01 / WITHOUT SEMANTICCOMPUTE
Inspect the output
The image looks plausible at normal scale. Isolated bad channels, NaNs concealed by rendering and one-step rounding differences have no quantified location or contract verdict.
Visual review establishes plausibility. SemanticCompute compares a candidate texture with an independent CPU reference, localises every failed pixel and channel, separates one-ULP drift from material divergence or special values, and rechecks the developer’s correction.
DETERMINISTIC ILLUSTRATIVE FIXTURE · NOT A GPU PERFORMANCE RUN
normal image view
01 / WITHOUT SEMANTICCOMPUTE
The image looks plausible at normal scale. Isolated bad channels, NaNs concealed by rendering and one-step rounding differences have no quantified location or contract verdict.
02 / WITH SEMANTICCOMPUTE
Compilation, dispatch, readback and parity become separate stages. Every channel is checked under a named policy, with the first failure, worst finite error and special-value count exposed.
03 / AFTER THE FIX
SC does not rewrite the shader. The developer applies the remedy, then the corrected candidate is compared against the same reference and policy.
Exact comparison treats any distinct Float32 bit pattern as a failure. ULP(1) accepts one representable step but never hides a NaN or a material finite divergence.
| Classified cause | Observed | Fails exact | Fails ULP(1) | Image reflection |
|---|
same reference · re-verified
RE-VERIFICATION
The corrected fixture is bit-identical to the trusted reference. A real receipt should also record the SC build identity, backend, precision, texture format, dimensions, tolerance and execution-stage status.
CURRENT SOURCE SNAPSHOT
| Lane | Without SC | Current SC | After requested additions | Before/after statistic |
|---|---|---|---|---|
| Image parity | Plausibility only | Location, channel, ULP, absolute error and stage receipt | Same verified path; storage/code families do not alter pixels | Failed channels, max ULP, max absolute delta |
| Disk history | Exact raw observations | Exact UInt64-pair cross-check | Irregular-time growth-onset evidence · local runtime receipt | 0 mismatches · 2.976 ms median · 9.36 MiB peak RSS |
| Dependency overlap | Host Set/Jaccard | Family discovery only | Variable-span exact intersection and union · local runtime receipt | 0 / 25,000 pair mismatches · 4.618 ms median · 9.75 MiB |
| Declaration matching | Structural clusters | Bounded dense assignment is source-present | Sparse lexicographic assignment · exhaustive small-graph oracle | 0 / 64 solution mismatches · 117.834 ms median for 12,800 solves · 6.23 MiB |
| Semantic graph | Consumer grouping and layout | CSR preparation plus bounded force/spring layout | Components, SCC condensation and topological strata · local runtime receipt | 0 mismatches over 12,000 nodes · 8.330 ms median · 9.42 MiB |
| Family and identical fixture | Without SC | With SC | Elapsed change | RSS change | Correctness |
|---|---|---|---|---|---|
| Variable-span overlap 25,000 pairs · Swift Set baseline |
16.584 ms 9.94 MiB |
4.618 ms 9.75 MiB |
−72.2% · 3.59× faster | −1.9% | 0 / 25,000 mismatches |
| Sparse lexicographic assignment 12,800 small solves · exhaustive 3×3 oracle |
56.370 ms 5.92 MiB |
117.834 ms 6.23 MiB |
+109.0% · 2.09× slower | +5.3% | 0 / 64 mismatches |
| CSR graph components 12,000 nodes · Set-backed host baseline |
9.632 ms 12.25 MiB |
8.330 ms 9.42 MiB |
−13.5% · 1.16× faster | −23.1% | 0 mismatches |
| Irregular-time growth burst 500,000 evaluated samples · direct host oracle |
2.568 ms 10.17 MiB |
2.976 ms 9.36 MiB |
+15.9% · 1.16× slower | −8.0% | 0 mismatches |
Every paired figure is the median of seven independent release-mode processes with a result-derived DCE barrier. Wins, washes and losses are retained: the small exhaustive assignment oracle beats SC at 3×3, while SC’s exact objective is intended to scale beyond that validation fixture; the growth family trades 15.9% elapsed time for explicit coverage and abstention semantics. The 36-checkout MacResilience fixture has not yet been replayed through these adapters.
LOCKED FIXTURE · EDITABLE AFTER-RECEIPT
The before column is the retained MacResilience fixture. The after column stays empty until the same fixture runs against an identified SC build. Count invariants guard against silent omission; resource changes remain unattributed until the evidence gate is runtime-validated.
| Core fixture measure | Before | After landing | Delta / guard |
|---|---|---|---|
| Same-repository checkouts included | 36 | pending | |
| Declaration records | 311,284 | pending | |
| Overlap clusters | 98,291 | pending | |
| Divergent clusters | 3,563 | pending | |
| Parent helper peak RSS MiB · lower is better | 774 | pending | |
| Largest semantic artifact MB · lower is better | 48.0 | pending |
| Requested SC family | Correctness mismatch | Elapsed ms | Peak RSS MiB | Result |
|---|---|---|---|---|
| Variable-span exact set overlap 25,000 candidate pairs · 7 processes |
pending | |||
| Sparse lexicographic assignment 12,800 solves · exhaustive 64-case oracle · 7 processes |
pending | |||
| CSR components / SCC / strata 12,000 nodes · 2,400 SCCs · 7 processes |
pending | |||
| Irregular-time growth-burst detector 500,000 evaluated observations · 7 processes |
pending |
The image comparison remains a deterministic browser illustration of sc_validate_metal_texture, not a customer-image or GPU-performance result. The paired family receipts came from release-mode local processes built from HEAD 55ae773 plus a dirty working tree whose 249-file package-source snapshot has SHA-256 4d5335104514…; the MCP executable still reports version 1.22.1 and buildCommit unknown. The figures establish local source/runtime evidence, not a clean commit, tagged release, or MacResilience 36-checkout integration result.