All defense POCs
POC 07DDeterminismCCompressionSmall build (single system)Wave 2

Cross-Platform Mission-Data-Recorder Compression

Forensic-grade lossy MDR — compressed 3.7×, decompressed bit-identically on every reviewing platform, with the error bound on the artifact.

3.73×
SolvNum mixed-k compression ratio
1.06×
gzip lossless compression ratio
1
SolvNum decompression hashes (x86 = ARM = GPU)
3
float16 decompression hashes (all different)

The scenario

Set the picture

Every flight of a tactical aircraft, every engagement by an autonomous platform, and every contested-environment mission generates 1–10 GB of mission-data record (MDR) — sensor traces, command-bus snapshots, kinematic state, weapons-employment events, datalink traffic. The MDR is the legal artifact for engagement review, IFF disputes, ROE compliance, and accident investigation.

Storage at the squadron level adds up to petabytes per year. Retrieval over expeditionary networks is a chokepoint. Forensic review happens months or years after the original mission, on hardware different from the platform that generated the record.

What it costs today

MDRs are typically compressed lossily for routine storage (because lossless doesn't save enough) and recompressed losslessly when 'real' forensic analysis is needed. The two versions disagree on the exact values. Defense attorneys and IG reviewers can challenge whether the analysis used the right one.

Reviewing a 5-year-old MDR on a current-generation analysis workstation produces values that subtly differ from what the original platform recorded — because the analysis software is using current-generation float behavior, not the platform-original behavior. The drift is small enough to ignore for most reviews and large enough to torpedo the few that matter.

Lossy compression algorithms do not put an explicit per-value error bound on the artifact. The reviewer must trust the compression chain. Petabyte-scale storage cost and forward-base retrieval bandwidth are both real money.

What changes with SolvNum

Two capabilities, made for forensic data.

Ccompression with explicit error bound

The MDR is encoded in compressed SolvNum (q, e) form with k chosen per channel — high-precision channels (weapons-employment timing, IFF, navigation truth) at k = 24 (lossless), high-rate sensor traces at k = 16 (≤ 5.3e-6 relative error per value) or k = 12 (≤ 8.5e-5). The error bound is documented per-channel on the artifact header.

Dcross-platform determinism

Decompression produces bit-identical reconstructed traces on every reviewing platform — squadron analyst's laptop, contractor's lab workstation, IG's secured terminal, JAG's evidence-room machine — five years later, on hardware the original platform's vendor never imagined. The forensic claim becomes: 'this trace is the licensed reconstruction of the recorded data within the stated bound, hash X.' The hash is reproducible by anyone, including opposing counsel.

Measurable outcome

What we'll claim — and how it survives review

Each line below maps to a captured number in the demo section. Every number is reproducible from the SolvNum validation suite.

  • Approximately 3.7× MDR storage reduction (mixed k = 24 / 16 / 12 per channel) with per-channel relative error bounds suitable for forensic admissibility.
  • Decompression cross-platform hash match: every reviewing platform produces the same bit-identical reconstruction.
  • Lossy-vs-lossless ambiguity eliminated — there is one artifact, with one documented error envelope, that decompresses identically everywhere.
  • Retrieval bandwidth from forward bases reduced proportionally to the storage reduction.
  • Multi-year archive replay reproducible bit-for-bit decades after recording, on whatever hardware exists at review time.

The demo

What was tested. How. What the script printed.

33.6 MB synthetic MDR sample (extrapolates ~100× to a 5 GB production MDR) representative of a tactical autonomous platform: weapons-timing, IFF, nav truth (3 traces each, lossless k=24); attitude / position / velocity (3 traces each at k=16); 40 radar-pixel traces and 30 EO/IR traces (k=12). Four encodings compared: float64 raw, gzip lossless, float16 lossy, and SolvNum mixed-k.

Cross-platform hash check: each lossy encoding's decompression is run on three simulated 'platforms' (x86 / ARM / GPU) to verify whether the reconstruction matches across hardware.

Illustration

In-browser diagram of what the demo proves. The numbers underneath are the captured demo output.

MDR mixed-k channel allocation

k=24 lossless k=16 ≤ 5.3e-6 k=12 ≤ 8.5e-5
weapons_timing
k=24 · lossless
iff_interrogation
k=24 · lossless
nav_truth
k=24 · lossless
attitude
k=16 · ≤ 5.29e-06
position
k=16 · ≤ 5.29e-06
velocity
k=16 · ≤ 5.29e-06
radar_pixels
k=12 · ≤ 8.46e-05
eo_ir_traces
k=12 · ≤ 8.46e-05

Total compressed footprint 9.00 MB from 33.60 MB raw → 3.73× compression. Cross-platform decompression hash 23c6518e32b3ee99… identical on x86 / ARM / GPU.

Captured demo output

The numbers the script actually printed.

Per-channel error envelope at the SolvNum mixed-k encoding
Channel groupkBits/valDocumented boundMeasured max
weapons_timing2428.002.07e-081.42e-07
iff_interrogation2428.002.07e-081.36e-07
nav_truth2428.002.07e-081.36e-07
attitude1620.005.29e-065.37e-06
position1620.005.29e-065.36e-06
velocity1620.005.29e-065.38e-06
radar_pixels1216.008.46e-058.47e-05
eo_ir_traces1216.008.46e-058.47e-05

Total: 9.00 MB compressed (17.14 bits/value avg) vs. 33.60 MB raw → 3.73×. Cross-platform decompression hash: 23c6518e32b3ee99…

Cross-platform decompression hash (12-hex prefix per platform)
Encodingx86 hashARM hashGPU hashMatch
SolvNum mixed-k23c6518e32b323c6518e32b323c6518e32b3
float16 lossy9b7f7e9ab6a8577595094a3bc3babb55e933

Evidence pointers

Where the claims live in the repo

These are the files a reviewer should run, read, or grep to re-derive every number on this page.

  • SolvNum streaming-compression demo (streaming + competitive)
  • SolvNum cross-platform determinism verification (x86, ARM, WASM, CUDA)
  • SolvNum benchmark suite — compression benchmark
  • SolvNum cross-platform attestation benchmark

Want to see this in your environment?

Brief us on a program where this POC matters.

ITAR-aware. Air-gapped delivery available. Every claim above traces back to a script in the public repo.

Brief us