Reproducibility
Version 9.0.0 (git tag v9.0.0). Repository: github.com/vishnuselvam101-hash/VSnexustechnologies-.
Environment we used. Linux x86-64 (Ubuntu 24.04), Python 3.12, a C11 compiler (gcc or clang) for the native kernels. Results were produced on a shared 4-core / 8-thread Intel Xeon Gold 6240 VPS; timing results will differ on other hardware, deterministic results must not.
Install
git clone https://github.com/vishnuselvam101-hash/VSnexustechnologies-.git vnx-dna && cd vnx-dna
git checkout v9.0.0
python3 -m venv .venv && . .venv/bin/activate
pip install '.[dev]' # builds the native kernels when a compiler works
python -m vnxdna.native --require-native # exits 1 unless every kernel is native
Tests: expected: 3,337 passed, 0 failed, 6 skipped (the count can differ on hosts without a C compiler).
pytest -q -n 6
Spot reproduction: re-runs a pre-specified subset (oracle EVAL seeds 91000-91001, coverage EVAL seed 91000, noisy 20 KB seeds 93000-93002) and compares every deterministic field with the committed results. Expected last line: REPRODUCED (29 rows, 0 differences).
experiments/v9/reproduce.sh spot
Full runs (reproduce.sh decoder | coverage | scale | channel) take hours; channel also needs the public D13 FIT files described in experiments/v9/d13/README.md.
What is compared. Timings (*_seconds), host load and environment fields are excluded; outcomes, hashes, counts and statistics must match exactly.
Benchmarks and their reproduction commands
| Benchmark | Class | Result | Version |
|---|---|---|---|
| Clean-channel encode and decode, 20 KB to 1 GiB | Verified (software) | 1 GiB: encode 105.8 s, decode 195.4 s, byte-exact; 9.78 nt per byte | V9 |
| Consensus V2 selection: candidate E vs the V8 decoder | Simulated | EXACT 283 / 600 vs 240 / 600; pooled ORE 0.575 vs 0.488; Holm-adjusted exact McNemar p = 5.2e-12; 0 false success | V9 |
| Native C11 consensus-polish kernel | Simulated | 60 / 60 decodes byte-identical to the NumPy reference; median speed-up 2.46× (bootstrap 95 % 2.41-2.51) | V9 |
| Adaptive computational coverage (stopping rule) | Simulated | 200 / 200 EXACT at the stopping point; 0 false terminations; 41.7 % (v4-balanced) and 31.2 % (v7-lowcov) fewer reads than the smallest fixed coverage with equal EXACT | V9 |
| Whole-archive recovery vs archive size under a stress channel | Simulated | 20 KB: 28 / 30 EXACT; 1 MiB: 2 / 30 EXACT; 0 false success at both sizes | V9 |
| Public-data nanopore channel model G1 (adequacy test) | Public data | INADEQUATE: fails the pre-registered FIT-only pre-check on M6r and M8 | V9 |
| V9 release verification | Verified (software) | 3,337 tests passed, 0 failed, 6 skipped; sanitizers PASS; 0 fuzz crashes; 29 / 29 reproduction rows | V9 |
| VNX-Secure simulated attack scenarios | Simulated | 10 / 10 included scenarios detected and contained; 0 findings over 270 benign requests; median detection 4.2 ms, recovery 12.1 ms | V9 |
| Outer code under i.i.d. strand loss (V6 product code) | Simulated | 20 / 20 seeds decoded up to 16 % strand loss (V5 code: 7 %), at redundant-strand overhead 0.249 vs 0.251 | V6 |