Benchmark
Whole-archive recovery vs archive size under a stress channel
Simulated
- Result
- 20 KB: 28 / 30 EXACT; 1 MiB: 2 / 30 EXACT; 0 false success at both sizes (archives recovered exactly)
- Classification
- Software-generated DNA strands passed through a software channel model. Not a laboratory result.
- Methodology
- Per-strand and per-row success measured; Wilson 95 % intervals (20 KB 0.79-0.98, 1 MiB 0.02-0.21). Larger sizes only by an analytic row model, labelled as extrapolation.
- Dataset
- Seeds 93000-93029; D13-F1 channel, mean coverage 10, NB dispersion 4, profile v4-balanced.
- Configuration
- Decoder E, 4 workers. A size is decoded only if a single decode is projected at ≤ 2 h on this host; 10 MB (projected 3.5 h) and larger were not decoded.
- Environment
- Development host: shared VPS, Intel Xeon Gold 6240 @ 2.60 GHz, 4 cores / 8 threads, Ubuntu 24.04, Python 3.12, CPU only. Host load 1.0-4.3 during the 1 MiB decodes.
- Limitations
- Per-row success ≈ 0.995 is enough for 9 rows but not for 411, so failures compound with size. The analytic model gives ≈ 0 at 100 MB in this setting: large archives under this channel need more coverage or parity. This is a known V9 limitation and a V10 work item.
- Version
- V9
- Source
- v9.0.0/experiments/v9/scale/results/noisy.jsonl
- Reproduce
experiments/v9/reproduce.sh scale· how to reproduce