Trophy case.
Twelve sealed Hardseal Edge evidence packets from a Jetson Orin Nano Super. Million-iteration soaks across multiple precision classes and power modes, plus an 8-hour sustained 25W run. Pinned-clocks replicate chain. Bundle SHA-256 below. Recompute the chain yourself — Python stdlib only, zero dependencies, short enough to read before you run it. Trust anchor lives on this page; the per-packet chain roots below are what you compare against.
4bb3343b336307a10ea0934128d6f46f098ddcd693493047c893b00ebd6cb7a1
3d0aee521f267b58d6d8f40a5dae56cace52f37aa06f7d4a8740878a0c1c1782
798a580ca212648cadf2be3e147d23e37251fca7886ff3180897d2c48fe0c13b
1597a26f43c1cd4615abffde095b2e3a439edece6eaafd094a39c1736c49f696
9ff39c3b2d5ae2f41a3db4d537c88d68f0e2d86b23bf2c29b46a7935b696e793
2c0723611e2ecf028dff656045b450df514d8eb129cfbdec1c463a81512022b4
8adec610c03ece2b62844790c0a6e06535218e6cb6db54564246d9f4fc3fbd67
1d9f06a194b5bf181e40a5efc15b06959f3b1525b7762edc46a56a5d6051473e
b73945d5…
ea0bd0a1…
de870b4e…
2a3acd73…
How to verify the packet chain.
Download the bundle.
It's 295 KB. The zip contains 12 packet.json files, the standalone verifier, and a manifest with all chain roots and the bundle SHA-256.
Verify the bundle SHA-256.
Open a terminal in the directory you downloaded to:
shasum -a 256 hardseal_edge_trophy_case.zip # macOS sha256sum hardseal_edge_trophy_case.zip # Linux
Output should match the bundle_sha256 value above. If it does not, do not run anything in the bundle. Re-download or contact us.
Unzip and verify the verifier itself.
The standalone verifier is intentionally readable. Hash it first against the verifier_sha256 above so you know it has not been tampered with inside the bundle, then open it and read it before you run it. The verifier is the entire trust surface for this bundle.
unzip hardseal_edge_trophy_case.zip cd trophy_case_bundle shasum -a 256 verify_standalone.py # compare to verifier_sha256 above less verify_standalone.py
Run the verifier on each packet — then compare the chain root.
Stdlib Python 3.8+. No pip install. No network calls.
python3 verify_standalone.py T1_thermal_saturation_yolov8n_int8_25W_1M_iter/packet.json
Expected output: a list of PASS: lines, ending with result: PASS. The verifier is Python stdlib only; shell examples shown are for macOS/Linux. Then compare the recomputed chain root against the published value in the specs block above. A PASS with a matching root means the packet is the one Hardseal published. A PASS with a different root means the packet was modified and re-hashed against a different chain root — reject it. Repeat for each of the twelve packets.
Or verify in your browser.
Open /verify.html, paste the contents of any packet.json, click Verify. The browser reproduces the same chain root the Python verifier produces; both paths are exercised by the same negative-test suite. Compare the output against the published root above. If the two paths ever produce different roots for the same packet, file a bug — we want to know.
Findings — yolov8x family.
The yolov8x INT8 evidence in this bundle covers two run classes on the same Jetson Orin Nano Super, same engine SHA-256, same TensorRT 10.3.0. Three MAXN_SUPER replicate runs at pinned clocks recorded 500,000 inferences each (1,500,000 total, ~8.23 wall-hours) at mean latency 19.733 ms, p99 32.368 ms, throughput ~50.64 FPS, p99/mean ratio 1.640. One 8-hour sustained run at 25W power mode recorded 2,000,000 inferences (~13.23 wall-hours, 200 blocks) at mean latency 23.812 ms, p99 39.235 ms, throughput ~41.98 FPS, p99/mean ratio 1.648, and stdev across the 200 block-means of 0.023 ms. The structurally similar p99/mean ratio across power modes (1.640 at MAXN_SUPER vs 1.648 at 25W) indicates the tail behavior is workload-bound, not power-bound — the same shape under both clock regimes. The chain roots above are what you recompute against; the verifier output is the only thing that should convince you these numbers are real.
What you should look for in the packets.
Read the limitations section first in any packet. The packet states explicitly what it does not certify. The benchmark numbers are real, the device fingerprint is real, the engine SHA-256 is a digest over the bytes that ran, and the chain root is independently recomputable — but the packet does not claim CMMC certification, FAA airworthiness, or any other regulatory outcome. That separation between integrity evidence and certification claim is the discipline.
And the trust anchor: even if a clever adversary mutates a packet and recomputes every internal hash, the resulting chain root will not match the value published on this page. That is why the chain roots above are out-of-band — the page is the trust anchor, not the verifier alone.
Use this for technical review.
Independent verification is the point of the artifact. Share it with your CISO, engineering lead, compliance owner, RPO, assessor, prime contractor, or internal red team. The packet is not a certification or readiness conclusion; it is a hash-chained, offline-verifiable artifact they can verify and interpret in context. A passing verification proves the packet is internally consistent and free of accidental change. The thing they cannot do is alter a packet and still have it match the chain root published out-of-band on this page — and we encourage them to try.