PoC 13F — 10,000-Frame Erf NPU Soak

Question

Can one sealed graph execute 10,000 consecutive diverse frames without numerical drift, buffer corruption, timeout, or device-health change?

Input and oracle

The native runner generates one continuous xorshift32 byte stream from seed 0x5733130f. Each of 10,000 frames contains 481 bytes. It reports cumulative input and output FNV-1a64 values every 1,000 frames.

The host independently generates the same inputs and maps them through its math.erf U8 table. The native runner contains no expected output table.

Final expected digests:

input:  0x1417bf4ceca8d26e
output: 0xcbd4ae7bafd1fccd

Safety envelope

• one context, graph, NBG node, and compile;
• exactly 10,000 sequential Run calls;
• fixed seed, shape, quantization, NBG, paths, and argument;
• ten bounded checkpoint lines;
• 32-byte output canaries on every frame;
• 30-second native watchdog and 40-second host deadline;
• no output file, retry, recovery, CLC, arbitrary input, or direct ioctl.

Result

status: PASS
execution attempts: 1
frames: 10,000 / 10,000
oracle-covered bytes: 4,810,000
checkpoint matches: 10 / 10
mean graph run: 84.8758 us
minimum graph run: 74 us
maximum graph run: 1,649 us
10 ms budget fraction: 0.00848758
canaries: PASS
health differences: none
kernel alerts: none

Evidence

Artifact	SHA-256
Runner binary	ef0d4ece1d44141b1254581aa5b05e5fd371f99f819cfac32b478cbcaaaf4cff
Soak manifest	ed64868e2c28b057066513c4b1f390d3686f2fd16c47f47fa9a3b469bb6acd74
Run result	e4f3fc3d0de3c833dadf912563585c3737d7dc30c7338f2e26d9936077ee6f8d

Regression

After the soak:

wave-onnx-hybrid: 187 tests passed
NPU harnesses:    338 tests and 302 subtests passed
Ruff:             passed
shell syntax:     passed

Interpretation

The exact partition has strong sequential stability and a large graph-run timing margin. This is not the originally planned full production Rust soak: it does not measure p50/p95/p99, RSS growth, CPU load, allocations, reset reproducibility, thermal behavior, or end-to-end audio-service scheduling.

Next gate

Port the exact model and latched fallback behind production Rust NpuInfer, then collect the omitted service-level statistics before enabling it outside the lab.