You do not have to trust our number. Recompute it.
Every Track A score artifact carries the full prediction matrix and the digest of that matrix. The script below re-derives the test split from the published rule, re-applies the trained head, and compares every prediction it computes against the ones the artifact reports — by digest, so a single changed prediction fails the check.
python3 and numpy. Nothing else — no GPU, no network, no
Nexido account, and no Nexido code beyond the published split rule.reproduce.py.python3 reproduce.py /path/to/the/artifacts
The features artifacts record a SHA-256 for every image. The script uses those, with the split it has just re-derived, to count the test images that are byte-identical to a training image, and names the classes they fall in. If the score artifact carries a leakage result of its own, the two must agree or the run fails. Finding duplicates does not fail the run: it is a fact about your archive, and you are the one who needs to know it.
A check that cannot fail proves nothing, so the script ships with four deliberate faults you can trigger yourself:
BREAK=flip python3 reproduce.py DIR # changes one prediction -> the digest must break BREAK=seed python3 reproduce.py DIR # alters the declared seed -> the rule must refuse BREAK=bias python3 reproduce.py DIR # drops the bias column -> the result must break BREAK=leak python3 reproduce.py DIR # plants one duplicate -> the count must move
All four must fail. If any of them passes, the clean run told you nothing.
parent. An artifact declaring
any other is refused rather than guessed at.An LLM fine-tune runs as three stages — training, validation scoring and test scoring —
and the job page offers, for each stage, its file (the adapter, or a score.json) and its
verification record. Two tools check them, and neither contacts Nexido.
In your browser: check a record here. It runs in your browser and sends nothing anywhere.
On your own machine: verify-record.mjs needs Node 18 or later and nothing else.
node verify-record.mjs record.json --artifact score.json --input your-file.jsonl
reproduce_llm.py scores the held-out conversations again on your own
machine, with the public model and the very scorer our nodes run, and compares every conversation with
the score.json you were given. Keep these five files side by side:
You also need the public Qwen/Qwen3-4B-Instruct-2507 at the commit our image holds, and
Python with the libraries the scorer uses. The versions below are our image's, on Python 3.10; with them, on the
same GPU class as ours, you get our score.json byte for byte. It runs on Linux or macOS; on Windows,
run it under WSL.
pip install torch==2.4.1 transformers==4.57.6 peft==0.20.0 safetensors==0.8.0 tokenizers==0.22.2 accelerate==1.14.0
huggingface-cli download Qwen/Qwen3-4B-Instruct-2507 \
--revision cdbee75f17c01a7cc42f958dc650907174af0554 --local-dir qwen3-4b
python3 reproduce_llm.py --record record.json --score score.json \
--adapter adapter.safetensors --input your-file.jsonl --model qwen3-4b
A GPU needs 11 GiB free; add --device cpu to run on a processor instead, with about 16 GB
of memory, more slowly.
score.json you kept is the output it names, as its signed receipt does. It names the part, the
split, the counted file and the adapter that the record names, every sum in it is a finite number, and its
decision digest is the digest of its own per-conversation decisions.score.json byte for byte.
Expected with the versions above on an RTX A4500, the GPU class our scoring runs on.Each control changes one thing after every check above has passed, so only the check named for it can catch it:
BREAK=flip python3 reproduce_llm.py ... # one decision changed in your copy -> the decisions differ BREAK=sum python3 reproduce_llm.py ... # one sum moved by 1 in 1,000 -> the sums differ BREAK=adapter python3 reproduce_llm.py ... # the adapter's B matrices zeroed -> the decisions differ BREAK=seed python3 reproduce_llm.py ... # the split's seed changed -> the scorer refuses it BREAK=file python3 reproduce_llm.py ... # one row dropped from your file -> the scorer refuses it
Each must end CONTROL CAUGHT, naming the check that caught it. If one ends CONTROL NOT CAUGHT, the clean run told you nothing.
score.json says, on held-out conversations the split names. Training is not reproducible
bit for bit across GPU models, and the product does not claim it is.verify-record.mjs does, on the same record.If the verifier disagrees with an artifact we gave you, that is a defect and we want to hear about it. A score we cannot hand over is not a product.