SCSEMANTICCOMPUTE / RECORDED IMAGE PARITY CASE STUDY

The image can look right and still be numerically wrong.

A real SC 1.20.0 family run makes an invisible implementation assumption measurable, locates its effect, and re-verifies the corrected path.

DETERMINISTIC 48 × 48 INSPECTION FIXTURE · NOT A GPU PERFORMANCE BENCHMARK

2,304 FLOAT32 PIXELS · SEPARABLECONV 2D

Candidate output versus trusted reference

Candidate Reference Deviation

Linear Log reveal

Undeclared zero boundary: visually plausible

implicit zero boundary

Worst pixel (0, 21) · reference 0.696996 · candidate 0.484007 · |Δ| 0.212988

01 / WITHOUT SC

Inspect

The output is recognizable, but the darkened outer band is easy to accept as ordinary filtering.

02 / WITH SC

Measure and diagnose

The family, precision, boundary model and tolerance become explicit; 368 pixels fail.

03 / AFTER THE FIX

Re-verify

The clamp-boundary family path passes the same comparison with zero failing pixels.

What the numbers differentiate

SC separates a material semantic mismatch from normal Float32 rounding instead of reducing both to “the images look close.”

Classification	Pixels	Share	How it reflects
Numeric divergence	368	16.0%	Visible dark boundary halo from the wrong edge policy.
ULP drift	720	31.3%	Normally invisible Float32 accumulation differences.
Exact Float32 match	1,216	52.8%	No representable difference at the compared precision.

Before and after the SC contract

Metric	Undeclared zero boundary	SC-prescribed clamp correction
Parity verdict	FAIL	PASS
Failing pixels	368	0
Mean absolute error	7.836e−3	1.249e−8
RMSE	2.705e−2	1.836e−8
PSNR	31.36 dB	154.72 dB
Maximum absolute error	0.212988	1.192e−7

Corrected family output

same fixture · same reference · re-verified

Re-verification 0 failing Maximum absolute error is 1.192e−7, a 1.79-million-fold reduction. The remaining difference is Float32 rounding within the stated tolerance.

Statistics are from the actual SemanticCompute 1.20.0 family and MCP parity run. The canvases recreate that deterministic inspection fixture for explanation. This is numerical-correctness evidence, not a GPU speed benchmark, and SC does not modify the kernel automatically.