Phase 0 run on CPU node

[bitzic]

You post test results, screenshots, measured numbers after each milestone. Only maintainers create threads; community comments underneath.

[bitzic]

Phase 0 — End-to-end MVP verified on CPU

What was tested

• Full miner → proof-test → validator → scoring → king-change cycle
• Determinism: two runs with seed=42 produce bit-identical loss (4.233057)
• Integrity: tampered checkpoint → rejected at op1; restricted-file patch → blocked before training

Measured numbers

Metric	Value	Notes
Noise floor (5 runs)	σ = 0.10 val_bpb	Suggested 2σ margin = 0.20
Proof-test wall-clock (CPU, 20 steps)	2.6s	Will re-measure on H100
Validator 4-ops time	<1s per submission	Excluding hidden-eval inference

Screenshots

What's next

Phase 0.5: rent H100, real FineWeb data, real Docker build, real attestation.

How to reproduce

git clone [git@github.com](mailto:git@github.com):KarpathianBase/karpathian.git
cd karpathian
python3 -m venv .venv && source .venv/bin/activate
pip install torch numpy tiktoken cryptography
python scripts/smoke_test.py

[bitzic]

Try to submit a patch that touches a restricted file:

mkdir -p submissions/bad_patch && cat > submissions/bad_patch/patch.diff <<'EOF'
--- a/eval/val_bpb.py
+++ b/eval/val_bpb.py
@@ -1 +1 @@
-"""val_bpb"""
+"""hax"""
EOF
cat > submissions/bad_patch/proof_request.json <<'EOF'
{"handshake_nonce": "0xfake", "seed": 1, "config_path": "configs/proxy_cpu_smoke.json", "miner_hotkey": "5BadMiner"}
EOF
.venv/bin/python -m proof.runner --submission submissions/bad_patch --out-dir runs/bad

Expect: RuntimeError: patch touches restricted paths: ['eval/val_bpb.py'] — the proof runner refuses to even start training.

:128: RuntimeWarning: 'proof.runner' found in sys.modules after import of package 'proof', but prior to execution of 'proof.runner'; this may result in unpredictable behaviour
Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in _run_code
File "/workspace/unicorn/karpathian/karpathian/proof/runner.py", line 341, in
main()
File "/workspace/unicorn/karpathian/karpathian/proof/runner.py", line 329, in main
bundle = run_proof_test(
^^^^^^^^^^^^^^^
File "/workspace/unicorn/karpathian/karpathian/proof/runner.py", line 179, in run_proof_test
raise RuntimeError(f"patch touches restricted paths: {violations}")
RuntimeError: patch touches restricted paths: ['eval/val_bpb.py']

[bitzic]

Measure the actual noise floor:

.venv/bin/python scripts/noise_floor.py --runs 10 --base-seed 5000

[noise_floor] run 10/10 (seed=5009)
[proof] running training: /workspace/unicorn/karpathian/karpathian/.venv/bin/python -m recipe.train --out-dir /workspace/unicorn/karpathian/karpathian/runs/noise_floor/proof_baseline_seed5009/training --seed 5009 --config /workspace/unicorn/karpathian/karpathian/configs/proxy_cpu_smoke.json --manifest /workspace/unicorn/karpathian/karpathian/data/data_manifest.json
[train] device=cpu params=6,761,216 (no embeddings: 328,320)
[train] manifest tokens=200,000 hash=d333a00be6e067a2…
[train] steps=20 batch=4 micro=4 seq=128
[step 0/20] loss=10.8627 lr=6.00e-04 |g|=2.88 tok/s=3,495
[step 2/20] loss=10.3505 lr=1.80e-03 |g|=2.51 tok/s=3,889
[step 4/20] loss=9.5606 lr=3.00e-03 |g|=2.04 tok/s=3,820
[step 6/20] loss=8.3964 lr=2.97e-03 |g|=2.07 tok/s=3,815
[step 8/20] loss=7.1801 lr=2.74e-03 |g|=2.03 tok/s=3,819
[step 10/20] loss=6.0589 lr=2.32e-03 |g|=1.97 tok/s=3,804
[step 12/20] loss=5.3098 lr=1.79e-03 |g|=1.64 tok/s=3,804
[step 14/20] loss=4.7867 lr=1.23e-03 |g|=1.37 tok/s=3,829
[step 16/20] loss=4.4904 lr=7.47e-04 |g|=1.26 tok/s=3,837
[step 18/20] loss=4.3295 lr=4.17e-04 |g|=1.15 tok/s=3,842
[step 19/20] loss=4.3123 lr=3.30e-04 |g|=1.06 tok/s=3,823
[train] done. final loss=4.3123 wall=2.7s

============================================================
NOISE FLOOR CALIBRATION — 10 baseline runs

val_bpb: mean=4.0206 std=0.1515 min=3.9131 max=4.2754
benchmark: mean=0.1740 std=0.0358
wall-clock: mean=5.27s total=52.71s

suggested noise-floor margin (2σ): 0.3030 val_bpb
written to: runs/noise_floor/noise_floor_summary.json

[bitzic]

Here is the summary json.

