A model gets packaged into a sealed .cnox container. SHA-256 is verified before a single op executes. Inference walks a fixed plan over a minimal opset. Every run can emit a per-op audit log: op type, output tensor hash, output sample — cryptographically linked to the exact container and input that produced it. If something goes wrong in production, you have a trail.
Scalar backend today — reference implementation and permanent fallback when hardware acceleration isn't available. Audit and verification is identical across all backends. SIMD next, GPU after that.
Input below is synthetic (all-ones) — pipeline is identical with real inputs.
github.com/Coelanox/CLF Audit example: { "schema": 2, "run": { "run_id": "59144ede-5a27-4dff-bc25-94abade5b215", "started_at_unix_ms": 1776535116721, "container_path": "/home/shark/cnox/models/output/bert_base_uncased.cnox", "container_sha256_hex": "184c291595536e3ef69b9a6a324ad5ee4d0cef21cc95188e4cfdedb7f1f82740", "backend": "scalar" }, "input": { "len": 98304, "sha256_hex": "54ac99d2a36ac55b4619119ee26c36ec2868552933d27d519e0f9fd128b7319f", "sample_head": [ 1.0, 1.0, 1.0, 1.0 ] }, "ops": [ { "op_index": 0, "op_type": "Add", "out_len": 98304, "out_sample_head": [ 0.12242669, -4.970478, 2.8673656, 5.450008 ], "out_sha256_hex": "19f8aa0a618e5513aed4603a7aae2a333c3287368050e76d4aca0f83fb220e78" }, { "op_index": 1, "op_type": "Add", "out_len": 98304, "out_sample_head": [ 0.9650015, 0.23414998, 1.539839, 0.30231553 ], "out_sha256_hex": "7ae2f025c8acf67b8232e694dd43caf3b479eb078366787e4fdc16d651450ad4" }, { "op_index": 2, "op_type": "MatMul", "out_len": 98304, "out_sample_head": [ 1.0307425, 0.19207191, 1.5278282, 0.3000223 ], "out_sha256_hex": "44c28e64441987b8f0516d77f45ad892750b3e5b3916770d3baa5f2289e41bdd" }, { "op_index": 3, "op_type": "Gelu", "out_len": 393216, "out_sample_head": [ 0.68828076, -0.0033473556, 1.591219, -0.16837223 ], "audit_elided": "hash_skipped: len 393216 > max 262144" }