Results with scope

What does the evidence prove?

01 / Software accuracy

How accurate is the calibrated proxy?

The pinned 320×320 YOLOv8n FLOAT export and signed-symmetric INT8 ONNX Runtime proxy were evaluated over 5,000 COCO val2017 images. The proxy loses 0.887 mAP50–95 points, within the project’s one-point budget. A deterministic 512-image train2017 subset supplied calibration data.

FLOAT baseline

28.50 mAP

YOLOv8n · COCO val2017 · 320 × 320

Exact report value: 0.284969
Calibrated INT8 proxy

27.61 mAP

Signed-symmetric INT8 proxy · 320 × 320

Exact report value: 0.276098
Within the 1.0-point budget

−0.89 mAP points

3.11% relative loss · 96.9% retained

Exact loss: 0.887 points

Scores use the familiar 0–100 mAP scale; exact 0–1 report values remain visible for traceability. YOLOv8n · 320 × 320 · 5,000 COCO val2017 images. Both software evaluations were repeated with identical predictions. The INT8 result uses an ONNX Runtime QOperator proxy; COCO-wide HASLAB command-path accuracy and hardware measurements remain future work.

Why the INT8 proxy is a separate result

ONNX Runtime uses QLinear sigmoid and multiply rather than HASLAB’s fused SiLU lookup, and requantizes learned-head outputs where v0 preserves an INT32 boundary. The proxy measures a calibration candidate; it does not establish whole-detector HASLAB command accuracy. See the INT8 report and calibration record.

02 / Command correctness

6,931,200 values. Zero mismatches.

The compiler and simulator execute accelerator nodes 0–217 through three raw INT32 learned-head boundaries. The synthetic diagnostic run compares 106 retained or aliased boundaries with the independent integer oracle; release mode matches all 302,400 final INT32 values. Across six additional diagnostic inputs, 37,900,800 stored values match exactly and 1,058,400 decoded host-tail values pass the stated tolerance.

Command records412,092ABI 0.1 candidate
Stored values compared6,931,200106 synthetic boundaries
Integer mismatches0Diagnostic + release
Peak output storage2,355,200Bytes · release mode

The detection-head report records 44,598,232 scheduled DMA bytes and exact final boundaries. The six-input report records independent accelerator and host-tail comparisons. These counts are not hardware timing or throughput. COCO-wide command-path accuracy, command delivery, RTL, and FPGA execution remain open.

Understand command execution →
03 / Workload accounting

What would the command path move?

The full candidate schedule makes command transport and package compaction a concrete pre-RTL decision.

Per pinned inference: 412,092 command records, 52.75 decimal megabytes of command bytes, 44.60 decimal megabytes of scheduled DMA, and 1.093 billion INT8 multiply-accumulates.
These are deterministic per-inference schedule counts, not measured hardware latency, bandwidth, power, or throughput. Command delivery and package size remain pre-RTL gates. Source report ↗
04 / Independent expectations

What do the 46 fixtures check?

The candidate conformance corpus stores binary commands, initial and expected memory images, completion state, errors, hashes, and provenance. It covers successful and rejected behavior so another implementation can consume the same expected bytes. Stable ABI freeze and final corpus release await independent review.

Numerical rules

Rounding, saturation, accumulation order, layouts, conversions, and SiLU behavior have documented semantics and tests.

Command behavior

Memory bounds, DMA, lifecycle, reset, completion, and architectural faults are specified in the v0 contract.

05 / Source reports

Reproduce each claim.

Reports below are linked to reviewed source revision e605bb0. Model weights and COCO images require their own setup; the workload guide gives commands and artifact requirements.