Status: evidence only. One cold run of the upstream small example; no source change. Nothing outside research/gpu/ is modified. Upstream source was the pinned commit 9e2bdd8ff5bf9c62f18263d7dac189b0400b260c, unchanged (logs/git-status-before.txt and logs/git-status-after.txt are empty).
examples/quickstart-gpu.sh, unmodified, on one rented H100 80GB (Vast.ai marketplace, on-demand). It builds the verifier and the default (serving) prover programs, then runsprover/examples/run_target_example.sh target_small_f2_roomy: synthetic ("made-up") weights, one commitment (-M 1), one request (own_choices.req), one proof, one verifier call.- No model weights were downloaded. No test seeds, no test build, no full-size config.
- Toolkit: the pip recipe from
prover/docs/build.md, line by line, into a fresh user directory. Resolved: cuda-toolkit 13.2.2, nvcc V13.2.86, cuBLAS 13.4.1.3 (logs/pip.log,logs/nvcc.txt). The image also shipped a CUDA under/usr/local/cuda; the recipe's nvcc was first on PATH and everyCompiled byline names it (3/3). - Everything ran as an unprivileged account. Compiler gcc/g++ 13.3.0.
| item | value |
|---|---|
| GPU | NVIDIA H100 80GB HBM3, compute capability 9.0, MIG disabled, 0 MiB used before start |
| driver | 615.71.09 (upstream tested 580; this is a newer driver, not the tested one) |
| CPU | Xeon Platinum 8480+, SHA-NI present; container cgroup cpu.max ≈ 26.9 cores |
| host memory | container memory limit ≈ 684 GiB |
| disk | 40 GB container overlay |
| host | one GPU of an 8-GPU marketplace host; the host was unverified in the provider's listing (no provider attestation of the machine was relied on) |
| cost | $1.186 all-in from the provider's final invoice (instance lifetime about 25.5 min, including the build). Correction: this row first said $1.026 (0.426 GPU-hours at $2.40/h plus $0.003 storage), read before billing had caught up; the final invoice is $1.186. |
| measurement | value | how measured |
|---|---|---|
| toolkit install | 11 s | stage timestamps (logs/stages.log) |
| verifier build | about 1 min | quickstart step log |
| prover default build | about 18.6 min (21:24:35 to 21:43:12 UTC) | quickstart step log, THE PROGRAMS ARE BUILT |
| whole quickstart | 20 min 04.6 s wall, 1611.8 s user CPU | /usr/bin/time -v (logs/whole-run.time) |
| cp_commit (made-up weights, M=1) | about 13 s | file timestamps (config copied 21:43:12, request copied after commit 21:43:25); phases in logs/cp_commit.log, key-file 8.1 s largest |
| cp_prove (one request) | about 3 s (±1 s) | request copy 21:43:25 to proof write 21:43:28, 1 s file-time resolution; not directly timed by upstream |
| peak GPU memory, whole run | 74,165 MiB sampled | nvidia-smi at 500 ms (logs/gpu-samples.csv); reached during cp_commit, not cp_prove |
| peak GPU memory, prove window | 5,293 MiB sampled | same samples, 21:43:26 to 21:43:30 |
| peak GPU power | 436 W sampled | same samples |
| peak host memory | 7.54 GB cgroup memory.peak (whole container, incl. builds); 4.6 GB max single-process RSS |
cgroup v2 / GNU time |
| proof size | 4,149,564 bytes | bundle/out/example.proof |
| key / leaf entry | 1,095,152 / 536,928 bytes | bundle/ivk/ |
| answer | token 284 | bundle/out/example.y |
| commitment ID | 5259d3dffde81fcf7b08634bfaca2bb217aa12286ef2aacd321648572660968c |
cp_commit output |
| verifier on GPU box | exit 0, YES |
logs/cp_verify.log |
Sampled GPU numbers are maxima of 500 ms samples, not true peaks. Commit/prove times come from second-resolution file timestamps around steps that the upstream script does not time individually, so treat them as ±1 s. n = 1, cold run; no repeat runs.
Coverage line from the verifier (exact-match policy): r=15 e=64 T=64 ... positions=15..63 tokens=16..63 y=yes next-token=no. Accounting line: 126.21 bits with proof of work counted, 96.23 without; the verifier itself states that no figure there is a security level.
