US Tech Automations

Replay last year's model answer, byte for byte

Model-risk and audit teams get one exact model kept running, so last year’s answer can be replayed byte for byte. Same model. Same settings. Same bytes. Next year. Cloud shops delete old versions. Your audit then cannot replay last year's answer. This is a frozen copy plus a paper trail.

Price
Not on sale
Built for
Teams who must replay last year's model answer and cannot if the cloud shop deleted that version
Read
No pin is on sale. A pin would keep one model fixed and let you re-run sealed inputs on it.
Newest sealed read
2026-09-01

Email us. A person replies with what we can and cannot do for that pin today.

Email a questionNot on sale right now.

Newest sealed read: 2026-09-01. Sample pin SAMPLE-2026-09-01-s2. Five re-runs of one sealed input produced one SHA-256. Engine fingerprint vllm-0.25.1-9f488e66 on all five. This is a hosted pin plus a paper, not a data feed and not a weights download.

What you get

One pin = one model version + one settings list + one machine. The pin includes the engine fingerprint printed on the certificate. A different engine is a different pin.

  • The pin stays up.We keep that exact model and those exact settings running on our hardware. We do not update it.
  • You may re-run sealed inputs on that pin.You save a question. You send it back later. You get the same bytes.
  • One signed certificate per run.Weights first/last 1 MiB hashes, settings, input hash, output hash, date.

First certificate the same day you run your first sealed input.

Five-repeat samplereal run 2026-09-01 · 5 identical hashes

Real run on our hardware, 2026-09-01, after 12:45 MST. Nothing was restarted. Five posts, about 1.1 seconds apart, same sealed input, temperature 0, seed 42, max tokens 64. Hash is SHA-256 of the reply bytes as UTF-8.

Output SHA-256 of each run (message content UTF-8) identical = true
RunOutput SHA-256Engine fingerprintfinish_reasonWhen (UTC)
167266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638vllm-0.25.1-9f488e66length2026-09-01T21:50:44Z
267266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638vllm-0.25.1-9f488e66length2026-09-01T21:50:51Z
367266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638vllm-0.25.1-9f488e66length2026-09-01T21:50:58Z
467266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638vllm-0.25.1-9f488e66length2026-09-01T21:51:05Z
567266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638vllm-0.25.1-9f488e66length2026-09-01T21:51:12Z
  • Runs5
  • Unique output hashes1
  • identicaltrue
  • Input SHA-256864227ab119fff2bc364478287b933b4a1f3eeb817b42809f627a470c0710f83
  • Sealed promptReply with exactly five words: frozen model custody holds.
  • Model servedqwen38-s2 · root unsloth/Qwen3.8-27B-NVFP4 · revision 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108
  • Content length285 bytes each run

Weights fingerprint on the sample paperfirst and last 1 MiB only

Read-only weights cache at /mnt/glm-weights, snapshot 7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108. SHA-256 of the first 1 MiB and last 1 MiB of each weights file. Full 22 GB file hash: UNKNOWN (not rehashed this run).

Weights ends hashed 2026-09-01 not a full-file hash
FileBytesSHA-256 first 1 MiBSHA-256 last 1 MiB
model.safetensors22,568,192,096c8d2d06d4cff3588cabdb630d6a060b3d72162d233ec6b20f975a11461507d3a83cda4b4730bdf6ae81dd1d4d94958ca0e6ff7f41e23564b97770fac6ef2d841
model_mtp.safetensors849,400,392f43c6b807aa1c068a7210a0b88c5bf28b74befc3a447fabe33f7b939523615435e1360631ecdbdfabf877bd415a4a0b75f2590f0a6139fbbb7ab22879bc99da1

What this cannot tell you

We do not guarantee the answer is correct. We do not warrant truth, safety, legal fitness, or that a regulator will accept the paper.

We do not promise a match on anyone else's hardware. A signed certificate is our statement about our box. It is not a court exhibit unless a lawyer later says so.

Cross-day replay is not proved by this sample. The window was about 28 seconds (2026-09-01T21:50:44Z to 21:51:12Z).

Replay under load is not proved. The sample pin's recipe still allows 4 parallel jobs.

The five-word instruction was not proved. All 5 runs hit finish_reason=length (the 64-token cap cut the text; the cut bytes still matched).

Full-file SHA-256 of the 22,568,192,096-byte weights file is UNKNOWN. First and last 1 MiB only.

Uptime, power cost, and a signed company key are UNKNOWN this run. The sample certificate says NOT SIGNED. It is a format sample.

Extra pins, other model families, phone support, and correctness are not included.

The model is Apache License 2.0. We host a frozen copy on our machine. Licensor provides the Work on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, including FITNESS FOR A PARTICULAR PURPOSE. We do not sell the NVIDIA / vLLM runtime.

See the sample

You do not have to take our word for the five matching hashes. Here are the real run rows from 2026-09-01, carrying the same hashes as the table above. Nothing in it is made up and nothing in it is tidied up.

