes, Fisher iris), no synthetic data
This is a sample. The full suite runs hundreds of these assertions across every module and finishes with 0 failures, each built NaN-honest so a mismatch can never be hidden.
Want the whole thing?
The full verification is available on request.
The complete per-check log, the tamper-evident SHA-256 manifest (engine source, method, every input file, and results), and the reproduction harness so you can re-run it yourself. Ask and we send it.
Amnesiac by construction
Read-only, zero disk writes.
The engine maps your file read-only with mmap() and
MAP_PRIVATE | MAP_NORESERVE. It never copies the file, never fully
decompresses it, and writes nothing to disk: no temp files, no caches, no logs.
The strace below shows every file syscall for a full run.
# strace -e trace=openat,read,write,mmap,munmap ./djehuti data.npy openat(AT_FDCWD, "data.npy", O_RDONLY) = 3 mmap(NULL, 1048576, PROT_READ, MAP_PRIVATE|MAP_NORESERVE, 3, 0) = 0x7f8a4c000000 # ... all computation runs off the read-only mapping ... munmap(0x7f8a4c000000, 1048576) = 0 close(3) = 0 # No open() with O_WRONLY or O_RDWR. No write() to any file (only the terminal). # No temp files, no caches, no logs. Every run is independent and leaves no trace.