ProtocolLedger v1.0.0

Image: ghcr.io/professor3333/protocolledger@sha256:d409389f086996acab756326252eae74287e35d0773f1cb7c6845d2401712cf3

Model scibert-uniform-negatives-s21-7dc7a1893e (bundle.json SHA-256 f96032c532caf6704dc6f943c91d725c03ab47ffd29d7f8b45ab880c96973d06), runtime torch-fp32, code 81d700a77d. Gates checked on GitHub Actions ubuntu-latest, 4 vCPU, dev split only (112 protocols); configuration: configs/release_gates.json.

Gate	Measured	Limit	Passed
dev entity F1 reproduces bundle.json	0.8003	0.8003 ± 0.002	yes
dev end-to-end relation F1 reproduces bundle.json	0.6389	0.6389 ± 0.002	yes
precision at 30% coverage reproduces the calibration report	0.9205	0.9205 ± 0.01	yes
ECE after isotonic calibration	0.0086	≤ 0.0087 + 0.01	yes
p95 latency, ≤ 500-word dev protocols (s)	2.8928	≤ 3.0	yes
peak resident memory (MiB)	1291.7	≤ 3072	yes

p50 latency 1.0839 s over 200 sequential requests. ONNX / INT8 parity is not gated: the served runtime is PyTorch FP32.

The gates use the dev split only. Locked-test results (evaluated once, before
this release) are in the README.