  • Open the sample as a CSVFive rows. One SHA-256. Engine fingerprint vllm-0.25.1-9f488e66 on every row.
  • The same rows as JSONThe same five runs, laid out for reading with code.
  • Sample certificateThe paper format, printed here in full. Status: SAMPLE. Not signed. Not a live pin.
    Read the sample certificate
    # Frozen Model Custody — sample certificate

Status: SAMPLE. Not signed. Not a live pin.

## What this paper claims

On our machine, this exact model and these exact settings produced the same bytes five times for one sealed input. It does **not** claim the answer is true, safe, legal, or useful.

## Pin

| Field | Value |
|---|---|
| Product | Frozen Model Custody |
| Pin id | SAMPLE-2026-09-01-s2 |
| Model name served | `qwen38-s2` |
| Model root | `unsloth/Qwen3.8-27B-NVFP4` |
| Weights revision | `7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108` |
| Engine fingerprint | `vllm-0.25.1-9f488e66` (returned on all 5 calls) |
| max_model_len (recipe) | 32768 |
| max_num_seqs (recipe) | 4 |
| gpu_memory_utilization (recipe) | 0.60 |
| prefix caching (recipe) | on |
| Offline weights | `HF_HUB_OFFLINE=1` (recipe) |

## Weights fingerprint (read-only)

Cache: read-only mount at `/mnt/glm-weights`.

Snapshot:

`/mnt/glm-weights/hub/models--unsloth--Qwen3.8-27B-NVFP4/snapshots/7d6f8d4d72f56b92b3cdbf22f156b90e1bab0108`

| File | Bytes | SHA-256 first 1 MiB | SHA-256 last 1 MiB |
|---|---|---|---|
| `model.safetensors` | 22,568,192,096 | `c8d2d06d4cff3588cabdb630d6a060b3d72162d233ec6b20f975a11461507d3a` | `83cda4b4730bdf6ae81dd1d4d94958ca0e6ff7f41e23564b97770fac6ef2d841` |
| `model_mtp.safetensors` | 849,400,392 | `f43c6b807aa1c068a7210a0b88c5bf28b74befc3a447fabe33f7b93952361543` | `5e1360631ecdbdfabf877bd415a4a0b75f2590f0a6139fbbb7ab22879bc99da1` |

Full-file SHA-256 of the 22 GB weights: **UNKNOWN** this run (not rehashed). Hugging Face blob id for `model.safetensors`: `c473512c70eace07e2256fe9fd76596ac03e3295bee7d54cfb72676416afcc05`.

`config.json` SHA-256: `1b3c71868d1299e52df6fc907deb202d5132b1ef0f72aae0ef6d15185dd53a5c` (22,564 bytes).

The model is Apache License 2.0. We host a frozen copy on our machine. We do not sell a weights download. Licensor provides the Work on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, including FITNESS FOR A PARTICULAR PURPOSE.

## Sealed request

| Field | Value |
|---|---|
| Prompt (UTF-8) | `Reply with exactly five words: frozen model custody holds.` |
| Input SHA-256 | `864227ab119fff2bc364478287b933b4a1f3eeb817b42809f627a470c0710f83` |
| temperature | 0 |
| seed | 42 |
| max_tokens | 64 |
| endpoint | `POST /v1/chat/completions` |

## Sealed answer

| Field | Value |
|---|---|
| Runs | 5 |
| Identical | **true** (1 unique hash) |
| Output SHA-256 (message content UTF-8) | `67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638` |
| Content length | 285 bytes each run |
| finish_reason | `length` each run (hit the 64-token cap; truncated bytes still matched) |
| Timestamps (UTC) | 2026-09-01T21:50:44Z … 21:51:12Z |

Five identical output SHA-256 lines:

```
67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638
67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638
67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638
67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638
67266ea3e0c06a4259669d9ce46fbad22104206059166eb73d405fa8cd7e3638
```

Engine fingerprint on all five: `vllm-0.25.1-9f488e66`.

Machine record: `sample-determinism.json`.

## Promise vs not

**What the sample shows:** We do not guarantee the answer is correct. The signed paper shows the same first/last 1 MiB weights hashes and the same settings.

**We do not promise:** that the answer is correct. That a regulator will accept it. That another vendor, GPU, or software version will match. That the model is safe. Uptime this run: UNKNOWN (not measured).

## Signature

```
USTA-SIGN: NOT SIGNED — sample only.
Operator applies the company key. This file is a format demo.
Signed-over: pin table + weights table + input hash + output hash + date.
Date: 2026-09-01
```
  • Machine record of the five runsThe raw determinism file would come with a pin. The sample rows above are cut from the 2026-09-01 run of it.

Not on sale right now

Not on sale right now. We cannot yet deliver this automatically after payment, so we are not taking payment.

Email a question

No payment is taken on this page.