{
"runs": 10,
"config": "configs/proxy_cpu_smoke.json",
"val_bpb": {
"mean": 4.020586799471742,
"std": 0.15152214768055,
"min": 3.913088474642254,
"max": 4.27538636258251,
"values": [
3.914982445869356,
3.915860463710027,
4.268525925501692,
3.923282863657131,
4.2067119320481146,
3.916056534851544,
3.913088474642254,
4.27538636258251,
3.945085677582512,
3.9268873142722804
]
},
"benchmark_accuracy": {
"mean": 0.174,
"std": 0.03583294573433783,
"values": [
0.16,
0.2,
0.14,
0.24,
0.16,
0.2,
0.14,
0.14,
0.14,
0.22
]
},
"wall_clock_s": {
"mean": 5.27147057056427,
"total": 52.7147057056427
},
"suggested_noise_floor_margin": 0.3030442953611,
"results": [
{
"seed": 5000,
"val_bpb": 3.914982445869356,
"benchmark_accuracy": 0.16,
"wall_clock_s": 5.207294702529907,
"bundle_hash": "eadcfeac9fd1a356e881117c3722b5b671f89950e55471a54f1cbbd162c0e632"
},
{
"seed": 5001,
"val_bpb": 3.915860463710027,
"benchmark_accuracy": 0.2,
"wall_clock_s": 5.247988224029541,
"bundle_hash": "84bc9373ad13e12c371e5cf13730be803640c8027ea3462cd7cf9aeb51e7afb7"
},
{
"seed": 5002,
"val_bpb": 4.268525925501692,
"benchmark_accuracy": 0.14,
"wall_clock_s": 5.141977071762085,
"bundle_hash": "2160f164bef8290a86127ba0ad40200313b7aaf87febccf55ff214f34f8b67e4"
},
{
"seed": 5003,
"val_bpb": 3.923282863657131,
"benchmark_accuracy": 0.24,
"wall_clock_s": 5.2615966796875,
"bundle_hash": "b6d9a8085c85e3d8396230b02a747cc331892e3530c469e50c1d1e4935e61e33"
},
{
"seed": 5004,
"val_bpb": 4.2067119320481146,
"benchmark_accuracy": 0.16,
"wall_clock_s": 5.780532598495483,
"bundle_hash": "ced59a7645ed77f8870e6a762bdc63a53b647aa1db8f526764ab60de58be0467"
},
{
"seed": 5005,
"val_bpb": 3.916056534851544,
"benchmark_accuracy": 0.2,
"wall_clock_s": 5.48534083366394,
"bundle_hash": "35a6c31bc123b8f9e8c7a2e0ec55bf00f0e69b51a0faaad2d96397269e897230"
},
{
"seed": 5006,
"val_bpb": 3.913088474642254,
"benchmark_accuracy": 0.14,
"wall_clock_s": 5.1017961502075195,
"bundle_hash": "e15bc59ad58001f27fc7863552f8aa9d2bbaa8dca33debeebe36d33f231ddb4a"
},
{
"seed": 5007,
"val_bpb": 4.27538636258251,
"benchmark_accuracy": 0.14,
"wall_clock_s": 5.213766574859619,
"bundle_hash": "b6c496bea2361ca0671df0cad3bee44aee6f7a30a89378826fa9c2e4c3171602"
},
{
"seed": 5008,
"val_bpb": 3.945085677582512,
"benchmark_accuracy": 0.14,
"wall_clock_s": 5.119000673294067,
"bundle_hash": "dfe2492693a70c4a49aa2ba2486b08101cf64867c7876681807d3d2b275614b1"
},
{
"seed": 5009,
"val_bpb": 3.9268873142722804,
"benchmark_accuracy": 0.22,
"wall_clock_s": 5.155412197113037,
"bundle_hash": "14215c6c448f733f9bded1694b82dc72e3669d10fdbf889a01ee83c105e8bc2d"
}
]
}

[bitzic]

Great, now let's read the chain like a historian!

[bitzic]

cat chain/events.jsonl | jq -r '"\(.timestamp | todate) \(.type) \(.miner_hotkey // "")"'

2026-05-25T13:41:39Z submission_scored 5MinerA_baseline
2026-05-25T13:41:39Z initial_king
2026-05-25T13:41:45Z submission_scored 5MinerB_raise_lr
2026-05-25T13:41:45Z king_changed