The public bundle (config/, ivk/, in/example.req, out/example.y, out/example.proof) was copied off the GPU box before it was destroyed. A separate cp_verify was built on a different (CPU-only) machine from a separate clean checkout of the same pinned commit, with gcc 13.3.0. Its source digest matches the GPU box's verifier build (build=93afdaf5...), and the two binaries are byte-identical (sha256 7f9efc1e...04394, see logs/binaries.sha256 and local-verify/result.json).
| case | exit | last line |
|---|---|---|
| honest bundle, correct ID | 0 | YES (1.53 s on CPU) |
| one bit of the proof flipped | 1 | NO: proof-refused |
one bit of the answer (example.y) flipped |
1 | NO: reply-differs |
| one bit of the request (last byte) flipped | 1 | NO: statement-range |
| last byte of the proof cut off | 1 | NO: proof-refused |
| wrong commitment ID (last bit flipped) | 1 | NO: model-statement |
The originals were never modified; each negative case ran on its own copy. Five rejections of single, crafted mutations are a check that the verifier is wired up, not soundness testing.
A reviewer who did not run the GPU job rebuilt cp_verify from a fresh clone at the pinned commit (verifier/build.sh, gcc 13.3.0, 16-vCPU AMD EPYC-Rome VM, no GPU). Its sha256 is again 7f9efc1e86a848a3715c6cc870e465ac2a1c5e8fd1eac09e4809023de7604394, byte-identical to both binaries above. review/replicate.sh <cp_verify> <scratch dir> reruns everything below; its output is review/result.txt.
- Honest bundle, correct ID: exit 0,
YESon 7 of 7 runs. Wall time per run, including process start: median 1.580 s, mean 1.587 s, SD 0.026 s, range 1.549 to 1.620 s (n = 7). A second batch of 7 earlier the same hour gave median 1.545 s, SD 0.039 s. - Eleven corruptions, none of them one of the five above, each on its own copy: all exit 1 with
NO.
| case | last line |
|---|---|
| proof, bit 3 of byte 1,000,003 flipped | NO: proof-refused |
| proof, bit 0 of the byte 17 before the end flipped | NO: proof-refused |
| proof header, bit 7 of byte 9 flipped | NO: cp-entry-missing |
| proof with one zero byte appended | NO: proof-refused |
| request with one byte appended | NO: statement-required |
| request and answer files swapped | NO: statement-range |
| proof's first 32 bytes zeroed | NO: cp-proof-head |
| proof's 8-byte magic zeroed | NO: cp-proof-head |
| key, bit 4 of byte 500,000 flipped | NO: key-refused |
| leaf entry, bit 1 of byte 1,234 flipped | NO: entry-refused |
| answer replaced by token 285 (honest answer 284) | NO: reply-differs |
The originals matched SHA256SUMS before and after. As above, this checks that a verifier built independently accepts this one proof and rejects simple tampering. It is not soundness testing.
- Not the upstream headline numbers. This is a new run of the same small example, not a reproduction of full-size results.
- Not the full-size config, real weights, batching, or serving.
- The driver (615.71.09) is newer than the one upstream tested. The run passed, but that says nothing about driver compatibility in general.
- One run, so there is no variance estimate.
- The host was an unverified marketplace listing. Nothing here says anything about the host's integrity or about erasure of the rental's storage; the data was synthetic and the secret side is retired.
- The commitment's secret side (
isk/) stayed on the rental and was destroyed with it. It is not published. The commitment is retired and must not be reused.
replication-1/ is a second, independent GPU run of this same example: a new rental (verified listing, 1-GPU machine, driver 595.91.07), a new key, no shared remote state, and n = 3 timed repeats. Verdict: agree. Answer token, coverage and accounting are identical. Peak GPU memory was 74,123 MiB, cp_prove took 2.35 s on average (SD 0.13), proofs were 4.149 to 4.162 MB, and the host verifier said YES on 4 of 4 proofs. cp_commit was slower there (mean 18.8 s against about 13 s). Proof bytes differ because each cp_commit draws a fresh seed. See replication-1/README.md.
bundle/: the public files a client needs, enough to reruncp_verify config ivk in/example.req out/example.y out/example.proof -id <ID>.logs/: quickstart transcript, build logs and records, program logs, GPU samples, timing. Host-specific paths, the GPU UUID and the PCI bus id have been replaced with placeholders (<REPO>,<TOP>,<CUDA>,<HOME>,<pci-bus>).local-verify/: the independent verifier's log for each case, plusresult.json.review/: the review replication script and its output.SHA256SUMS: covers every file here except itself.